Changelog
Changelog
What’s new and changed in the Primores AI Wiki.
July 2026
Week of July 20 – 26
New glossary term + article — why a YouTube “89× ROAS” is fake 📉 (July 22) — Prompted by a real situation: a team presented an 89× ROAS on ~€1,000 of YouTube spend at ~€60 AOV — with more conversions than clicks. That last detail settles it: on a single purchase event, one click can’t produce more than about one order, so if conversions exceed clicks, most of the credited conversions had no click — they’re engaged-view conversions (EVC), Google crediting a watch-then-buy that happened anyway. The new glossary/engaged-view-conversions page explains the mechanism (EVCs sit inside the main Conversions column for YouTube video-action/Demand Gen, which is why the count can exceed clicks), the two secondary tells (fractional counts = modeled conversions; restate an absurd ROAS as an absurd CPA — 89× at €60 AOV implies a €0.67 CPA), and why inflation is worst on high-traffic sites with strong organic demand (the ad stands in front of a river of existing demand and claims the water). The diagnostic ladder — backend order reconciliation → conversion-type segmentation → click-through-only floor → geo-holdout incrementality — gives you an afternoon-long proof and the one number a CMO will actually respect (incremental lift, ideally from a holdout you run yourself, not a platform study grading its own homework). The paired public article, “Why Your YouTube ROAS Is Fake (and the Number to Trust Instead),” makes the same case for the high-intent practitioner search. Back-links wired into incrementality-testing, multi-touch-attribution, performance-max, and the analytics pillar.
Week of July 13 – 19
Two new pages from a marketing-pain sweep: the email deliverability squeeze + the ad-account suspension black hole 📬🚫 (July 14) — A verified research pass asked “what are marketers actually struggling with in 2025–26 that this wiki doesn’t cover?” — and two problems earned pages. marketing/email-deliverability: inbox placement fell to 83.5% and spam placement nearly doubled across 2024 — roughly 1 in 6 legitimate marketing emails never reaches the inbox, and the failure went silent (junk-foldering your ESP reports as “delivered”). The Gmail/Yahoo bulk-sender rules — DMARC alignment, branded domain, sub-0.3% complaint rate, permanent rejections from Nov 2025, Microsoft joining — hit small senders hardest, because they inherit enterprise requirements with none of the infrastructure. marketing/ad-account-suspensions: Google suspended 39.2 million advertiser accounts in 2024 (with the honest caveat most coverage omits — the majority were fraud caught pre-serve), the vague “Circumventing Systems” policy is the #1 suspension reason, and both platforms have now admitted platform-side error — Google implicitly (its claimed 80% reduction in incorrect suspensions), Meta explicitly (a >100% false-positive spike in Q4 2025). The genuinely scandalous part: nobody publishes the wrongful-suspension rate — no audit exists. Also shipped: the survey data making measurement-pain the #1 marketing challenge folded into the analytics page, SMB capacity-strain numbers (42% have under an hour a day) into the automation overview, the Performance Max channel-reporting concession onto its glossary page, and a new public article — “Google Ads Suspended for ‘Circumventing Systems’ — What It Actually Means” — aimed at the desperate search query whose results are currently all recovery-agency lead-gen.
July monthly review — the wiki gets its two missing front doors 🚪 (July 14) — The month’s health check found the structure sound but the entrances missing: the SEO and Automation domains — 18 and 20 pages each — had no overview page at all. Both now exist: seo/overview tells the AI-era SEO story in three numbers (the ~47% measured CTR collapse under AI Overviews, the 13.7% overlap between AI citations and classic rankings, and the ~2/3 of ChatGPT prompts that get rewritten before retrieval) and routes into the visibility stack, measurement, access control, and the articles engine. automation/overview fronts the automation-eats-execution thesis, the 1,048-case evidence base (median ~50% improvement; the 90% Club eliminates tasks rather than optimizing them), and the adoption path. Also this review: two tool pages promoted from seedling (the AI email and video production stacks), the managed-agents break-even question parked as answered (~2,000–5,000 sessions/month on May pricing), the marketing overview’s Related refreshed with the paid-creative production cluster, and five April-era tool reviews queued for a freshness sweep — starting with the MCP page, which predates the entire ads-MCP wave. July so far: 6 new pages, 1 public article, 4 verified research passes in 14 days.
New glossary term — Query Fan-Out (why AI search doesn’t search what you typed) 🔀 (July 14) — AI engines rewrite the user’s question into multiple synthetic sub-queries and answer from the combined pool — Google documents this as AI Mode’s core retrieval mechanism, and ChatGPT fans out on roughly two-thirds of prompts (vs ~30% for Perplexity). The new glossary/query-fan-out page pins down the GEO consequence: you can rank #1 for the phrase a user types and still be invisible in the AI answer, because the model searched a reformulated version — so titles should match the model’s likely rewrites, and covering the sub-question cluster beats nailing one keyword. Also documented: the injection pattern (models add “best” and the current year to sub-queries — the structural reason listicles dominate AI answers), the engine-behavior split (ChatGPT non-deterministic, Perplexity mostly pass-through), and which numbers are Google-primary vs vendor-directional. Separately, the “research → 360 ad variants” article got a substantive refresh: the one-batch question discipline, the saturated-vs-open competitor map, situation+pain audiences with a mandatory risk field, the five angle checks, and the results-as-queue close.
Alt-channels — the pricing columns, filled (with honest gaps) 💵 (July 14) — A fourth research pass added a credibility-flagged CPM/CPC benchmark table to marketing/alternative-ad-channels, closing the roster’s last open column. The verified picture splits into three cost tiers against Meta ($8.19 CPM) and Google ($5.26 CPC): cheap discovery — Pinterest ($3–4.67 CPM, $0.50–0.70 ecom CPC) and native ads ($5–12 CPM, $0.10–0.90 CPC) are where the CPM arbitrage lives; premium social — Snapchat is the priciest major platform (~$8.60 CPM, spiking to $12.84 in Q4); high-attention inventory — streaming audio ($15–25 CPM) and CTV ($15–65 CPM) cost the most per impression. The neat through-line: the channels that are cheap to enter (Pinterest, native) are also cheap per impression, while the newly-self-serve premium inventory (CTV, audio) is cheap to enter but expensive per impression. Honest about the holes: only the social CPMs and Google’s CPC rest on large samples — everything in CTV/audio/native is triangulated agency data (no platform publishes official CPMs) — and X, Quora, and Nextdoor produced no verifiable pricing at all, while Reddit’s much-quoted “$0.59 CPC” couldn’t be re-triangulated (its $2–6 CPM is solid). Six inflated/wrong pricing figures were caught and listed as do-not-cite. Refresh-quarterly caveat on-page: ad pricing is a snapshot, not a stable rate.
Alt-channels — Pinterest & X verified, plus a new “what Telegram Ads actually cost” article 📡 (July 14) — A third research pass filled the last two open cells in marketing/alternative-ad-channels, both from the platforms’ own docs. Pinterest turns out to be a genuinely low-friction ecom channel — no approval gate, no hard minimum, and rare among alt-social in offering keyword/search-intent targeting plus catalog shopping ads. X (Twitter) produced the sharper finding: it’s a third case of the page’s core thesis — its “$0 spend floor” understates the barrier, because you now need a paid verification subscription (X Premium / Verified Organizations) just to be eligible to advertise, and its brand-safety reputation is the weakest of any major platform (only 4% of marketers rate it brand-safe vs 39% for Google). Separately, the research became a public explainer article — “How Much Do Telegram Ads Actually Cost in 2026?” — because the search results for that question are a mess of ad-network blogs quoting figures we’d already refuted (the “€2 million minimum,” the “€2 CPM,” the “160-character text-only ads”). The article sets the record straight from primary sources: TON-denominated pricing at a 0.1 TON minimum CPM, the €2M as a legacy-account myth rather than an entry ticket, and the “$150 Telegram ads” being third-party Mini App networks rather than the official platform. Remaining gap across the whole channel roster: live CPM/CPC benchmarks, which barely survive verification anywhere and need a dedicated pricing pass.
New page — Alternative advertising channels, ranked (and the Telegram-Ads myths, debunked) 📡 (July 14) — Prompted by a practitioner question — everyone’s talking about channels beyond Meta/Google, and Telegram Ads especially feels under-addressed. A verified deep-research pass turned it into marketing/alternative-ad-channels: a landscape of credible self-serve channels plus a ranked “try first” shortlist for DTC/ecom. The load-bearing insight: published entry floors mislead in both directions. Telegram’s famous “€2 million to advertise” overstates the barrier — a TON-denominated self-serve path now starts near $50. Reddit’s “$5/day minimum” understates it — the real gate is an advertiser-approval wall: whole verticals (alcohol, dating, gambling, finance, crypto, pharma) must work directly with a Reddit sales rep, and weapons/tobacco/vaping/“dupe” products are banned outright. So the ranking: Microsoft/Bing is the frictionless first move (no approval gate, CPC 33–42% below Google), Reddit has the best economics but only if your category clears policy, then Pinterest/Snapchat/Amazon, with Telegram as a specialist/arbitrage play (cheap attention, but contextual channel-level targeting and a confusing three-path access model). The page carries a full Telegram-Ads paid-product deep-dive (0.1 TON minimum CPM, what targeting actually exists, the reseller-vs-TON-vs-€2M access split). Sharpest result of the research: most of the “Telegram Ads facts” that circulate online are wrong — the 160-char text-only limit, the fiat €2 CPM floor, “€2M is the only way in,” and “channel-only targeting” all failed adversarial verification (13 claims confirmed / 12 refuted). Honest gaps flagged: Pinterest/Snapchat/CTV/native-DSP/newsletter mechanics didn’t survive verification this round, and Telegram’s exact current targeting set needs a fresh primary-source pass. (Update, same day — a second verified pass filled most of those cells: 22 confirmed / 3 refuted, mostly from primary sources. Snapchat $5/day, Quora fully self-serve, CTV went self-serve — Roku + Hulu both $500 floor no-revenue-requirement, Spotify $250, Nextdoor ~$1/day with rare audience/behavioral targeting for a low-floor channel, Taboola/StackAdapt self-serve-but-committed. Sharpest catch: beehiiv’s Ad Network is managed and sales-gated with a hard $15,000 minimum — not the low-floor self-serve channel people assume. Still open: Pinterest, X/Twitter, Netflix/Prime Video, Outbrain, Paved/Sponsy, and live CPM/CPC benchmarks.)**
New page — The Articles Engine + the AI-visibility adoption gap, quantified ⚙️ (July 13) — Born from a practitioner observation: client audits keep finding businesses that have never checked how AI describes them and have no systematic content program. A two-track verified research pass turned both halves into wiki material. The adoption gap is now quantified in seo/ai-visibility: in the broadest available sample, only ~24% of marketers track AI-search visibility (barely up from 22% in 2025), and while 54% say they prioritize GEO, only 12% have measurable results. The high-adoption numbers you’ll see quoted elsewhere (51% platform adoption, 12% of budgets) turn out to describe enterprise-only leadership samples surveyed by vendors — and there is no representative small-business data at all, which is itself the finding. The new seo/articles-engine page is the playbook for the other half: the six-stage loop (map → produce → publish → interlink → measure → refresh) that turns the wiki’s existing pieces into an operating system, with the measured freshness evidence (AI assistants cite fresher content — but updating existing articles captures much of the advantage, and the effect is strong on ChatGPT while absent on Google AI Overviews), and Google’s own line between an articles engine and content spam: purpose and user value, not production method — a line that since May 2026 explicitly governs what gets cited in AI answers too. Honestly flagged: nobody has measured publishing-cadence economics or time-to-citation; the viral “200× at 12+ pieces/month” stat failed verification outright.
New page — Ad Transparency Surfaces: every platform’s ad library as a CI data source 🔍 (July 13) — The wiki’s Meta Ad Library coverage was deep but Meta-only; the new competitor-analysis/ad-transparency-surfaces page maps the whole cross-platform landscape, adversarially verified. The organizing insight: EU-targeted ad data is structurally richer on every platform because the EU’s DSA (Article 39) forces large platforms to disclose targeting parameters and reach for EU ads — so TikTok’s library covers only EEA/Switzerland/UK, LinkedIn shows its remarkable B2B targeting disclosures (~12 parameter categories, per-country impressions) only for EU-targeted ads, and Microsoft’s free no-signup API covers only EEA-served Bing ads. Two universals worth tattooing on every CI workflow: no platform exposes spend data publicly (anything claiming competitor “ad spend” is modeled, not measured), and ads vanish ~1 year after last view — archive on sight. Also covered: Google’s July 9 AI-generation labels (they live in the consumer-facing My Ad Center, not the Ads Transparency Center, and only Google-AI-made ads get auto-labeled — label absence means nothing), and X’s ad repository being formally ruled deficient (€120M fine, the first-ever DSA non-compliance decision). Honest gaps flagged rather than filled: Google’s Transparency Center field set, TikTok’s Creative Center, and the minor platforms didn’t survive verification this round. Practical note: the intended “competitor-analysis stack page” turned out to already exist (competitor-analysis/overview) — this page is the data layer under it, not a duplicate.
Weekly health check — clean bill, two small fixes 🧹 (July 13) — Full lint after four ingest sessions: zero broken links, zero orphan pages, zero frontmatter issues across all 216 pages. Two drift items corrected: the index’s domain-page count was off by one, and marketing/ecommerce-to-tiktok-ai-pipeline still described TikTok’s Ads MCP as “not generally available” — updated to reflect the June 30 Agentic Hub launch (the page’s real point stands: there’s still no official MCP for organic TikTok posting).
