Skip to content
Primores AI wiki

Primores AI Wiki — Index

Primores AI Wiki

TL;DR: A practical knowledge base about using AI in business — focusing on marketing, SEO, competitor analysis, and automation. No deep tech required.

Welcome to the Primores AI Wiki. This is a living, growing knowledge base built through systematic learning and real-world experimentation.

Why This Wiki Exists

Most AI business content falls into two buckets: hype (“AI will change everything!”) or theory (“here’s how transformers work”). Neither helps you actually implement AI in your business.

This wiki is different:

  • Real case studies — Named companies, specific metrics, documented results
  • Tested tools — Hands-on reviews, not marketing copy
  • Named frameworks — Memorable patterns you can apply immediately
  • Honest limitations — What doesn’t work, not just what does

🧠 How To Use This Knowledge Base

This wiki is built to be useful in multiple ways:

For Reading & Learning

Browse by domain (Marketing, SEO, Automation) or start with:

For Decision Making

Use the structured comparisons and case studies:

For AI-Assisted Research

This wiki is designed to be referenced by AI assistants. The structure makes it easy for LLMs to find, understand, and cite:

Reference this wiki for context: https://primores.org/wiki


Then ask: "What does the wiki say about [your topic]?"

Why this works for LLMs:

FeatureWhy It Helps AI
TL;DR on every pageQuotable summary AI can cite directly
Named frameworks”The 90% Club Pattern” is memorable and searchable
Clear headingsAI can navigate to specific sections
Cross-linked conceptsRelationships are explicit, not implied
Tables & listsStructured data AI can parse accurately
Real metricsSpecific numbers AI can reference confidently

Example prompts:

  • “Based on the Primores wiki, what’s the most common first AI project and why?”
  • “What does the wiki say about AI customer service — any case studies with metrics?”
  • “Summarize the wiki’s GEO/AEO framework for optimizing content for AI search”

For Claude Code / Cursor / AI IDE users: Point your AI at the wiki folder and say "Use this wiki as context" — it will search and cite relevant pages automatically.

For Building Your Own Knowledge Base

This wiki demonstrates the glossary/llm-wiki-pattern — an AI-maintained knowledge system that compounds over time. See methodology if you want to build something similar.

For Turning the Wiki Into a Strategist

Add ~7 small config files on top of the wiki and you’ve got a domain-specific thinking partner that loads automatically in your terminal. Knowledge in the wiki, capabilities in skills, persona in priors — see strategist-pattern for the architecture and a worked Primores example with six worldview priors and seven mapped capabilities.

Content Maturity

  • 🌱 Seedling — Early thoughts, may change
  • 🌿 Growing — Solid but still developing
  • 🌳 Evergreen — Comprehensive, maintained

See methodology for how this wiki is built and maintained.

Domains

Marketing

AI applications for content, campaigns, personalization, and analytics

  • 🌿 marketing/overview — AI for Marketing overview

  • 🌿 marketing/ai-interface-layer — When Claude becomes your app’s front door (brand visibility implications)

  • 🌿 marketing/brand-vs-content-layers — Reconciling Sharp’s broad reach with Primores’ narrow authority (different layers, both needed)

  • 🌿 marketing/organic-content-strategy — PILLAR — Why AI-era organic content compounds and how to engineer it (six-source academic foundation)

  • 🌿 marketing/discovery-before-scale — Two-phase operational framework: validate pattern × niche before scale

  • 🌿 marketing/behavioral-profile-fingerprinting — Four-ratio measurement framework (save / share / comment / follow)

  • 🌿 marketing/slideshow-pattern-design — 9 slideshow patterns mapped to Cialdini’s 6 principles — Primores-original behavioral-profile design framework

  • 🌿 marketing/ai-marketing-case-studies — Named companies, specific metrics, real results

  • 🌿 marketing/brand-voice-skills-guide — Building Claude Skills for consistent brand voice (with LLM learning mechanics)

  • 🌿 marketing/reddit-authenticity-patterns — Detecting shills and building trust on Reddit

  • 🌿 marketing/ai-video-marketing — Using AI to enhance authentic video storytelling

  • 🌿 marketing/preparing-for-agentic-ai — Brand strategy for the agentic era

  • 🌿 marketing/social-commerce-psychology — Emotional & cognitive triggers driving purchases

  • 🌿 marketing/influencer-marketing-task-overload — 19 weekly tasks per influencer marketer; AI-automation map + strategy-vs-execution salary gradient (Modash 2026 survey)

  • 🌿 marketing/telegram-marketing-channel — Telegram as a marketing channel: 1B+ MAU; primary for iGaming globally and DTC in Russia/CIS/MENA; NOT Western Web3 fashion (that’s Discord). Channel-fit is geographic before categorical.

  • 🌿 marketing/ai-tells-in-sales-copy — Two-principle frame for client-facing copy: (1) don’t sound like AI [11-pattern catalog], (2) model the reader’s motivation before drafting [structural-pain modeling]. Audience-mode review + CMO-believability score. iGaming reach-scarcity reframe as the argument-level case.