Week of July 6 – 12
Mid-July sweep — the independent trust data finally lands, and it’s construct-dependent 📊 (July 12) — A second deep-research pass this week (every claim adversarially verified) chased the July 9 sweep’s open questions. The headline: independent longitudinal AI-trust data now exists — and it points both ways. The KPMG × University of Melbourne global study (48k+ respondents, university-led) shows trust in AI systems genuinely declining (trustworthiness 63%→56%, willingness-to-rely 52%→43%, 2022→2024) even as workplace adoption doubled — while Stanford HAI’s AI Index 2026 shows general AI-product optimism rising (55%→59%). seo/geo-aeo-benchmarks-2026 now frames the trust story as construct-dependent: systems-trust declining, product optimism rising, answer-trust low — and the specific answer-trust decline still rests on one vendor survey. (Edelman 2026 was checked and eliminated — no longitudinal AI measure.) The FTC’s AI-accuracy statement hit the Federal Register (July 7, 91 FR 41638): the primary text asserts state AI laws like Colorado’s are impliedly preempted — a legally untested theory — with only ~24 docket comments so far and none from notable filers; glossary/ftc-ai-marketing-enforcement carries the full mechanics and an early-August recheck flag. The first real-world agent-reliability numbers for ad work arrived: AD-Bench (225 real ad-analytics tasks) finds the best frontier model passes only 76.9% overall and 61.4% of complex tasks, failing mostly on tool operation, not reasoning — exactly the risk write-capable ad agents expose (marketing/ad-platform-agent-surfaces; Tencent-authored, so the fully independent measurement is still missing). And two negatives re-verified from primary sources: Gemini 3.5 Pro still hasn’t shipped (no API-changelog entry; July 17 remains rumor — comparisons/ai-tools-when-to-use) and Google’s Ads MCP is still strictly read-only (docs unchanged since June 29 — though community MCP servers do offer unofficial write access, a distinction now on the page).
Early-July freshness sweep — four open questions answered, folded into 7 existing pages ✅ (July 9) — The June 29 sweep left four items explicitly unresolved; a deep-research pass (every claim adversarially verified) closed all four. (1) Gemini 3.5 Pro did not ship. Google promised a June rollout; as of July 9 it’s still in limited preview with no GA date, benchmarks, or pricing — comparisons/ai-tools-when-to-use now says plainly: don’t build a routing decision around it yet (the July 17 launch date circulating is unconfirmed rumor). (2) The independent AI-trust data arrived — and it says LOW, not falling. Oxford’s Reuters Institute (48 markets) measured trust in AI-chatbot answers at just 20% globally (UK: 6%; among actual users: 44%), and Pew found 42% of US adults now use chatbots for information search. seo/geo-aeo-benchmarks-2026 carries the key distinction: independent sources confirm trust is low, but the widely-quoted decline (82%→54%) still rests on one vendor survey. (3) TikTok’s agent layer became a real product. On June 30 TikTok launched Agentic Hub — a live marketplace of AI Skills on its ads MCP server, with 14+ partners and full write access (agents can create and optimize campaigns). The sharp contrast: Google’s Ads MCP is strictly read-only. marketing/ad-platform-agent-surfaces now tables who actually lets agents operate an ad account. (4) Mastercard’s “Agent Pay for Machines” is real — but not shipped. The June 10 announcement (~32 partners spanning card networks, stablecoins, and agent infrastructure) is a validation-phase initiative with no availability date — glossary/agent-payment-protocols files it with the announcement-vs-GA caveat prominent. Two bonus items: the FTC’s July 1 proposed policy statement extends §5 deception to covertly-steered AI outputs and asserts state AI laws (Colorado’s) are federally preempted — proposed, not law, comment open through July 31 (glossary/ftc-ai-marketing-enforcement); and SparkToro’s 2026 clickstream reading puts US zero-click at 68.01% (cross-panel caveat carried) — seo/zero-click-strategy. Two circulating claims were refuted in verification and are marked do-not-cite: “SparkToro found AI Overviews cut CTR ~60%” (the study makes no such finding) and “Google has an agent-facing ad-management surface via MCP” (read-only ≠ control).
June 2026
Week of June 23 – 29
Late-June freshness sweep — 12 verified developments folded into 7 existing pages 🆕 (June 29) — A deep-research sweep (every claim adversarially verified, vendor sources flagged) pulled the genuinely-new late-June items into the pages that already own each topic, rather than spawning thin new pages. (1) Anthropic shipped Claude Fable 5 / Mythos 5 (June 9) — a new top tier above Opus 4.8, at premium pricing (~2× Opus); comparisons/ai-tools-when-to-use now flags it as a “reach for the hardest jobs” model while Opus 4.8 stays the workhorse. (2) automation/agentic-commerce gained the sell-side: Meta launched its Business Agent globally (June 3) — an AI that talks to customers across WhatsApp/IG/Messenger and recommends, qualifies leads, books, and (aspirationally) closes sales. (3) glossary/agent-payment-protocols picked up Adyen Agentic (a cross-protocol “universal translator” across UCP/AP2/ACP/Meta checkout) and Shopify’s removal of the UCP approval gate (self-serve agent commerce). (4) marketing/ad-platform-agent-surfaces added Meta Brand Memory (infers your brand from past ads to drive AI creative) and TikTok Search Hubs (paid top-of-search for brand pages). (5) glossary/performance-max now notes Google AI Max for Search exiting beta (the DSA successor — but the “DSA sunset date” claim was refuted and excluded). (6) seo/geo-aeo-benchmarks-2026 gained Google’s new Search Console generative-AI report (June 3, impressions-only), a Fractl survey showing AI-search trust falling (82%→54% YoY; Google still ~3× AI for product recs), and a GEO-at-scale preprint (78% of citations go to corporate sites; sentiment far more volatile than presence) — both flagged vendor-interested. (7) glossary/human-anchored-ai-multiplication picked up two peer-reviewed findings: AI-disclosure is double-edged (novelty↑ vs authenticity↓), and creative homogenization is fixable with diverse-persona deployment (not an inherent GenAI limit). Honest gap noted: no net-new FTC/EU AI-marketing enforcement this window — the field is quiet there since the May Cox Media settlement.
New page — Funnel Mechanics 📉 (June 26) — A first-principles funnel framework for 2026: marketing/funnel-mechanics. The core reframe — treat the funnel as a measurement instrument, not a literal path. Use AARRR (Acquisition/Activation/Retention/Referral/Revenue) to diagnose which stage is leaking (a single “1.4% conversion” number hides the actual problem), and know that real buyer behavior loops (McKinsey, Google’s “messy middle,” Gartner all showed this years before AI). Two things matter most: activation is the highest-leverage stage — and reaching value isn’t the same as completing a task; and the AI era is compressing the top of the funnel — independent Pew panel data shows people click a result in just 8% of searches when an AI summary appears (vs 15% without), ~60% of searches are now zero-click, and B2B buyers self-serve and arrive later and more informed. The honest verdict: “the funnel is dead” is overcorrected — the literal straight-line was always a simplification, but stage diagnosis still works; what changed is where the top of the funnel lives. Every figure is credibility-tiered (Pew strongest; Amplitude/Bain/Littledata flagged directional) and date-stamped.
New: the static-ad template system 🧩 (June 26) — A new page, marketing/static-ad-template-system, plus two skill reviews (tools/ad-template-foundry, tools/ad-template-apply), capture a reusable way to produce static ads. The core idea: a finished ad is angle × template × product — what to say (from audience research), how to lay it out (a brand-agnostic template), and whose brand (colors, fonts, product photo). One step deconstructs existing ads into a catalogued template library (compressing many similar ads into a few reusable layouts — “compress, don’t fork”); the other fills a chosen template with an angle and a brand and renders the ad in every aspect ratio, pulling the copy from the research rather than inventing it. The honest framing throughout: the library is a durable, compounding asset, but it lays out a message — it can’t supply one — and the specific templates stay experimental until real results promote them. Cross-linked to the existing reverse-engineering/casting cluster rather than duplicating it.
Four mid-2026 developments ingested 🆕 (June 26) — A freshness sweep (deep web research, every claim adversarially verified) pulled in four genuinely-new May–June 2026 items. (1) A new page, glossary/ftc-ai-marketing-enforcement, on the other AI-marketing legal risk — not failing to label AI, but claiming AI you don’t have. The anchor case: the FTC’s May 2026 Cox Media “Active Listening” settlement ($930K), where a service marketed as “AI that targets ads by listening through your microphone” used no voice data at all and just resold marked-up email lists. The lesson — “AI-washing” is deceptive under FTC §5. (2) A new page, marketing/ad-platform-agent-surfaces, on how TikTok, Reddit, and LinkedIn all shipped AI ad tooling in one tight window — TikTok an Ads MCP server (so AI agents can run ad accounts) plus AI video creative, Reddit four “Community Intelligence” ad products at Cannes, LinkedIn a Brand Kit for on-brand AI creative. The honest framing: these are announcements, not proven results. (3) The comparisons/ai-tools-when-to-use guide now reflects the mid-2026 frontier — Claude Opus 4.8 (top of the intelligence index, stronger browser-agent scores) and Gemini 3.5 Flash (cheap, fast, agentic) — with a clear caveat that the agentic benchmarks are vendor-reported, not independently audited. (4) glossary/ai-agent-behavior picked up peer-reviewed backing (MIT’s ABxLab, ICLR 2026): AI shopping agents are “strongly biased choosers” steerable by price, ratings, and nudges — confirming what was previously only a working-paper finding.
New page — Content Provenance & AI Disclosure 🏷️ (June 25) — A single home for a topic that was scattered across the wiki: glossary/content-provenance. It untangles four things people constantly mix up — C2PA / Content Credentials (a signed “nutrition label” in the file’s metadata, easily stripped by a screenshot), SynthID (an invisible watermark baked into the pixels, robust but Google-only), platform AI labels (TikTok and YouTube both say labeling doesn’t cut your reach), and disclosure law (the enforceable layer). The law is converging fast: the EU AI Act’s transparency rules and California’s AI Transparency Act both bite 2 August 2026, and New York’s synthetic-performer disclosure law is live 9 June 2026. The practical rule: disclose realistic AI content conspicuously, keep your claims substantiated, and don’t treat a platform label as legal cover. Every date and claim is primary-sourced and credibility-graded.
Correction: how AI-disclosure law actually works ⚖️ (June 25) — Tightened the AI-image disclosure guidance on marketing/ai-product-image-generation after a primary-source check. The earlier version (drawn from vendor summaries) overstated it — there was no 2026 update to the FTC Endorsement Guides (last revised 2023) and no FTC “AI unit.” The accurate picture: undisclosed AI imagery can be deceptive under FTC Act §5, AI deception is an active enforcement priority (Operation AI Comply, Sept 2024), and civil penalties attach to order or rule violations (like the 2024 Fake Reviews Rule), not automatically to a first finding. Better to be exactly right than impressively wrong.
Consistency fix: static vs. video product imagery ✅ (June 25) — A follow-up sweep caught two spots in marketing/ecommerce-to-tiktok-ai-pipeline that flatly said “AI can’t show your real product” — true for video, but no longer for static images. Scoped both to video and pointed readers to the new static-image page, so the wiki tells one consistent story: static stills are now reliably re-staged from your real product, while video and hard SKUs still need the composite method.
New page — AI Product Image Generation 🖼️ (June 25) — A dedicated guide to marketing/ai-product-image-generation: how to crawl your real product photos and re-stage them across endless on-brand scenes with reference / image-to-image models (Nano Banana Pro, Seedream 4.5, Flux 2), while keeping the product itself accurate. The honest line: this is now the default for ordinary static products (verify every output), but hand-compositing the real photo is still the fallback for reflective metal, glass, jewellery, fine on-pack text, and all video — and the model re-stages a product you give it, it doesn’t invent one. Covers the model line-up, what’s reliable vs what still breaks, feeding TikTok Photo-Mode carousels, and the new AI-disclosure rules. The static-image companion to the AI product video page.
Product-image generation: the fidelity rule moved 🖼️ (June 25) — Updated glossary/reference-image-conditioning and the TikTok pipeline page to reflect a real 2026 shift: for ordinary static products, you can now feed a real product photo into a reference/image-to-image model (Nano Banana Pro, Seedream 4.5, Flux 2) and generate endless on-brand scenes that keep the product accurate — reliable enough to be the default, as long as you verify each output. Hand-compositing the real photo is now the fallback for the hard cases (reflective metal, glass, jewellery, fine on-pack text, regulated goods) and all video. Also corrected a TikTok fact: the AI-content label does not reduce your reach (TikTok’s own words) — only unoriginal/bulk content is suppressed — and pinned the Sept-2025 enforcement dates.
Three measurement gap pages: ATT, Performance Max, and Multi-Touch Attribution 📊 (June 25) — A wiki lint surfaced three concepts the wiki leaned on constantly but never actually defined — so we wrote them. glossary/app-tracking-transparency is the root-cause page: Apple’s 2021 iOS change where most users declined tracking, which quietly reshaped all of paid media (it’s why attribution broke, why marketing-mix-modeling came back, and why “creative is the new targeting”). It also untangles ATT from Google’s third-party-cookie deprecation — the one that actually happened vs. the one that kept getting delayed and was ultimately scrapped. glossary/performance-max is Google’s answer to Meta Advantage+, written as a clean parallel — same AI-automation trade, same critiques (the black box, the branded-search cannibalization quantified in a 503-account study, attributed-return ≠ incremental lift). glossary/multi-touch-attribution completes the attribution trio with MMM and incrementality: what MTA is, why ATT and cookie loss demoted it from “source of truth” to a tactical in-platform signal, and why the honest 2026 answer is to triangulate all three rather than trust any one. Every figure is credibility-graded; refuted vendor stats flagged do-not-cite.
Consistency pass: AI UGC framing corrected ✅ (June 25) — Same lint reconciled two contradictions. The glossary/ai-ugc-ads page was retitled and rewritten to lead with the honest case — AI UGC’s advantage is cost and iteration speed, not performance (there’s no independent evidence it out-converts human UGC, and recognized undisclosed AI carries a trust penalty) — bringing it in line with the rest of the wiki. And the creative-fatigue numbers across marketing/channel-economics and marketing/andromeda-era-creative-strategy now agree: a single top ad can fade in ~5–7 days, but a diversified set rotates every 2–3 weeks — onset vs. cadence, an honest 1–3-week range.
Turning an e-commerce store into TikTok content with AI — the honest playbook 🎬 (June 23) — New page marketing/ecommerce-to-tiktok-ai-pipeline tackles the “Claude turns your store into viral TikToks” pitch and separates what’s real from the hype. The pipeline is real — scrape product pages and reviews, have an LLM write hooks and scripts, generate AI-UGC video, auto-caption, schedule — and the text layer is genuinely strong and ~10–50× cheaper than human UGC ($2–30/video vs ~$200). But the load-bearing finding is sobering: AI compresses everything except product truth and platform trust. Current AI video can’t render your actual product reliably (logos and packaging morph between frames), and there’s no independent evidence AI UGC out-converts human UGC — the one peer-reviewed study finds a trust penalty when viewers spot undisclosed AI. The page fact-checks the tool stack (Blotato’s Claude→TikTok MCP is real; “v0 for video” is a myth; CapCut has no API), debunks the “TikTok suppresses AI” fear (it suppresses unoriginal/spam content, not AI as a category), ranks the legal risks (unsubstantiated product claims under FTC §5 are the real danger; the avatar-lawsuit panic is overblown; note New York’s new synthetic-performer disclosure law from June 2026), and explains why small DTC brands run this end-to-end while big brands keep AI behind the scenes. Closes with a low-risk pilot. Every number is credibility-graded; refuted vendor stats are flagged do-not-cite.
Week of June 16 – 22
The Meta health-ad policy reference, made deployable ⚖️ (June 19) — New page marketing/meta-ad-policy turns “Meta keeps rejecting our weight-loss ads” into a checklist. It names the four rules that reject the most health/fitness/beauty/finance creative — negative self-perception (can’t make the viewer feel bad about their body to sell), personal attributes (the “you have X” trap, and 2026 enforcement now catches the implied “you,” not just the explicit one), before/after & idealized results, and close-up problem-area body shots — and gives the one reframe that reliably passes: spotlight the food/product/environment, agitate the food (hidden sugar/calories), never the viewer’s body. Includes the narrow before/after exception (fitness services showing activity are tolerated; weight-loss products showing physique change are not), the account-level risk for repeat offenders (lockdowns/bans — heaviest on young accounts), a nutrition-app brand-spec compliance block as a worked example, and a pre-flight checklist. Verified against Meta’s Transparency Center plus 2026 health-ad guides.
Advantage+ Creative and the brand-control problem ⚠️ (June 19) — New page glossary/advantage-plus-creative goes deep on Meta’s creative-automation feature — the one that applies “enhancements” to your ads, from harmless touch-ups to generative AI that creates whole new images. It’s on by default, and the page is honest about why that matters: in late 2025/early 2026 brands reported Meta swapping in AI-generated images they never made (one apparel brand’s top ad was replaced with an AI “grandmother in an armchair”), brand colors drifting, and disabled enhancements turning themselves back on. Includes the full enhancement list (what’s live vs. rolling out), Meta’s own performance claims, the AI-labeling policy, and why this is the concrete case for keeping a human-anchored creative you control. Also corrected an earlier page: the “+22% ROAS” figure is genuinely Meta’s own (from its Dec 2024 engineering blog), not a vendor number.
What Meta Advantage+ actually is — the AI ad-automation suite, demystified 🧭 (June 19) — New page glossary/advantage-plus clears up a term that gets thrown around loosely. Advantage+ isn’t one campaign type — it’s Meta’s umbrella brand for AI ad automation: end-to-end campaigns (Sales, App, Leads) plus single-step features (Audience, Creative, Placements). The big 2025–2026 story is that it became the enforced default — Advantage+ Shopping was renamed to Sales (Feb 2025) and the old manual-campaign APIs were deprecated (Oct 2025 → Q1 2026). The page is deliberately honest about evidence: every performance number Meta publishes is self-reported and dates to 2022 (the famous “32% ROAS” only holds with its specific test methodology attached), and it presents the incrementality critique and its counter-evidence rather than picking a side. Researched via a multi-source, adversarially-verified deep-research pass.
Meta Andromeda and the “build a template library, not single ads” playbook 🤖 (June 19) — Two new pages on the engine that quietly reshaped paid social. glossary/meta-andromeda explains Meta’s ad-retrieval engine (detailed Dec 2024): it picks a few thousand ads from tens of millions of candidates, and it was built to handle the creative-volume explosion that Advantage+ automation creates. The careful bit most write-ups get wrong — Meta didn’t say “targeting is commoditized / creative is the new targeting”; that’s the agency synthesis layered on top, and the page attributes it honestly. marketing/andromeda-era-creative-strategy is the playbook that follows: when the engine chooses from a huge pool, the unit of work becomes a reusable template library — 10–15 genuinely different concepts (archetype × angle), templated across 1:1/4:5/9:16, rotated every 2–3 weeks, with diversity at the formula level because near-duplicates get suppressed. Every circulating ROAS/lift number is flagged vendor-unverified.
A durable Meta ad placement reference — safe zones, dead zones, placement-aware copy 📐 (June 19) — New page marketing/meta-ad-placement-safe-zones codifies the craft of building creative that survives auto-placement: aspect ratios per placement, the pixel safe/dead zones (≈270px top and ≈672px bottom reserved on a 9:16 Reel), and the March 2026 Stories+Reels unification — one 9:16 master now clears all four vertical surfaces if you design to the Reels 35%-bottom reserve. Covers the classic failure mode (a 4:5 ad’s footer landing under the Reels like/comment rail) and the fix (keep load-bearing copy in the center-square intersection). Numbers drift as Meta changes its UI, so the page carries a re-verify-before-a-big-run note and links the official Meta sources.
The auto-generated-claims compliance gate ⚖️ (June 19) — automation/staged-compiler-pattern gained a subsection on a non-obvious gate: when an AI pipeline’s final stage produces finished copy (not just a brief), checkable and policy-sensitive claims need a mandatory human fact-check before any ad spend — and it’s easy to skip precisely because auto-generated copy looks finished and trustworthy. The human creative decision doesn’t disappear; it relocates to gate-selection and compliance.
The reverse-engineering cluster grew up — 18 pages 🌱→🌿 🌿 (June 17) — The full AI-creative-reverse-engineering cluster (pillar + 6 SEO clusters, 2 glossary entries, 8 Meta-Ad-Library Q&A pages, and the seo/meta-ad-library-api guide) moved from seedling to growing. The content was already deep and sourced; what was missing was navigation — several pages listed “related topics” as plain sentences instead of links. Those are now a real cross-linked graph, and the Meta Ad Library how-to pages (questions/how-to-access-meta-ad-library, questions/how-to-search-meta-ad-library, questions/how-far-back-meta-ad-library, questions/how-to-download-meta-ad-library-video) now form a navigable sub-cluster hubbing through the API guide and up to the capstone methodology. Easier for both readers and AI assistants to traverse the whole topic.
Seven new glossary entries for the AI-crawler vocabulary 📖 (June 16) — The glossary grew by seven plain-English definitions covering the bots and tools behind AI crawler access: glossary/ai-crawler, glossary/gptbot, glossary/ccbot, glossary/bytespider, glossary/waf, glossary/llms-txt, and glossary/pay-per-crawl. Each is a short, quotable definition (what it is, why it matters, how it works) cross-linked into the seo/ai-crawler-access cluster — the kind of canonical, definition-first page AI assistants like to cite. Glossary now at 82 entries.
Who can read your site, and what it costs you — a new AI-crawler-access page 🤖 (June 16) — New page seo/ai-crawler-access covers the access-control side of AI visibility: which AI bots to allow vs block. The taxonomy that matters — training bots (turn your content into model weights, give nothing back), retrieval/search bots (turn it into cited answers with referral links), and user-fetch bots (fire when a real person asks an AI to open your page — never block these). The load-bearing, under-known bit: robots.txt only asks compliant bots to stay away; a WAF actually blocks the request at the edge. Real enforcement is the firewall, not the txt file. Current 2026 user-agent strings for OpenAI, Anthropic, Google, Perplexity, and Meta, plus the infra-layer shift (Cloudflare now default-blocks AI crawlers and offers HTTP-402 pay-per-crawl). Pairs with the tools/ai-visibility-audit skill.
ChatGPT shopping is ~80% Google Shopping optimization 🛒 (June 16) — seo/agentic-search-optimization gained a section on a triangulated 2026 finding: ChatGPT’s product-discovery carousel is predominantly populated by querying Google Shopping’s organic index — three independent investigators converged on this (one study put it at 83% of carousel products in Google’s top 40 Shopping results; an independent replication found ~75% in the top 3). The practical upshot: there’s no separate “ChatGPT shopping” discipline — it’s Google Merchant Center feed quality + Google Shopping rank. And one feed won’t win five engines: Gemini reads Merchant Center directly, Perplexity and Copilot use Bing, Amazon Rufus uses Seller Central.
Paid-channel economics — the intent/cost map + owning what you build 📊 (June 16) — New page marketing/channel-economics lays out a vertical-agnostic framework: paid channels trade purchase intent against cost and access (search = highest intent and cost, and gatekept; social = cheap reach, weak intent; native + messenger = the “golden middle”). The structural answer to whatever wall you hit is owned-channel diversification — and the trigger differs by vertical: restricted verticals (iGaming/crypto) are blocked from paid (gambling ads are license- and certification-gated on both Google and Meta), while DTC/ecom hits a ceiling (CAC up 40–60% since 2023, Meta-dependence, 5–7-day creative fatigue). Same “build a channel you keep” thesis, different trigger. Benchmarks are labeled with how well-evidenced they are — including one claim (“20–60% registration-to-deposit”) flagged as refuted (the real figure is ~2–4%), and affiliate-network CPC numbers flagged as self-interested.
Week of June 9 – 15
A full AI-creative-reverse-engineering pillar published — 18 new pages 🧩 (June 14) — A drafted content cluster that had been staged for months went live in the wiki: an end-to-end methodology for reverse-engineering winning ads. The pillar, seo/ai-creative-reverse-engineering-complete-methodology, walks the whole flow — find a winning competitor ad → deconstruct its structural formula (10 layers) → cast that formula onto your product → ship 5 testable variations. Around it: six cluster pages (seo/surface-vs-structural-mimicry the core craft distinction, seo/ai-reverse-engineering-vs-creative-briefs, seo/ai-template-casting-workflow, seo/when-not-to-reverse-engineer, seo/ai-creative-team-structure, seo/ethics-of-reverse-engineering-ads), two glossary entries, and a Meta Ad Library how-to set (8 questions + the seo/meta-ad-library-api page). Honestly marked 🌱 seedlings — freshly published, not yet battle-tested — but a substantial, internally-linked body of practical creative-ops knowledge. The wiki grew 168 → 186 pages.
AI for knowledge management — grown into a current, sourced page 🧠 (June 13) — The wiki’s last thin April seedling, automation/knowledge-management, grew up (🌱→🌿). Now covers the 2026 landscape: retrieval got smarter (GraphRAG + agentic RAG), enterprise knowledge search became an agentic category (Glean’s reported $7.2B valuation on a permissions-aware knowledge graph; Guru’s “Verified Truth” human-approval model), and the personal side caught up (NotebookLM as RAG→agentic researcher with a ~2M-token context; Granola/Reflect/Mem for capture). The load-bearing honesty: the hard problems are permissions, verification, and curation discipline — not retrieval, RAG quality degrades past ~100 documents, and “compounding” is a maintenance discipline, not an emergent property (this wiki is the worked proof). Six dated sources added; cross-linked into the automation/glossary cluster. — The wiki’s last thin April seedling, automation/knowledge-management, grew up (🌱→🌿). Now covers the 2026 landscape: retrieval got smarter (GraphRAG + agentic RAG), enterprise knowledge search became an agentic category (Glean’s reported $7.2B valuation on a permissions-aware knowledge graph; Guru’s “Verified Truth” human-approval model), and the personal side caught up (NotebookLM as RAG→agentic researcher with a ~2M-token context; Granola/Reflect/Mem for capture). The load-bearing honesty: the hard problems are permissions, verification, and curation discipline — not retrieval, RAG quality degrades past ~100 documents, and “compounding” is a maintenance discipline, not an emergent property (this wiki is the worked proof). Six dated sources added; cross-linked into the automation/glossary cluster.
The brand email design system — the durable anchor for AI email production 🎨 (June 13) — Companion to the email production stack and the answer to “what do you actually need to make good AI emails”: a brand design system. New page marketing/email-design-system codifies the seven layers a reusable one needs — brand tokens, email anatomy, a tested HTML component library, formatting rules, flow maps, voice, and ESP technical notes — and frames it as the human-authored anchor AI multiplies against. With it, the collapsed production loop outputs your brand; without it, generic AI-template slop that costs trust. Build it by extracting from your best existing sends. It’s glossary/human-anchored-ai-multiplication for email and the durable layer of the automation/staged-compiler-pattern. — Companion to the email production stack and the answer to “what do you actually need to make good AI emails”: a brand design system. New page marketing/email-design-system codifies the seven layers a reusable one needs — brand tokens, email anatomy, a tested HTML component library, formatting rules, flow maps, voice, and ESP technical notes — and frames it as the human-authored anchor AI multiplies against. With it, the collapsed production loop outputs your brand; without it, generic AI-template slop that costs trust. Build it by extracting from your best existing sends. It’s glossary/human-anchored-ai-multiplication for email and the durable layer of the automation/staged-compiler-pattern.
The DNA-skincare article cluster, completed — the science behind the relief 🔬 (June 13) — A fourth public piece closes the cluster: “What Your Genes Actually Tell You About Your Skin (And What They Don’t).” The rigorous companion to the buyer-psychology article — the tests measure real genetic variants, but measurement isn’t prescription: the recommendations are mostly generic best-practice (sunscreen, retinoids, hydration), the field’s one robustly actionable finding is sun protection, and dermatologists see “no scientific rationale” for DNA-tailored routines today. The honest reconciliation: the real value is education and decision structure, not predictive precision — and overclaiming precision is what produces the accuracy-mismatch reviews that read as fraud.