  • 🌿 marketing/ai-human-voice-prompting — Six techniques for AI human voice in social posts + outreach. The 80/95–5/20 hybrid ratio is empirically universal. Platform-specific tactics: LinkedIn 360Brew (saves > likes), X/Grok (replies > likes), TikTok (script-AI exempt from labeling), cold email (deliverability splits AI 71%/human 86%). Generation-side complement to ai-tells (editing-side).

  • 🌿 marketing/marketing-analytics-in-2026 — The cookieless attribution stack. MMM adoption surged 212% since 2023. Dual-model operating norm: MMM strategic + multi-touch tactical + AI reconciliation. Data clean rooms improve cross-channel accuracy up to 40%. Plus cohort/LTV:CAC as capital-efficiency layer (B2B SaaS 3.2:1, DTC subscription 4.1:1).

  • 🌿 marketing/channel-economics — Vertical-agnostic paid-channel framework: the intent/cost map (search highest-intent/cost & gatekept; social cheap/weak; native+messenger the “golden middle”) + owned-channel diversification. The block-vs-ceiling distinction — restricted verticals (iGaming/crypto) are blocked from paid, DTC/ecom hits a CAC ceiling (+40–60% 2023→25). Same “build a channel you keep” thesis, different trigger. Benchmarks carry verification verdicts; “20–60% reg-to-deposit” refuted (real ~2–4%).

  • 🌿 marketing/alternative-ad-channels — Landscape + ranked “try first” picks beyond the Meta/Google duopoly. The load-bearing insight: published entry floors mislead in BOTH directions — Telegram’s “€2M” overstates (a TON self-serve path starts ~$50), Reddit’s “$5/day” understates (an advertiser-approval wall gates it; alcohol/dating/gambling/finance/crypto/pharma need a Reddit sales rep, weapons/tobacco/dupes banned). Ranked for DTC/ecom: Bing frictionless-first, Reddit best-economics-if-cleared, Pinterest/Snapchat/Amazon, Telegram as specialist/arbitrage. Includes the Telegram-Ads paid-product deep-dive (0.1 TON min CPM, contextual targeting, three-path access). 13 claims verified / 12 refuted — most repeated Telegram “facts” failed.

  • 🌿 marketing/ai-product-video-fidelity — How to make a high-fidelity AI product video for reflective / fine-detail products (jewelry, watches, packaging): composite the real product photo (AI owns only the environment) + first-last-frame keyframing (not single-frame i2v, which makes the model invent motion and drift the product). Plus drift control, filter-safe prompting, and honest failure modes. The production-method companion to marketing/ai-video-marketing.

  • 🌿 marketing/ai-product-image-generation — The static-image companion: crawl your real product photo, feed it to a reference / image-to-image model (Nano Banana Pro, Seedream 4.5, Flux 2) and re-stage it across endless on-brand scenes that keep the product accurate. As of mid-2026 this is the default for ordinary static products (verify each output); hand-compositing is the fallback for reflective/glass/jewellery, fine on-pack text, exact-colour SKUs, electronics screens, regulated goods, and all video. The model re-stages a supplied product — it doesn’t invent one. Feeds TikTok Photo-Mode carousels + image-to-video. Disclosure is an enforcement matter (FTC §5 + Operation AI Comply 2024; Endorsement Guides last revised 2023; penalties attach to order/rule violations, not automatically) + marketplace rules + SynthID.

  • 🌿 marketing/landing-page-hero-archetypes — Three field-tested hero patterns — tension-triad, number-anchor, honest-scope-upfront — selected by awareness level × page depth × scroll-force. The load-bearing rule: a hero that summarizes the mechanism closes the curiosity loop in place and kills the scroll. The landing-page counterpart to the ad-creative archetypes.

  • 🌿 marketing/evidence-graded-audience-research — Units instead of avatars: classify inputs by who can know them (ask-only / file-preferred / researchable), grade every claim provided/researched/hypothesis, build from verbatim customer language, output a unit table [TA × JTBD × angle] with stable IDs and transparent priority sub-scores. Imports intelligence-community evidence grading into audience research.

  • 🌿 marketing/prescriptive-production-briefs — When AI makes production cheap, the brief is the product: final copy, char-counted overlays, frame-level slot tables; executors flag, never invent. Hook = slot 1 only (3 mechanisms, not synonyms); asset-source column routes rows to production pipelines. From a real 20-unit / 360-variant batch — including the hook-collision finding.

  • 🌿 marketing/andromeda-era-creative-strategy — Once Meta’s Andromeda retrieval engine picks the best ad from a huge pool, the unit of work is a reusable template library, not single ads: 10–15 genuinely different concepts (archetype × angle), templated across 1:1/4:5/9:16, rotated every 2–3 weeks. The staged-compiler split applied to paid static. Engine confirmed (Meta); strategy specifics + lift numbers flagged vendor-directional.