AI email production, mapped — the email-channel companion to the video stack 📧 (June 13) — Email/Klaviyo was a confirmed gap: the wiki covered AI video production but not email. New page tools/ai-email-production-stack closes it. With a generative-image MCP wired in, email production collapses to a single-session loop — brand design system + MCP hero image + channel-ready HTML assembly — because the historical bottleneck (a designer or photoshoot for the hero) becomes one generate_image call. Same collapsed-loop shape as tools/ai-video-production-stack, and the MCP’s generation-only surface actually fits email (its one automatable visual step is generation — no placement/keyframe step like product video). But the loop only collapses production: copy/voice, brand fidelity, and deliverability stay human — and two of those are where a careless loop loses money. Verified counter-findings on the page: AI-looking imagery costs brand trust ~4:1 (eMarketer 2026), and image-heavy emails hurt deliverability (keep ≥400–500 chars of text, images under ~40%). The defense for both is the brand design system itself — the human anchor in glossary/human-anchored-ai-multiplication, a new channel. Scoped honestly: N=1 capability map, no open-rate claims.
Three public articles from a DNA-skincare research run — voice-of-customer, mined from public reviews 🧬 (June 13) — A category-research engagement for a DNA-skincare brand (anonymized) produced 100 graded reviews and ~580 archived community posts; three articles mine the parts worth publishing. “What DNA Skincare Customers Actually Buy” — the dominant reason people buy isn’t better skin, it’s decision relief (ending “product roulette”); the science sells as a permission structure, not a prediction, and the warmest reviews involve a human guidance layer. “Why Customers Call DNA Skincare a Scam” — the angriest reviews follow silence, not product failure: a missed lab turnaround plus an unreachable support channel flips “this is late” to “I’ve been robbed.” Communication and honest operations are the category’s unowned moat. “How to Map a Consumer Category Without a Research Panel” — the method behind the other two: mine public review corpora plus Reddit archives (via pullpush.io), keep everything verbatim, grade each finding by source credibility, and treat the gaps (near-zero organic discussion; a latent “it’s genetics” belief) as findings — honestly capped where the data is thin.
Human-anchored AI multiplication — a named framework, counter-findings included ⚓ (June 11) — The wiki-side companion to the Shoot Multiplication article, and the day’s most contrarian entry: glossary/human-anchored-ai-multiplication. The market uses AI as a creator; the data says it’s an amplifier — fully AI-generated ads underperform on brand effectiveness (Ipsos: −14%/−17%), AI video lags AI images, and the moat argument now has three academic layers: AI-assisted work measurably converges (Science Advances 2024), AI delegation produces “diversity collapse” (MIT), and autonomous model loops collapse onto ~12 clichéd motifs (Patterns 2025) — so the professional shoot is the one input competitors can’t prompt into existence. The page carries its counter-evidence in full: one working paper found pure-AI ads winning on short-run CTR; the ~500M-impression reconciliation is that AI creative wins only when it doesn’t look AI-made — perceived humanness is the real variable. Plus the calibrated Klarna case ($6M = stock-imagery replacement, self-reported) and the AI-slop fatigue data (Gartner: 49% say AI made content worse).
New public article: “Shoot Multiplication” 📸 (June 11) — Why the evidence says AI should multiply your professional shoot, not replace it. The platforms reward creative volume and concept diversity (TikTok: 5–7 creatives ≈ 1.5×, weekly refresh +10–12%; Meta penalizes near-duplicates by grouping them) — but fully AI-generated ads measurably underperform human-made (Ipsos May 2026: −14% short-term, −17% long-term), the trust penalty concentrates on AI-fabricated people, and “AI made it cheaper” is the one disclosed framing consumers punish. The conclusion is an architecture: AI works the scenes, formats, and concepts around real shoot assets — the one input competitors can’t prompt into existence. Honestly scoped: platform-doctrine-aligned, no audited public case study yet. At primores.org/articles/shoot-multiplication.
The two skills behind the cluster, reviewed as tools 🛠️ (June 11) — The strategy→production cluster documents the methods; two new tool reviews document the shipped skills that run them, both dogfooded on a real engagement: tools/target-audience-research (brand inputs → evidence-graded unit table; the first run caught a dead competitor the client’s own files listed as live) and tools/scenario-compiler (frozen unit → 18 named ad variants; the first batch ran 20 units → 360 variants and surfaced the hook-collision lesson). Tool catalog grows to 15.
New public article: “From Research to 360 Ad Variants” 📰 (June 11) — The day’s methodology cluster, packaged for readers: how a real engagement turned evidence-graded audience research into 360 testable ad variants via fully prescriptive briefs and one frozen contract — including the four-cell hook-collision story and the guardrails most teams miss (FTC testimonial rules, ad-licensed music, compliance flags that travel with the card). Published at primores.org/articles/research-to-360-ad-variants.
From research to 360 ad variants — a 3-page strategy→production methodology cluster 🏭 (June 11) — How do you run paid-social creative work when AI makes production nearly free? A real engagement (anonymized DTC client; 20 research units → 120 creatives → 360 named variants) yielded three connected pages. marketing/evidence-graded-audience-research replaces persona theater with units: every claim graded provided | researched | hypothesis with a source (importing intelligence-community evidence grading — apparently uncodified in marketing), audiences built from verbatim customer language, output as a [TA × JTBD × angle] table with stable IDs and transparent priority sub-scores. marketing/prescriptive-production-briefs is the downstream layer: when execution is cheap, the brief carries every decision — final copy, char-counted overlays, frame-level slot tables, an asset-source column routing each visual to the right pipeline; hooks live in slot 1 only, three per creative, each a different mechanism. The batch lesson worth the page alone: hook collisions happen at copy level even with centrally allocated concepts (one stat opened hooks in four test cells — contaminating the test read), so signature hooks are cell-exclusive and a similarity audit is mandatory. automation/staged-compiler-pattern names the architecture underneath: two compilers — research (expensive, durable, compounding) and production (cheap, disposable, regenerable) — joined by a frozen JSON contract, with human gates only at the points of irreversibility. None of it is glamorous AI; that’s the point.
The empty paid-social lane in DNA-personalized beauty — a worked market note 🧬 (June 11) — The competitor-analysis domain gets its second page, and its first worked market note: competitor-analysis/dna-beauty-paid-social-whitespace. As of June 2026, US paid-social DNA-skincare has one live paid funnel (test-only ClarityX) — the formulation+subscription model is unmarketed. But the lane is empty for a reason: GeneU wound down in 2018, ALLÉL’s domain now redirects to a retailer, and Know Beauty dropped its DNA angle — empty ≠ easy. The note documents the skepticism headwind (“snake oil with a digital signature”), the subscription churn line (cancellation talk clusters at $50–70/mo, so premium personalization must price-anchor against $150–400 treatments, never products), and the two clearest whitespaces (show-the-process reveal video; advertorial funnel). The verification pass earned its keep: the category’s favorite defense — “60% of skin aging is genetic” (Frontiers in Genetics 2025) — turns out to be co-authored by a DNA-skincare vendor’s founder, now flagged on-page as a disclosure-required citation.
Landing-page heroes get their own archetypes 🦸 (June 11) — The wiki’s creative archetypes (glossary/framing-archetype, glossary/focal-hierarchy) cover ads; landing-page heroes solve a different problem — earning the scroll. New page marketing/landing-page-hero-archetypes codifies three patterns from real B2B collateral builds: tension-triad (stacked pain lines → resolution tease → open-loop close — the mechanism reveals below the fold, because a hero that summarizes the mechanism closes the curiosity loop in place and kills the page), number-anchor (a specific stat carries the information scent), and honest-scope-upfront (a scope box under the hero that deliberately filters wrong-fit visitors — glossary/honest-assessment at the hero level). Selection runs on awareness level × page depth × scroll-force. Prior-art research found names for the ingredients (curiosity gap, information scent, exclusionary personas) but not for the assembled patterns — the triad, the scope box, and the depth-matched CTA ladder are codifiable territory.
What 1 billion data points say about AI search — 14 Ahrefs studies, calibrated 🔍 (June 9) — A deep-research pass over Ahrefs’ H1-2026 AI-search research run (14 studies) produced a new synthesis hub, seo/ahrefs-ai-search-studies-2026, plus two glossary entries — and corrected a standing wiki claim. Six findings: AI surfaces are separate discovery layers (Google AI Mode and AI Overviews reach the same answer 86% of the time but share only 13.7% of citations; only 38% of AIO citations rank in Google’s top 10; 28% of ChatGPT’s most-cited pages have zero Google visibility) — so “rank #1 → get cited” is broken. The CTR collapse is accelerating (position-1 −58%, up from −34.5% eight months earlier). The top visibility correlate is YouTube mentions (0.737) — above brand mentions, far above backlinks and Domain Rating; content volume correlates with almost nothing. “Best X” listicles are the single most-cited content format (43.8% of ChatGPT citations) — now a glossary entry, glossary/best-x-listicle. And being retrieved by an AI engine isn’t being cited — ChatGPT cites only ~50% of what it fetches, with readable titles and URL slugs predicting which half (glossary/retrieval-vs-citation). The honest correction: the wiki had called schema markup “essential / non-negotiable” for AI — but the best-designed study here (a quasi-experiment) found schema has no measurable effect on AI citations. Schema is now framed across the SEO pages as useful for traditional rich results and author-entity verification, not as an AI-visibility lever. All figures are graded as Tier 2 single-vendor measurement — large-scale and disclosed, but correlational and not peer-reviewed; the YouTube correlate in particular is flagged as possibly confounded.
AI product video without wrecking the product — a new 3-page production cluster 🎬 (June 9) — The wiki knew the strategy of AI video (marketing/ai-video-marketing) and the analysis of competitor creative (glossary/creative-reverse-engineering), but had nothing on actually making a high-fidelity AI product video. Three new pages, drawn from a real luxury-jewelry production and a cross-tool research pass, close that. The method — marketing/ai-product-video-fidelity — rests on two moves: composite the real product photo (let AI generate only the environment — never let the model render reflective metal or gemstones, which it hallucinates every time and which makes a luxury product look counterfeit), and first-last-frame keyframing instead of single-frame image-to-video (pin both ends so the model interpolates rather than inventing motion and drifting the product). Plus the counter-intuitive bits — i2v only animates forms already in the frame, so plan the motion into the keyframes; write filter-safe positive prompts; keep clips 3–5s and run multiple seeds — and honest limits (transparent gems, on-product text under motion, and any occlusion still break it). The companion tools/ai-video-production-stack is a dated capability→job map (Kling 3.0, Seedance 2.0, Nano Banana Pro, Veo, Luma Ray3, Midjourney, higgsfield as a hub), with one load-bearing finding for anyone trying to automate this: higgsfield’s MCP is generation-only — the placement, inpaint, and keyframe steps that make product video fidelity-safe are still UI-only. And glossary/reference-image-conditioning covers show-don’t-tell — controlling AI aesthetics by feeding reference images instead of prose, whose highest-leverage use is brand coherence: feed the client’s own glossary/distinctive-assets so output reads as them, not as generic AI.
Week of June 2 – 8
E-E-A-T gets a home — the lint’s one real gap, closed 📘 (June 7) — A full wiki lint came back clean (no contradictions, no broken links, no orphans, frontmatter 100%) but surfaced one genuine gap: E-E-A-T — the load-bearing AI-search quality framework — was referenced across ~17 pages with no page of its own. New glossary entry glossary/e-e-a-t fixes that: what the four letters mean (Experience, Expertise, Authoritativeness, Trustworthiness), why in 2026 it acts less like a soft ranking signal and more like a near-binary gate on whether AI engines cite you at all, and the four load-bearing 2026 signals (earned media, author-entity verification, Wikipedia, topical-authority depth). Calibrated honestly: the famous 96% / 3× / <4%-DA figures are vendor estimates, while the direction (earned third-party authority beats brand-owned content) has independent preprint support. Wired bidirectionally into the SEO/GEO cluster. The lint also fixed a stale index stat (domain pages 35 → 45).
Two open questions grow up 🌱→🌿 — a personal-AI-advisor reliability framework and a sharpened automation-eats-execution rule 🧭 (June 7) — A synthesis session (no new external source). questions/ai-as-personal-advisor gained a reliability framework that finally answers “when is AI advice trustworthy?” by wiring in glossary/appropriate-reliance: trust an advisor output only when it’s a high-validity, abundant-data task and your expertise/stakes don’t demand verification — because trust tracks the advisor’s confidence, not its accuracy, and AI advisors are uniformly confident. The durable Primores deliverable is the calibration discipline, not the tool stack. Separately, questions/automation-eats-execution-next-domains gained a candidate-scoring matrix across nine domains and a sharper rule: AI eats execution that’s high-volume and structured, but not execution that’s cumulative (brand-building) or relational (B2B sales) — so execution-layer compressibility, not the strategic layer, is what decides whether a domain bifurcates at all. Customer Success / RevOps flagged as the cleanest untested next probe.
SEO numbers get a reality check — Pew Research anchors the AI-Overview CTR collapse; vendor stats labeled as such 🔍 (June 7) — The wiki’s SEO pages carried striking headline numbers (64.82% zero-click, AI Overviews cut CTR up to 58%, 96% of AI citations from strong-E-E-A-T sources, brand mentions correlate 3× more than backlinks, Google AI Mode at 75M users) — all sourced to single SEO-vendor blogs. A deep-research pass (101 agents, 24 claims confirmed, 1 killed) asked which survive contact with primary measurement. The big win: the AI-Overview CTR-collapse claim is now anchored by a named research institution — the Pew Research Center July 2025 clickstream study (real browsing behavior, 68,879 searches) measured users clicking a traditional result on just 8% of pages with an AI summary vs 15% without (~47% reduction), and clicking the AI summary’s own citations on only 1% of visits. The honest part: the other four figures stay vendor-only — the exact 64.82% (Similarweb), the 58% (Ahrefs), the 96% E-E-A-T share, the 3× brand-mention correlation, and the 75M AI Mode users have no primary source, and seo/zero-click-strategy now says so in a “primary evidence vs vendor estimates” calibration table. A second institutional source (Reuters Institute’s Generative AI and News Report 2025) adds a useful honesty flag: people self-report clicking AI-overview links 33% of the time, but Pew measured 8% — trust the clickstream over the survey. The brand-authority direction got independent (preprint) support from Chen et al. 2025 (U. Toronto), though not the specific vendor coefficients. seo/geo-aeo-benchmarks-2026 and seo/ai-visibility calibrated to match. Net effect: the SEO domain’s single most-cited claim is now primary-anchored, and every remaining vendor number is labeled as one.
Win-loss follow-up — the two remaining gaps, honestly answered 🔬 (June 7) — A targeted second pass chased the two gaps the first win-loss ingest flagged. Both came back honest-negative on the headline number, which is the point. The 5–15pp win-rate lift is confirmed unmeasured — no peer-reviewed, quasi-experimental, or longitudinal study measures win-loss-program adoption against objective win rate (and a vendor “up to 50% lift” claim was refuted); glossary/win-loss-analysis now says so and sketches the clean study that would settle it. The “interview within 30–90 days” rule is mechanism-anchored but the exact window is unmeasured — and some survey-methodology evidence cuts against a naive “fresher is always better” (optimal recall is a forgetting-vs-telescoping tradeoff; one 10-year panel refuted time-based decay outright). The useful twist: the real threat to win-loss accuracy is buyer reconstruction and face-saving, not the calendar — so triangulating stated reasons against contemporaneous behavioral evidence matters more than interview speed, and reflective “why did you choose them?” questions can actually amplify biased memory (choice-supportive distortion). Net: the win-loss page is now as anchored as the evidence honestly allows — mechanism and practices grounded, both the magnitude and the exact timing marked unmeasured.
Win-loss analysis — academic anchors close the gap (honestly, halfway) 🎯 (June 7) — The longest-standing “needs rigorous backing” flag in the competitor-analysis cluster. A targeted deep-research pass (108 agents, 25 claims verified, 0 killed) asked whether the page’s load-bearing claim — win-loss is the only CI layer that reliably moves win rate, 5–15pp in 6 months — could be anchored in peer-reviewed research. Honest answer: no, not directly — no study measures win-loss → win rate causally, so the magnitude stays a practitioner figure, and the page now says so in its TL;DR, “Why it matters,” and Honest limits. But the mechanism and the practices anchor cleanly to tier-1 evidence, and glossary/win-loss-analysis gained a new Academic foundations section to carry them: win-loss is a structured retrospective review — the same family as the after-action review/debrief, which two independent meta-analyses show reliably improves performance (Tannenbaum & Cerasoli 2013, d=.67, ~20–25%; Keiser & Arthur 2021, d=.79). The “interview buyers, not reps” practice is backed by dyadic sales research (Endres et al. 2023 — rep self-perception explains only 4.7–7% of how customers actually perceive them; the buyer’s view predicts purchases better) resting on the seminal Nisbett & Wilson (1977) finding that people lack introspective access to their own decision processes. Two further anchors (Vieira et al. 2023; Kirca et al. 2005, market-orientation r=.32) place CI inside a real-but-conditional chain to performance — and caution that no single CI practice is a clean direct lever. Honest notes carried on-page: the “failing fast” construct (Friend et al. 2019) is forward-looking, not win-loss, and disqualifies itself as an anchor; and the “interview within 30–90 days” timing practice still lacks a verified decay-curve study. competitor-analysis/overview Layer 1 calibrated to match. No new pages.
Week of May 26 – June 1
CI follow-up — academic anchors for the competitive-intelligence cluster 🎯 (May 29) — A targeted deep-research pass closed the competitive-intelligence gap from the earlier ingest (and verified four candidate sources — all CONFIRMED real, none fabricated). New page glossary/ai-competitive-analysis — can you hand competitive/strategic analysis to an LLM? Yes, with three disciplines: aggregate many runs rather than trusting one biased single-shot (Doshi et al. 2025, Strategic Management Journal — aggregated LLM rankings match experts r=0.675), keep humans on the judgment dimensions where LLMs score worst (Csaszar et al. 2024, Strategy Science — innovation r=0.21, PMF r=0.24), and let AI triage signal rather than decide; plus the anchoring trap — using AI for both problem-framing and ideation cuts strategic quality 15 points (Wu et al. 2025, INSEAD RCT). Two existing pages moved from practitioner-backed to peer-reviewed-anchored: glossary/share-of-model gained the canonical GEO paper (Aggarwal et al. 2024, KDD), the brand-bias finding (LLMs favor global/luxury brands — share-of-model isn’t a neutral mirror), and the non-determinism result (measure as a distribution, not once); competitor-analysis/overview gained its theoretical roots (Day 1994 market sensing; Madureira 2023 CI construct). Honest notes: two monitoring papers’ method claims were refuted in verification and ingested as limitations only; and win-loss analysis still lacks rigorous academic backing — flagged for a future pass.
Deep-research ingest — AI-reliability + service-recovery evidence 🔬 (May 29) — Ran a multi-agent deep-research pass over trusted/academic sources (23 fetched, 25 claims adversarially verified, 2 refuted) and ingested the top results. Two new pages: glossary/appropriate-reliance — the goal with AI is calibrated reliance, not maximal; mere AI labeling triggers costly over-reliance (Klingbeil 2024) and suppresses critical thinking (Lee et al. 2025, CHI), yet experts under-rely, and disclosing AI use erodes trust −16–20% via reduced legitimacy (Schilke & Reimann 2025, 13 experiments) — reconciled via expertise × stakes moderation. And glossary/review-response-strategy — how to reply to reviews, backed by two ISR studies: responses lift review volume (not sentiment) via a third-party effect, detailed-for-negative/brief-for-positive (Chen et al. 2019), and tone must match the justice type violated — rational for procedural complaints, empathetic for interactional (Ravichandran & Deng 2023). Three existing pages gained fresh field evidence: glossary/jagged-frontier (Ju & Aral 2025 MIT field experiment + “diversity collapse”), glossary/ai-skill-leveling (Alibaba production RCT — skill-leveling and a top-performer decline), and questions/what-ai-tools-actually-deliver-roi (Copilot RCT: time saved ≠ output; Stanford HAI 2025 adoption-vs-ROI baseline). Honest note: the competitive-intelligence sources didn’t clear verification this pass — flagged for a follow-up.
Customer-perception cluster gets a framework hub + ROI question developed 🧭 (May 29) — A consolidation session. New framework page glossary/customer-perception-moments ties the two May-26 behavioral-evidence ingests (glossary/weekend-review-effect + glossary/ai-humor-forgiveness) into one navigable lens: customer perception crystallizes at three moments of judgment — the decision moment, the review-writing moment, and the failure-recovery moment — which form a feedback loop with reviews as the connective tissue. The transferable lesson is the moderator-flip meta-pattern: every headline behavioral finding (weekend negativity, humor-helps-forgiveness) comes with a context-dependent moderator (hedonic-vs-functional, severity, focal-vs-observer, review-count stage) that can reverse it — so the discipline is identify-your-moment-and-moderators-before-applying. glossary/honest-assessment is named as the unifying mechanism (visible imperfection out-converts polished perfection at every moment). Separately, the open question questions/what-ai-tools-actually-deliver-roi was developed with a two-axis ROI model (frontier position × error cost) — highest ROI is the unglamorous inside-frontier/low-error-cost quadrant; the −19pp danger zone is outside-frontier/high-error-cost — plus a function-by-function map and free-vs-paid reasoning; upgraded 🌱→🌿. Full lint run (no contradictions, no orphans, clean frontmatter); one genuine missing cross-link fixed and an llms.txt duplicate-line bug corrected. No new external source this session — the value was in consolidation, development, and health.
Weekend review-ops cluster — Science Says newsletter ingest with full cross-reference integration ⭐ (May 26, second session) — Second Science Says newsletter same day, user direction: integrate all the cross-referenced findings (not just the primary study) — saved as new feedback discipline. New page glossary/weekend-review-effect (~4,500 words) covering Bayerl, Schoenmueller, Goldenberg, Stahl 2026 (Journal of Marketing Research 63(2), n=400M reviews / 33 platforms / field experiment n=11,667): weekend reviews average -3% 5-star share and +6% 1-3 star share vs weekdays. The “Eleanor Rigby” mechanism: weekend reviewers self-select into a more socially-isolated population (fewer sociality-related words in reviews). Three caveats the newsletter missed, surfaced during the user’s “validate claims” pass: (1) hedonic-product counter-finding — effect reverses for entertainment/food/travel categories (2023 ScienceDirect, n=588K); (2) industry response-rate data conflicts — Saturdays are among highest-volume send-days per Bazaarvoice/Yotpo/PowerReviews — academic paper measures star rating outcome, industry measures response rate; (3) 0.04-star magnitude is below noise floor for products with 100+ reviews — effect matters operationally for low-review-count items + weekend-busy businesses. The wiki integrates the timing finding with three cross-referenced review-ops levers (first-review anchoring + incentive-positivity transfer per Woolley & Sharif 2021 +83.4% effect + display-order effects with 52%-prefer-mix trust nuance) into a four-lever practitioner playbook. Retrofit to glossary/ai-humor-forgiveness with two adjacent service-recovery tactics the original ingest missed (thanks-not-sorry pattern from JTTM 2022; chatbot interjections). Back-links to honest-assessment + ai-humor-forgiveness. Source archived to raw/articles/_ingested_2026-05-26_dont-ask-for-reviews-on-weekends.eml. New feedback discipline saved: when ingesting curated newsletters, follow the cross-referenced findings — newsletter author already built the cluster, wiki should inherit it.
AI humor-forgiveness — Science Says newsletter ingest with counter-finding integration 😅 (May 26) — User flagged a Science Says newsletter summarizing Xie et al. 2025 (Journal of Business Research, n=1,919) on AI agent service-failure humor and forgiveness. Verdict: worth ingesting. New page glossary/ai-humor-forgiveness (~4,500 words) covers: self-deprecating humor produces +47.8% forgiveness uplift vs no humor on low-severity failures (and +25.6% on after-sales) — outperforming positive humor by ~10pp. Two load-bearing boundary conditions the newsletter didn’t fully surface: (1) severity gate — effect vanishes (not weakens) for high-severity failures like refund refusal; (2) focal-customer gate — Honora et al. 2025 (J. Business Ethics) counter-finding shows humor directed at the burned customer reads as sarcasm, reduces perceived company morality, and forgives less. Independent corroboration from Nature Sci Reports 2025 (n=780; hedonic-motivation moderator + perceived-warmth mechanism). The wiki integrates the Hosanagar forgiveness-asymmetry quote — “People are not as forgiving of AI errors as they are of human errors” — into glossary/agent-adoption-frictions as the trust-friction repair cost. Substantial extensions to glossary/honest-assessment (self-deprecation as conversational-layer extension of the honest-assessment mechanism) and marketing/ai-human-voice-prompting (humor-on-failure as recovery-side seventh-technique). Back-links wired per schema discipline. Source archived to raw/articles/_ingested_2026-05-26_humor-makes-it-easier-to-forgive-AI-mistakes.eml. The wiki value-add: the newsletter missed the focal-customer counter-finding; the wiki page integrates it.
Week of May 19 – 25
Gemini Omni — same-day capture of Google I/O 2026 launch 🎬 (May 20, second session) — Google I/O 2026 was May 19; Gemini Omni is Google’s first any-to-any multimodal model unifying video/image/audio/text generation with Gemini’s reasoning baked in. Fuses Veo (video) + Nano Banana (image editing) + Project Genie (world simulation) under one architecture. Ships today in Gemini app + Google Flow for AI Plus/Pro/Ultra subscribers ($20/$30/$100/mo); YouTube Shorts + YouTube Create this week; Vertex AI / Gemini API / Agent Platform API in coming weeks. World-model physics understanding (gravity, fluid dynamics, collision behavior) inherited from DeepMind Project Genie — predicts outcomes from physical intuition, not running a physics simulation. New page: tools/gemini-omni (~3,500 words covering capabilities, pricing, competitive context, marketing/ad applications, honest limits). Distinct strengths versus Sora 2: prompt adherence on multi-clause briefs + text rendering reliability — both load-bearing for advertising. The 2026 split: Omni is the publisher’s tool (efficient, distribution-embedded, ad-scale variant generation); Sora 2 is the artist’s tool (cinematic, social, audio-sophisticated). Substantial extensions to marketing/ai-video-marketing (publisher-vs-artist-tool framing), comparisons/ai-tools-when-to-use (Gemini side refresh with two new video-generation rows), glossary/creative-reverse-engineering (Omni as the third vision-LLM closing the analysis-to-generation loop). Back-links wired to 5 target pages per the schema discipline. Same-day capture cadence matches the May 7 Claude Managed Agents release-wave ingest.
Competitor-analysis five-layer methodology now fully backed 🎯 (May 20) — Two new glossary entries close the remaining layer gaps in the competitor-analysis cluster. The May 18 pillar rebuild named five operational layers but only three (win-loss, battlecards, share-of-model) had dedicated practitioner glossary entries; Layers 2 and 5 were described in pillar prose without their own deep-dive pages. New: glossary/continuous-monitoring (~3,500 words on the 2026 CI discipline replacing the quarterly competitive-landscape deck — Crayon tracks 100+ signal types per competitor; the bottleneck has shifted from collection to signal-to-noise triage and decision routing; Klue’s 2026 Compete Agent + Deal Tips push competitive guidance to reps in near-real-time; tooling tiers from free Meta Ad Library + Changedetection.io stack to $20K–$100K+/yr enterprise platforms; three operational cadences — real-time + weekly + monthly — needed in parallel) and glossary/creative-reverse-engineering (~3,500 words on the systematic discipline distinct from the formula-vs-skin framework — pattern extraction across volume is the load-bearing move; 10–20 ads per pass minimum; the vision-LLM hybrid stack of Claude for copy + GPT-4o for visuals + Foreplay/Atria/Motion for collection; the formula-vs-skin distinction is the legal safety line). The competitor-analysis cluster is the wiki’s first domain with methodology pillar + every operational layer backed by a dedicated practitioner glossary entry. Back-links wired to 9 target pages per the schema discipline; index Stats block fixed (was stale: said ~126 pages / 45 glossary; actual 144 / 62).
Week of May 12 – 18
Prompt caching glossary entry 💾 (May 18, fourth session — late evening) — Tier 2 #5 from the morning gap analysis, the last remaining priority. New page glossary/prompt-caching (~2,500 words) covers the production-cost-optimization layer that the agent-engineering cluster was missing. Three distinct caching mechanisms (often conflated in practice): prompt cache (vendor-level prefix reuse — Anthropic cache_control markers at 5-min default / 1-hour extended TTL; OpenAI automatic on GPT-4o/GPT-5.x; cache reads at 10% of base price = 90% input-token discount), semantic cache (vector-similarity response reuse via Redis/GPTCache — typical 50%+ cost reduction with repetitive query patterns), KV cache (model-internal, not practitioner-facing). The Anthropic February 2026 TTL change (60-min → 5-min default, increasing many production costs 30-60% overnight) documented as a real failure mode. The 70-80% combined-strategy cost-reduction headline numbers contextualized as upper-bound. Proactive cache warming flagged as critical under-applied practice. Distinct from but adjacent to glossary/agentic-memory — prompt caching is per-request cost optimization; memory is cross-session persistence. Closes the May 18 gap-analysis cycle (5 of 5 priorities done in one day).
SEO/GEO refresh against 2026 zero-click data 🔍 (May 18, third session) — Tier 2 #4 from the morning gap analysis. The SEO domain was last touched April 2026 and pre-dated the most material 2026 findings: Google AI Mode hitting 75M daily users (92-94% zero-click), E-E-A-T as binary AI visibility filter (96% of AI Overview citations from strong-E-E-A-T sources), brand-mentions-3x-stronger-than-backlinks correlation (0.664 vs. 0.218), Domain Authority predicting less than 4% of AI citations, and earned-media citation value lasting 18-24 months. New page seo/zero-click-strategy (~5,000 words) covers the strategic operating model for a 64.82%-zero-click world — brand-and-visibility-first instead of traffic-first, the dual mandate (traditional rankings + AI engine citations), and the hardest part (admitting traffic-first KPIs are broken). Substantial extensions to seo/geo-aeo-benchmarks-2026, seo/ai-visibility, and seo/agentic-search-optimization absorb the May 2026 data shifts. Back-links wired to all 6 existing SEO pages per the schema discipline. PR strategy is now SEO strategy.
Competitor-analysis domain build-out 🎯 (May 18, second session) — Tier 1 #3 from the May 18 deep research, closing the longest-flagged structural gap (competitor-analysis as a one-page domain). User framing was explicit: methodology, not tool reviews. Four pages shipped — rebuilt pillar competitor-analysis/overview (~4,500 words on the 5-layer operational methodology: win-loss analysis + continuous monitoring + battlecards + share of model + creative reverse engineering, tied to specific decision triggers; SWOT/Porter’s critique; AI compresses execution while strategy stays human), plus three named-framework glossary entries: glossary/win-loss-analysis (the only CI layer that reliably moves win rate, 5-15pp improvement in mature programs; Klue/Crayon 8-practice convergence; interview buyers directly, not sales reps), glossary/battlecards (2026 living-battlecard evolution from static PDFs; modular + AI-assisted + role-specific + governance metadata; one-screen rule; success = win-rate-against-named-competitor), glossary/share-of-model (the AI-search competitive dimension that didn’t exist in 2024; ChatGPT ~79% of generative AI traffic; sector concentration extreme — Apple 54.38% mention share in consumer electronics — specialization is the only path for non-dominant brands). The wiki’s longest-flagged structural gap is now closed methodology-first. Back-links wired to 8 target pages per the schema discipline.
Agent payment protocols + Claude Managed Agents May release wave 💰 (May 18, morning session) — Monday session after weekend gap. Executed top two ingest priorities from the May 18 deep research: agent-to-agent payment infrastructure (which shipped April–May 2026) and Claude Managed Agents feature wave (May 7).
- New page: glossary/agent-payment-protocols — Comprehensive treatment of the four-protocol stack constituting production agentic-commerce infrastructure: AP2 (Google authorization, JSON-LD W3C Verifiable Credentials, September 16 2025), x402 (Coinbase + Cloudflare HTTP-native stablecoin protocol, ~$600M annualized volume by March 2026), UCP (Google + Shopify + retailer coalition commerce schema), Visa TAP (agent identity at Cloudflare edge, October 2025). AWS Bedrock AgentCore Payments shipped May 7, 2026 with native x402 support and 10,000+ x402 Bazaar endpoints discoverable by agents at runtime. OpenAI killed Instant Checkout March 2026 — the aggregator-with-own-checkout approach lost; market consolidated to Amazon closed-loop vs. Google-coalition open-protocol. Anthropic Project Deal (April 2026) documented the “agent quality gap” — users represented by less capable models get objectively worse outcomes and don’t notice; model capability dominated prompt framing. US legal void on consumer dispute rights covered explicitly.
- Extended: automation/agentic-commerce — Originally written April 14, the page described the agentic-commerce category but had zero coverage of the infrastructure that shipped between September 2025 and May 2026. Added major new sections on the four-protocol stack, the closed-loop vs. open-protocol dynamic, the agent quality gap, and the legal void. The page now reflects May 18 infrastructure reality.
- Extended: tools/claude-managed-agents — May 7 2026 release wave. Outcomes promoted from Research Preview to Public Beta. Multi-Agent Orchestration promoted to Public Beta. New Dreaming section covering between-session memory consolidation (research preview), the two operating modes (automatic vs review-before-landing), and the cluster connection to glossary/agentic-memory. Cowork enterprise features (RBAC, group spend limits, agent view in Claude Code). 10 finance agent templates + 20+ legal MCP connectors as the verticalization move.
- Extended: glossary/agentic-memory — Added a substantial new Anthropic Dreaming section. The cluster prediction (memory must be engineered, not built-in) is being validated and partly automated by vendor releases within days of the wiki entry. Dreaming shipped 8 days after this glossary entry was created and consolidates episodic memory into procedural memory at the platform layer.
- Architectural payoff: the agentic-commerce cluster now has all four sides covered (category + agent biases + user-side trust + infrastructure). Plus the agent-quality-gap finding cuts across all four as a new asymmetry dimension. Back-links wired to 9 target pages per the schema discipline.
Marketing analytics gap closed 📊 (May 15, after foundational glossary) — Tier 1 #2 from the May 15 gap analysis. The wiki had zero coverage of marketing analytics — MMM, incrementality, attribution-in-2026, cohort/LTV — despite this being core CMO-targeted content. Four new pages close the gap: marketing/marketing-analytics-in-2026 (comprehensive pillar — the cookieless attribution stack: dual-model operating norm with MMM + multi-touch + AI reconciliation + data clean rooms + platform AI like Advantage+/PMax; cohort/LTV:CAC as capital-efficiency layer underneath), plus three named-framework glossary entries: glossary/marketing-mix-modeling (top-down statistical attribution, no cookies needed; +212% adoption since 2023; Google Meridian + Scenario Planner + Meta Robyn democratized the methodology), glossary/incrementality-testing (causal-validation layer; three test designs; 2026 standard 10-20% holdout + synthetic controls; “fewer tests that materially change decisions”), glossary/cohort-analysis (the shape of the retention curve > M12 endpoint; M0-M3 onboarding cliff; surviving-cohort LTV (M3+) + NRR + CAC payback > aggregate LTV; 2026 LTV:CAC benchmarks B2B SaaS 3.2:1, DTC subscription 4.1:1). The cluster forms the marketing-analytics-in-2026 operating stack and provides citation anchors for future CMO-facing pages.
Foundational AI glossary gap closed 📚 (May 15, after human-voice page) — Tier 1 #4 from the May 15 gap analysis. Five new glossary entries: glossary/hallucination (signature LLM failure mode, structural cause = probabilistic next-token prediction without internal “I don’t know”), glossary/agentic-memory (4-layer architecture — working/episodic/semantic/procedural; memory is engineered, not built-in), glossary/tool-use (the capability that constitutes the agent category; basic primitive under all 2026 agentic work), glossary/guardrails (production safety layer paired with tool-use; six basic categories per Pimenov; persuasion-based bypass risk per Cialdini-on-AI 28K-prompt study), glossary/embeddings (numerical representations preserving semantic similarity; foundational layer under semantic search and RAG). Together: any future page that wants to cite these has a glossary anchor rather than a parenthetical explanation. The agent-engineering cluster (jagged-frontier → agent-engineering → tool-use → guardrails → agentic-memory → ai-agent-behavior → agent-adoption-frictions) is now self-contained — no out-of-page hops needed for foundational vocabulary.
New page: AI human voice for social + outreach 🎙️ (May 15) — Deep research + comprehensive new page. User flagged the human-voice question as “very important” and asked for deep research before building. Eight WebSearch passes across detection science (Nature 2025 stylometry, F1 ≈ 0.94 combined; em-dash 3.28× human rate; “absent lived experience” as the load-bearing fingerprint), empirical platform performance (cold-email deliverability AI 71%/8% spam-flag vs human 86%/3%; AI follow-up dies at 1.6 exchanges vs human 3.7), platform algorithms (LinkedIn 360Brew 150B-parameter LLM deployed March 12 2026, X Grok ranking, TikTok C2PA but script-AI exempt, EU AI Act August 2 2026 deadline), and prompting techniques (few-shot 2–5 optimal; Ruben Hassid Taste Interviewer 100-question approach; banned-word 2026 consensus; SuperWhisper voice-from-audio as sixth technique; Adore Me + Unilever case studies as 80/95–5/20 hybrid ratio anchors). The new page marketing/ai-human-voice-prompting is the generation-side complement to marketing/ai-tells-in-sales-copy’s editing-side discipline. Together they form a three-page cluster with marketing/brand-voice-skills-guide covering generation→editing→trust at distinct layers. The deepest single-topic cluster the wiki has built.
Cluster lint 🧹 (May 14) — Focused lint after yesterday’s heavy-creation day (4 new pages + 6 substantial extensions + ~25 new cross-links). Five dimensions checked. Wikilink validity 100% (38 unique targets, all resolve). Frontmatter compliance 100%. Substantive consistency 100% across five cross-cluster checks (jagged-frontier vs jagged-intelligence distinguished without unit-confusion; Karpathy’s >10x flagged as directional vs Dell’Acqua’s measured +12.2% properly bounded; Goldilocks-autonomy aligned across pages; Pirolli-Card invocations consistent across the validation-before-volume triangle; user-side/agent-side terminology symmetric between agent-adoption-frictions and ai-agent-behavior). Significant finding: bidirectional cross-link asymmetry. Yesterday’s new pages had rich outbound Related sections but target pages mostly lacked back-links. 16 back-links added across 14 target files to close the structural gaps. Second finding: llms.txt stats stale (same recurring failure mode the May 12 lint flagged). Refreshed page/section counts (marketing 14→16, glossary 45→47), bumped Stats section, added entries for the 4 new pages. Both recurring failure modes (ship-day back-link asymmetry + llms.txt drift after index updates) are now visible enough to deserve CLAUDE.md INGEST-workflow improvements.
AI-tells page — two-principle frame extension 🎯 (May 13, immediately after the page shipped) — Strategist task 2026-05-13-ai-tells-page-proactive-reader-motivation argued that audience-mode REVIEW is necessary but not sufficient: argument-level mismatches (wrong structural pain for the vertical) survive any amount of voice polish and only get caught when reader-motivation is modeled BEFORE drafting. The page now opens with a “Two principles for client-facing copy” section: (1) don’t sound like AI [existing catalog], (2) model the reader’s motivation before drafting [new section]. New “Model the reader before drafting” section covers the DTC-vs-iGaming structural-pain contrast and the iGaming reach-scarcity reframe as the canonical reference case — first-pass copy used DTC-style “rented FTDs vs. owned audience” framing, which implicitly assumes a working paid channel; iGaming doesn’t have that (paid bans on Meta/Google/Telegram official). The fix was argument-level, not voice-level. The verbatim user-articulation trigger now cited in Sources: “The key is the ability not to sound like AI and also look from the perspective of the reader.”
New page: AI tells in sales copy ✍️ (May 13, after Telegram extension) — Strategist task 2026-05-13-ai-tells-in-sales-copy requested a wiki home for the eleven-pattern catalog + audience-mode review discipline + CMO-believability score heuristic surfaced during a sales-page audit that moved CMO-believability from ~6.5/10 to 9/10 in one cycle. New page marketing/ai-tells-in-sales-copy codifies all three layers. The eleven tells include factual overreach for rhythm (the most damaging — one false claim discredits the whole doc), strategist-memo voice (meta-commentary instead of substance), em-dash overuse, coined-term over-use, parallel-construction overdensity, and seven more. The page is the negative trust-signal counterpart to glossary/honest-assessment (positive trust signal) — both depend on the same mechanism: audience reads writer’s judgment through the prose. Back-links added to honest-assessment and brand-voice-skills-guide (which is the LLM-instruction-side complement: what to sound like vs. what not to sound like).
Telegram page — operational-unlock section 🗺️ (May 13, after Karpathy ingest) — Strategist task requested an operational extension to the Telegram page; pure reorganization of existing sources into methodology. Added “Operational unlock: keyword × region discovery” (placed between Audience footprint and iGaming sections). The new section names the operational constraint that shapes everything: Telegram has no platform-level discovery surface, so entry is per-group rather than per-impression, and the unlock is a 2-4 week scout phase using TGStat + vertical indexes (Data40), $50 minimum / $350 proper-read per channel test (PropellerAds 2025), then scale on validated channels only. Key architectural move: this is marketing/discovery-before-scale applied to channel selection rather than to content patterns — same Pirolli-Card Independence of Inclusion from Encounter Rate math. Cross-link added in both directions. The Telegram page now sits at the intersection of two named frameworks (channel-fit-is-geographic + discovery-before-scale).
Karpathy Sequoia AI Ascent 2026 ingest — agent-engineering cluster completed 🎤 (May 13, later in session) — User flagged the YouTube talk From Vibe Coding to Agentic Engineering and a Pimenov writeup of the same content. Triangulated with two additional independent summaries (Travis Media, AI Agents Simplified) to verify quotes and framing.
- New page: glossary/agent-engineering — Karpathy’s professional discipline of coordinating AI agents reliably and safely, framed as the complement to vibe-coding rather than its successor. Load-bearing quotes: “Vibe coding raises the floor. Agentic engineering raises the ceiling.” and “You can outsource your thinking, but you can’t outsource your understanding.” Covers the Software 1.0/2.0/3.0 framing (“LLM became the computer, prompt became the program”), the directional >10x developer claim, the “neural networks as operating systems” architectural prediction (OpenAI Codex + MCP as early evidence), and the human skills that stay distinctly human (taste, architectural thinking, oversight, contextual understanding). Most important move: explicitly names the connection between Karpathy’s “jagged intelligence” (model-side) and Dell’Acqua’s “jagged frontier” (human-side) as the same structural insight viewed from two sides.
- Extended: glossary/vibe-coding — Added a “May 2026 update: the floor-vs-ceiling distinction” section. The 2025 vibe-coding discourse had collapsed two distinct disciplines into one phrase; Karpathy’s update separates them. Both are Software 3.0; they sit at different points on the floor-to-ceiling axis. Added the Software 1.0/2.0/3.0 table.
- Cross-linked: glossary/jagged-frontier — Added a “Karpathy’s ‘jagged intelligence’ — the model-side cousin” section with side-by-side comparison table. Unifying mechanism: Klein-Kahneman’s high-validity-environment + rapid-feedback conditions (glossary/recognition-primed-decision) predict both jaggedness shapes. The wiki now has a tight three-page cluster — jagged-frontier (human-side empirical anchor) + agent-engineering (model-side framing + engineering response) + vibe-coding (accessibility-side framing) — all referencing the same underlying mechanism and the same cross-domain thesis (glossary/automation-eats-execution).
Raw queue drained + two new pages + cluster extension 📥 (May 13) — Session goal was to clear the 8-article raw/ queue, ingest the strategist Telegram task, and refresh two stale seedlings. Useful finding: most raw/ items were already ingested in earlier sessions but had stayed in the queue because there’s no archive-on-ingest convention. Filed as a maintenance-protocol improvement.
- New page: glossary/agent-adoption-frictions — Wharton × Science Says AI Agent Adoption Blueprint (April 2026). Three psychological frictions block agent adoption: perceived competence, trust, delegation of control. The barrier is psychology, not tech. Draws on 700,000+ employees surveyed across Google, ServiceNow, Wolters Kluwer, Workato, Concentrix, Zapier. This is the user-side counterpart to glossary/ai-agent-behavior: where Columbia/Yale documents what agents choose, Wharton documents whether users let agents choose at all. Connects to glossary/jagged-frontier and glossary/recognition-primed-decision — the user-side calibration response to AI capability asymmetry.
- New page: marketing/telegram-marketing-channel — Built from the May 7 strategist research task. Telegram has 1B+ MAU (Durov, March 2025) and is the dominant marketing channel for iGaming globally and for fashion DTC in Russia/CIS/Iran/MENA — but NOT for Western Web3 fashion (that’s Discord, empirically: RTFKT, BAYC × BAPE, SYKY, Lacoste UNDW3 all run Discord). The page names the generalizable framework-level insight: channel-fit is geographic before categorical. Connects to glossary/super-niche as the channel-selection application of specificity-beats-generality. Also covers the iGaming pricing reality (official Telegram Ads bans gambling, so iGaming runs on third-party networks; affiliate channels eclipse official brand channels) and the Trezor channel-deactivation case showing operational tax can cancel even strong-fit channels.
- Extended: glossary/ai-agent-behavior — Added the underlying Allouah et al. (Columbia + Yale, Working Paper Dec 2025) study with full researcher names and study scale (1,000 experiments × 8 categories). Added model-improvement curves on the “obvious better deal” test: Claude Sonnet 3.5 → Opus 4.5 failure rate 63.7% → 4.3%; GPT-4o → GPT-5.1 25.8% → 1.0%; Gemini 2.0 → 2.5 2.8% → 0%. Added the position-bias reversal finding (GPT-4.1 favored top-left; GPT-5.1 reversed). Added the 0.1-rating sensitivity. Added the Cialdini × Wharton 28,000-prompt study (compliance 33.3% → 72%) — implication: product pages optimized for human Cialdini-persuasion are also optimized for agent persuasion. Added the ZDNet BlancPottery real-world replication. Sources section restructured with full citation depth.
- Refreshed: tools/obsidian — Was 32 days stale. Bumped 🌱 → 🌿. Added a “Five-week field report” section documenting actual usage of Obsidian as the wiki’s IDE: graph view as health monitor (made zero-orphans lint enforceable rather than aspirational), wikilink autocomplete as link-rot prevention, real-time edit preview as LLM-output calibration. Free tier sufficient; git handles sync.
- Refreshed: experiments/ai-visibility-ecommerce — Was 27 days stale. Bumped 🌱 → 🌿. Added “Status update (2026-05-13)” section. Honest finding: next-steps remain open (no new field work) but findings haven’t lost relevance — they gained relevance. The pigu.lt WAF block on AI bots was an “interesting SEO oddity” in April; with 25-30% of US online purchases now reaching AI agents (Columbia/Yale), the same technical fact has become a load-bearing commerce assumption violation. Cross-linked to ai-agent-behavior, jagged-frontier, agentic-commerce, agent-adoption-frictions.
- Architectural finding for next lint: The “raw/articles already ingested but still appearing in queue” pattern wasted real session time. The fix is an archive-on-ingest convention (
_ingested_YYYY-MM-DD_filename prefix, parallel to the drafts/ folder’s_archived_convention). To be applied to this session’s processed articles immediately and proposed as an INGEST workflow update to CLAUDE.md.
Weekly lint + four targeted fixes 🧹 (May 12) — First session in seven days; lint confirmed wiki health is strong (100% frontmatter compliance, zero orphans, zero private-path leakage, zero pages stale past 30 days). All four fixable findings closed in the same session.
- Refreshed
llms.txt— was last updated April 23, claimed 68 pages / 21 glossary. Now reflects current state (125 pages, 45 glossary, 10 academic foundations), adds the May framework cluster (automation-eats-execution, jagged-frontier, ai-skill-leveling, advisor-strategy), and includes a new “How to use this wiki as an AI assistant” guidance section. - Fixed a broken-wikilink in strategist-pattern pointing at the repo-root
CLAUDE.md(which lives outside the wiki folder, so wikilink resolution would render as broken). Now plain prose with a contextual link. - Wired five missing cross-links to address low inbound-density pages: cases/intercom-fin-support now linked from glossary/automation-eats-execution as the customer-service domain anchor; cases/binti-social-services now linked from glossary/ai-skill-leveling as the field-data analog for skill leveling in regulated documentation work; seo/new-site-ranking now linked from both glossary/super-niche and glossary/topical-authority; experiments/overview gained a new “Niche-Discovery Experiments” section pointing to both niche-hunter case studies.
- Substantially extended glossary/ai-agent-behavior with a new section “Why agent decisions follow patterns: connection to the academic foundations.” Applies the May 5 academic wave (jagged-frontier + recognition-primed-decision) to the agent-decision layer: agent purchasing biases are the inside/outside-frontier asymmetry playing out at decision time. Bumped 🌱 → 🌿.
- No stats changes from this session (the ai-agent-behavior page was already in the glossary count).
Week of May 5 – 11
Drafts cleanup + advisor-strategy ingest 🧹 (May 5, late evening) — Fifth and final productive session of the day. Cleaned up the drafts/ folder (three stale files renamed with _archived_2026-05-05_ prefix) and ingested two raw articles. The bigger of the two ingests — Anthropic’s April 2026 Advisor Strategy — produced the day’s most architecturally interesting synthesis: the strategy-vs-execution pattern is fractal, recurring at three layers (org chart / individual workflow / model architecture) for the same structural reasons.
- Created glossary/advisor-strategy — Anthropic’s April 2026 model-pairing pattern. Cheap executor (Sonnet/Haiku) drives tasks; Opus advisor consulted only on hard decisions. Server-side, single API request,
max_usescost ceiling. Benchmark: Haiku + Opus advisor = 41.2% on BrowseComp (vs. 19.7% solo) at 85% lower cost than Sonnet alone. The architecture is the model-level fractal of the wiki’s comparisons/strategy-vs-execution-ai thesis. - Substantially enriched comparisons/strategy-vs-execution-ai with a new “The pattern recurs at three layers (a fractal observation)” section. Names the org-chart / individual-workflow / model-architecture layering as evidence the frame captures something structural, not a 2026 incidental.
- Updated questions/ai-as-personal-advisor with a new section drawing the model-architecture parallel: a personal AI advisor is structurally the same architectural insight at a different scale.
- Updated questions/managed-agents-break-even with two new dimensions: Anthropic’s task-horizon thesis (if sessions trend toward hours/days/weeks, the break-even math shifts toward Managed) and the advisor-strategy economics (executor-tier flexibility makes cheaper-model executors more viable for agent work).
- Updated tools/claude-managed-agents with the explicit task-horizon positioning + cross-links to advisor-strategy and jagged-frontier.
- Added glossary/advisor-strategy to Related on automation/ai-agent-organization.
- Stats: ~125 → ~126 pages, glossary 44 → 45.
- Drafts cleanup: three files (organic-content-pillar-draft, two reddit-response files) renamed with
_archived_2026-05-05_prefix. Pillar was shipped April 30; Reddit response insights already covered by strategist-pattern and glossary/llm-wiki-pattern.
Academic AI-productivity wave 📚 (May 5, evening) — Ingested four peer-reviewed sources to anchor the automation-eats-execution framework with academic empirical evidence. The framework previously rested on three industry data points (Seufert/Modash/Karpathy); it now also rests on three randomized/quasi-experimental academic studies plus a theoretical foundation. Six independent anchors across multiple methodologies — unusually broad empirical support for a 2026 management framing.
- Ingested Dell’Acqua et al. (2023, BCG × Harvard, n=758 consultants) — published as glossary/jagged-frontier. Inside the AI capability frontier: +12.2% tasks, +25.1% faster, +40% quality. Outside it: −19pp accuracy. The frontier is invisible from a task description.
- Ingested Brynjolfsson, Li & Raymond (2023, NBER, n=5,179 customer-support agents) + Noy & Zhang (2023, Science, n=444 writers) + Dell’Acqua’s bottom-half-skill finding — synthesized as glossary/ai-skill-leveling. Three independent studies, three methodologies, same finding: AI raises low-performer productivity disproportionately. The skill premium compresses.
- Ingested Noy & Zhang’s task-decomposition finding specifically — published as glossary/ai-task-restructuring. AI compresses rough-drafting; idea generation and editing become the new bottleneck. The mechanism behind why novices benefit most.
- Ingested Klein 1998 + Kahneman-Klein 2009 — published as glossary/recognition-primed-decision. The theoretical foundation: pattern-matching judgment (human or AI) is reliable only in high-validity environments with rapid feedback. Predicts where AI itself will struggle for the same structural reasons humans do.
- Substantially upgraded comparisons/strategy-vs-execution-ai with a major new “Peer-reviewed academic foundation” section. The framework now distinguishes academic empirical evidence from industry observation in its sourcing.
- Added an empirical-anchors-table-for-academic-studies to glossary/automation-eats-execution. Reorganized Sources into industry data / peer-reviewed academic / synthesis.
- Substantially upgraded questions/what-ai-tools-actually-deliver-roi with direct ROI numbers from all three studies. The seedling now has a real empirical answer for inside-frontier tasks.
- Added theoretical-reliability section to questions/ai-as-personal-advisor (Klein-Kahneman conditions).
- Added jagged-frontier reliability caveat to questions/managed-agents-break-even.
- Cross-linked glossary/dual-process-thinking ↔ glossary/recognition-primed-decision for the complete fast-intuition picture (when it works, when it doesn’t).
- Stats bumped (~121 → ~125 pages, glossary 40 → 44). Added “Academic foundations cited” stat: 10 papers.
- The cross-domain framework architecture is now an eight-page cluster: comparison synthesis + named-framework glossary + open question + three domain anchors + four academic foundations. The wiki’s most empirically anchored framework.
Lint follow-up wave 2 🛠️ (May 5, late) — Closed the remaining lint findings except those requiring hands-on real-world work. The cross-domain “automation eats execution” framework now has a four-page architecture: synthesis comparison page + named-framework glossary stub + open question tracking adjacent-domain expansion + three domain anchors. This is the most complete framework architecture the wiki has built around a single Primores synthesis.
- Created glossary/automation-eats-execution — names the cross-domain pattern as its own framework so other pages can cite it without rebuilding the argument. Three empirical anchors in a single table.
- Created questions/automation-eats-execution-next-domains — working hypotheses for 5 candidate domains (email/CRM, SEO, brand-building, B2B sales, analytics). The brand-building negative case is the most useful: framework predicts it’s NOT on the curve, and empirically it isn’t.
- Closed the orphan-cluster lint finding by adding cross-references to all 7 industry-cases pages from automation/ai-implementation-patterns (was 3 → now 11; including 3 previously-orphan pages on developer-tools, healthcare, and security).
- Upgraded 4 glossary entries (🌱 → 🌿):
lighting-recipe,focal-hierarchy,framing-archetype,creative-formula-vs-creative-skin. All are well-integrated (7 inbound links each) and substantively complete. - Stats bumped (~119 → ~121, glossary 39 → 40, questions 3 → 4).
- Three lint growth opportunities transparently deferred to real sessions: hands-on creator-discovery experiment, influencer-platform tool reviews, and the Modash vs Aspire vs CreatorIQ comparison. Wiki schema requires hands-on testing for experiments and tool reviews; speculative versions would violate the wiki’s own quality bar.
Lint sweep + 3 follow-up pages 🧹 (May 5, later) — Ran the first full wiki lint since April 20. Wiki health is genuinely strong (zero frontmatter violations, zero private-path leakage, zero true orphans, zero stale seedlings). Lint surfaced two real broken links and a comparison-page synthesis opportunity unlocked by the May 4-5 ingests. All resolved in the same session.
- Created getting-started — public orientation page for new readers (who the wiki is for, how it’s organized, quick paths into the content). Closes a broken link referenced from CLAUDE.md and the maintenance protocol.
- Created glossary/vibe-coding — Karpathy’s February 2025 term for AI-assisted coding by intent. Where it works (prototypes, personal tools, learning), where it breaks (production, security-sensitive, long-lived). Closes a broken link from the dev-tools cases page.
- Created comparisons/strategy-vs-execution-ai — the cross-domain synthesis the May ingests made writable. AI eats execution work; strategy, judgment, integration stay human-leveraged. Three empirical anchors: paid media (Seufert), influencer marketing (Modash 2026), software (Karpathy / vibe coding). The wiki now has three named anchors for the same pattern at three different layers — this is the meta-page that ties them together.
- Wired 3 missing inbound links to marketing/influencer-marketing-task-overload from
marketing/overview,marketing/ai-marketing-case-studies, andautomation/finding-ai-use-casesper lint recommendation. - Stats bumped (~116 → ~119 pages, glossary 38 → 39, comparisons 3 → 4).
Influencer Marketing Task Overload 📋 (May 5) — Ingested Modash’s 2026 salary survey (n=499 influencer marketers). Captured the report’s central operational finding: the role is structurally overloaded with ~19 distinct weekly tasks, and the salary data shows execution-style tasks pay the least while strategy/leadership tasks pay the most. Mapped the 19 tasks against current AI tooling capability (Tier 1: AI eats it; Tier 2: AI assists, human owns the call; Tier 3: stays human-leveraged) — and the pattern that emerges is the same as paid-media’s “creative is the new targeting”: automation eats execution, strategy stays the lever.
- Created marketing/influencer-marketing-task-overload (~280 lines, 🌿 growing) — challenge-first framing, the full 19-task weekly stack, strategy-vs-execution salary gradient (+$14,830 for strategy ownership), Primores-synthesis AI-fit tiering, the “just go do social” failure mode (cross-functional creep → 12% lower pay, 15% lower satisfaction).
- Updated glossary/creative-is-new-targeting with a new section showing the same automation-eats-execution pattern in influencer marketing — generalizing the framing beyond paid media.
- Stats bumped (~115 → ~116 pages, 34 → 35 domain pages).
- Out-of-scope per direction: salary geography, gender pay gap, freelance economics, job satisfaction breakdowns. The original Modash report covers these.
Week of April 28 – May 4
“Creative is the new targeting” glossary entry 🎯 (May 4) — New glossary anchor for the phrase that’s been used informally in Primores work since 2023. Captures the structural shift in performance marketing (post-iOS 14.5 ATT) where Meta Advantage+, Google PMax, and TikTok Smart+ commoditized channel-side work, leaving creative variation as the dominant remaining lever.
- Created glossary/creative-is-new-targeting — origin (Seufert / Mobile Dev Memo), the three things that got automated, the leverage shift, the AI-era twist, and honest-limits scoping (works for DTC/mobile-app; weakens for B2B/regulated/local/brand-building).
- Wired into glossary/creative-formula-vs-creative-skin as the upstream “why” for the formula/skin distinction.
- The wiki now has named anchors for all three layers a client engagement might cross between: paid-performance (this entry), brand-building (Sharp’s mental availability + distinctive assets), and organic content-marketing (organic-content-strategy pillar).
Organic Content Strategy pillar published 🚀 (April 30, end of day) — The day’s foundation reading culminates: the pillar shipped from drafts/ to wiki/marketing/ as full prose, plus two operational spokes.
- Created marketing/organic-content-strategy (~3,400 words) — the pillar. Full prose across seven sections, grounded in all six foundation sources (Ajzen, Pirolli & Card, Granovetter, Sharp, Cialdini, Kahneman). Three time horizons + cognitive substrate. Recipes + iGaming cases.
- Created marketing/discovery-before-scale (~1,800 words) — operational spoke covering the two-phase architecture (Discovery 2-4 weeks → Scale indefinite). Decision tree, failure modes, the math (9K × 100 = 6-9M views only if validated).
- Created marketing/behavioral-profile-fingerprinting (~1,900 words) — the four-ratio measurement spoke (save/share/comment/follower). Each ratio with industry baseline + Cialdini-principle activation. Recipes case worked example.
- Three of four planned spokes now exist (slideshow-pattern-design earlier today, discovery-before-scale + behavioral-profile-fingerprinting tonight). Fourth (tiktok-user-behavior-fundamentals) deferred — the academic foundations now live in glossary entries.
- Stats bumped (~111 → ~114 pages).
Kahneman Thinking, Fast and Slow ingest 🧠 (April 30, late) — The pillar’s foundation reading is now substantively complete. Five major sources (Ajzen, Pirolli & Card, Granovetter, Sharp, Cialdini, Kahneman) all ingested in one day. Kahneman is the substrate layer underneath the others, not a parallel framework.
- Created glossary/dual-process-thinking — Kahneman’s System 1 / System 2 framework, cognitive ease as a persuasion lever, the lazy-System-2 thesis, and three explicit cognitive-substrate connections to existing wiki concepts.
- Connected three existing entries to their cognitive substrates: Sharp’s mental availability ↔ Kahneman’s availability heuristic (same mechanism, different layer); Cialdini’s Scarcity ↔ Kahneman & Tversky’s loss aversion; Pirolli & Card’s information scent ↔ Kahneman’s WYSIATI. The wiki now has explicit cross-disciplinary connections, not just adjacent frameworks.
- The three-time-horizon model now has a substrate row: years (Sharp/availability heuristic) — months (Primores/cognitive ease + repeated exposure) — the moment (Cialdini/System 1 trigger features). All three run on the same dual-process substrate.
Cialdini Influence ingest 📘 (April 30, late) — Six persuasion principles + a Primores-original mapping page that converts intuitive slideshow-pattern picking into deliberate behavioral-profile design.
- Created glossary/persuasion-principles — Cialdini’s six (reciprocation, commitment & consistency, social proof, liking, authority, scarcity) plus the click-whirr meta-framework. Each principle gets a mechanism + canonical experiment + AI-era content relevance.
- Created marketing/slideshow-pattern-design — Primores-original mapping of 9 recurring slideshow patterns onto specific Cialdini principles. Sorts patterns by behavioral profile (save-bait, share-bait, comment-bait, follow-bait). The first wiki page that converts pattern selection from gut-feel to deliberate design.
- Extended marketing/brand-vs-content-layers from a two-layer to a three-time-horizon model: years (Sharp/brand-building), months (Primores prior #4/content authority), the moment (Cialdini/compliance triggers). All three compose; none replaces the others.
Sharp’s How Brands Grow ingest 📕 (April 30, evening) — The book labeled “THE load-bearing source” in the pillar draft now has its three core frameworks as wiki citizens, plus a top-level reconciliation page resolving an apparent worldview-prior conflict.
- Created glossary/mental-availability, glossary/distinctive-assets, glossary/double-jeopardy-law — Sharp’s three-point empirical conclusion (growth = popularity = light buyers), the distinctiveness-vs-differentiation argument, and the empirical anchor that smaller brands get hit twice (fewer buyers + less loyalty).
- Created marketing/brand-vs-content-layers — top-level architectural page reconciling Sharp’s “broad reach builds mental availability” with Primores’ worldview prior #4 (“narrow + exhaustive beats broad TOFU”). Both are correct at different layers; both are needed for a complete strategy. Includes layer-specific decision criteria.
- Updated glossary/super-niche and glossary/topical-authority with “What This Doesn’t Replace” sections explicitly scoping them to the content-marketing layer.
- Architectural payoff: Sharp + Granovetter + Pirolli & Card + Ajzen now form a complete causal chain for brand growth in the AI-era — cross-cluster diffusion → attention allocation → mental availability + distinctive recognition → purchase intention → behavior.
Foundation reading ingests 📚 (April 30) — Three academic papers ingested to ground the forthcoming Organic Content Pillar.
- Ajzen 1991 — Theory of Planned Behavior — Refreshed glossary/tpb with the foundational paper as primary citation (was citing the 2014 defense paper). Added empirical baseline (R≈.51 for behavior, R≈.71 for intention prediction across 16-19 studies), PBC vs Locus of Control disambiguation, past-behavior + moral-obligation extensions, and the methodological caveat that belief-based attitude measures only explain 10-36% of variance. Reframed the “subjective norms are weak predictors” finding from AI-specific to general.
- Pirolli & Card 1999 — Information Foraging Theory — Created glossary/information-foraging. Covers information scent, patches + Charnov’s Marginal Value Theorem, diet selection (Independence from Encounter Rate), and enrichment. Frames the 200-400ms scroll-decision window as the empirical signature of Charnov’s MVT applied to feed environments. Reinforces glossary/super-niche and glossary/topical-authority with foraging-math grounding.
- Granovetter 1973 — The Strength of Weak Ties — Created glossary/weak-ties. Bridge argument (strong ties cluster → bridges are necessarily weak ties), diffusion implication, and the Primores-original extension that algorithmic feeds operate as synthetic weak-tie bridges. Reinforces glossary/smra with the same framing.
Strategist Pattern documentation 🧠 (April 29-30) — New top-level meta page on turning a wiki into a thinking partner. Initial publish (three-layer architecture, six worldview priors, seven mapped capabilities, maintenance loops) on April 29; updated April 30 with the session-signals architecture (third loop), self-pressure beat discipline, and confidence calibration. See strategist-pattern.
Workflow tightening 🔧 (April 30) — Added wiki/changelog.md to the INGEST workflow (step 7) and Session End Checklist in CLAUDE.md. Reason: the log gets touched atomically with each ingest, but the changelog had been drifting day-by-day. This change aligns the two surfaces.
Production sprint ⚡ (April 28-29) — Six new wiki pages in two days, plus integration sweep.
- Schwartz’s Awareness Levels — New glossary page synthesizing Eugene Schwartz’s “Breakthrough Advertising” (1966) frameworks (Five Levels of Awareness, Five Stages of Market Sophistication) with AI-era applications. Enhanced same day with direct book quotes via the new tools/pdf-streamer.
- AI Interface Layer — New marketing strategy page on Claude becoming the “front door” between users and apps. Argues SEO success doesn’t transfer to AI visibility — different channels, different mechanisms. See marketing/ai-interface-layer.
- Claude Connectors — Tool documentation covering 200+ ready-to-use integrations (Blender, Adobe, Spotify, Uber, HubSpot, Salesforce). Available on every plan including Free. See tools/claude-connectors.
- GEO/AEO Benchmarks 2026 — Comprehensive 2026 data on AI search impact: 31.3% generative AI adoption, 48%+ of Google queries showing AI Overviews, 61% CTR drop when AI Overviews appear, +35% CTR for cited brands. See seo/geo-aeo-benchmarks-2026.
- PDF Streamer tool — New Primores Claude Code skill for converting large PDFs (30+ pages) to markdown page-by-page. Resumable, vision-fallback aware, strips repeated headers. See tools/pdf-streamer.
- 4 creative-cluster glossary stubs — Lifted glossary/creative-formula-vs-creative-skin, glossary/focal-hierarchy, glossary/framing-archetype, and glossary/lighting-recipe from the published article cluster into the glossary so the vocabulary has a definition home.
- Integration sweep — Cross-linked the six new pages into eight neighbor pages so today’s work isn’t orphaned. Refreshed stats.
Week of April 21-27
Reverse-engineering pillar 🎨 — Largest content release to date.
- AI Creative Reverse-Engineering pillar + 6-cluster article series — 1,185 lines of public content on the Formula vs Skin framework, surface vs structural mimicry, AI template casting workflow, and ethics of reverse-engineering ads. See cases/ad-alchemy-creative-reverse-engineering.
- Brand Voice Skills Guide — How to build Claude Skills for consistent brand voice with LLM-learning foundations. See marketing/brand-voice-skills-guide.
- Niche Hunter case study + skill — Three niches evaluated against five-axis validation (AI visibility = GO, Reddit workflow = GO, e-commerce content = MAYBE). See cases/niche-hunter-fresh-2026-04 and tools/niche-hunter.
- Creative reverse-engineering article cluster (a01–a15) — 15 articles published on Meta Ad Library workflow, focal hierarchy, framing archetypes, lighting recipes, and the legality of reverse-engineering ads.
- Cross-repo dispatch to primores-web on content push (operational improvement).
- Reddit Thread Analyzer skill + substance ranking concept — See tools/reddit-thread-analyzer and glossary/substance-ranking.
- Niche Hunter primores-creative case — First five-axis validation run for our own brand. See cases/niche-hunter-primores-creative.
- Public methodology + LLM usage guide — See methodology and contributing.
Week of April 14-20
Major content growth 📈 — Wiki crossed 50 pages.
- Agenica.ai case study — AI agent vs manual competitor ad monitoring.
- Google Cloud AI dataset ingest — 1,048 cases analyzed; 232 metric-rich cases distributed across 10 industry pages.
- AI Implementation Patterns (meta-analysis page) — what actually works across the 1,048-case dataset. See automation/ai-implementation-patterns.
- Reddit shill detection synthesis — Detecting astroturfing patterns. See marketing/reddit-authenticity-patterns.
- TPB framework dissertation ingest — Multi-model synthesis (S-O-R + TAM + TPB).
- Vietnamese Gen Z TikTok research ingest — Algorithm impact on mental well-being.
- New glossary entries: glossary/smra, glossary/tpb, glossary/astroturfing, glossary/super-niche, glossary/topical-authority.
- Strengthened competitor-analysis domain (was the thinnest); added experiments/ad-alchemy-competitor-piggyback and experiments/seo-geo-content-ecommerce.
Week of April 7-13
Wiki launch 🚀
- Initialized the wiki structure
- Established content methodology (see methodology)
- Created templates for consistent page formatting
- Ready to start building knowledge!
Stats (as of 2026-05-04)
| Metric | Count |
|---|---|
| Total pages | ~115 |
| Glossary entries | 38 |
| Tool reviews | 11 |
| Comparisons | 3 |
| Domain pages | 34 |
| Case studies | 8 |
| Experiments | 4 |
| Open questions | 3 |
| Google Cloud AI cases ingested | 1,048 (232 with metrics) |