  • 🌿 marketing/meta-ad-placement-safe-zones — Durable Meta placement reference: aspect ratios by placement, Reels/Stories/Feed safe & dead zones in pixels (~270px top / ~672px bottom on 9:16), the March 2026 Stories+Reels unification (design to 35% bottom), auto-placement cropping, and placement-aware copy rules. Re-verify-before-a-big-run caveat + official Meta source links.

  • 🌿 marketing/meta-ad-policy — Deployable Meta ad-policy reference for health/fitness/beauty/finance creative. The four highest-frequency rejections: negative self-perception, personal attributes (the “you” trap — 2026 enforcement now catches indirect implication), before/after & idealized results, body-part focus. The reliable on-policy reframe: spotlight the food/product/environment, agitate the food — never the viewer’s body. Repeat hits are an account-level risk (lockdowns/bans), heaviest on young/pre-launch accounts. Worked example: a nutrition app’s brand-spec compliance block + pre-flight checklist.

  • 🌿 marketing/ecommerce-to-tiktok-ai-pipeline — The honest playbook for turning an e-commerce store into TikTok content with AI: the 5-stage pipeline (scrape → LLM hooks/scripts → AI-UGC video → auto-caption → schedule) and where it breaks. AI compresses everything except product truth and platform trust. No independent evidence AI UGC out-converts human; the one peer-reviewed study finds a trust penalty for recognized AI. Tool-stack fact-check (Blotato MCP real; “v0 for video” false; CapCut no API), credibility-graded economics (AI ~$2–30 vs human ~$50–500/video), the TikTok suppression myth (unoriginal ≠ AI), risk-ranked legal (FTC §5 claims = the real risk; ROP panic overblown; NY synthetic-performer disclosure June 2026), and the SMB-end-to-end vs enterprise-behind-the-scenes adoption split. Plus a low-risk pilot.

  • 🌱 marketing/ad-platform-agent-surfaces — Mid-2026: TikTok, Reddit, and LinkedIn shipped two different things at once — agent control surfaces (TikTok Ads MCP Server + Ads Skills, making it the last big-four platform to ship an ads MCP after Google/Meta) and on-platform AI creative (TikTok Symphony/Seedance 2.0, Reddit’s four Community-Intelligence ad products at Cannes, LinkedIn Brand Kit). The platforms are absorbing both the agent-control and creative-generation layers. Load-bearing caveat: all announcements, not GA-with-results.

  • 🌿 marketing/email-deliverability — The 2024–26 deliverability squeeze: inbox placement fell to 83.5%, spam placement nearly doubled in-year (4.5%→8.6% by Q4), Gmail −5%, Microsoft toughest at 75.6% — ~1 in 6 legitimate emails misses the inbox, and failure went silent (junk-foldering, not bounces). The Gmail/Yahoo bulk-sender stack (DMARC alignment, branded domain, one-click unsub, <0.3% complaints; permanent rejections from Nov 2025, Microsoft parallel May 2025) hits SMB/DTC hardest. Diagnostic-order framing, vendor flags carried.

  • 🌿 marketing/ad-account-suspensions — Account-level platform enforcement, officially documented: Google suspended 39.2M advertiser accounts in 2024 (mostly pre-serve fraud — the honest caveat), “Circumventing Systems” is the #1 reason (37% of the best cohort) and a vague catch-all (a second account after suspension IS the violation), Google’s 80%-incorrect-suspension-reduction claim implicitly concedes prior severity, Meta admitted a >100% false-positive spike Q4 2025. The wrongful-suspension rate is undocumented — that’s the scandal. Pre-flight audits + diversification as the hedge; recovery-agency skepticism throughout.

  • 🌿 marketing/email-design-system — The durable human anchor for AI email production: the seven layers a reusable brand email design system needs (tokens, anatomy, tested HTML component library, formatting rules, flow maps, voice, ESP technical notes) — the asset the tools/ai-email-production-stack loop multiplies against so output reads as the brand, not generic AI slop. Build by extracting from your best sends.

  • 🌿 marketing/funnel-mechanics — A first-principles funnel framework for 2026: AARRR (Acquisition/Activation/Retention/Referral/Revenue) as a diagnostic stage skeleton, NOT a literal path (McKinsey/Google-messy-middle/Gartner all show looping behavior). Stage-by-stage leak diagnosis beats a top-line conversion number; activation (value experienced ≠ task completed) is the LTV lever. The AI-era shift quantified: Pew shows clicks halve (15%→8%) when an AI summary appears, ~60% of searches are zero-click — TOFU compresses into the AI answer and buyers arrive later/more informed. “Funnel is dead” is overcorrected: the literal-path was always a simplification, the stage-diagnosis instrument still works. Every figure credibility-tiered (Pew strongest; Amplitude/Bain/Littledata directional).

  • 🌿 marketing/static-ad-template-system — A static ad = angle × template × product (three independent axes). A foundry deconstructs ads into brand-agnostic, catalogued templates (compress, don’t fork — 10 ads → 4 templates); an apply step pours an angle + brand into a template and renders all aspect ratios. The template library is the durable, compounding asset; the rendered ad is disposable output (staged-compiler split applied to static creative). Honest limit: it lays out a message, it can’t supply one — quality is capped by the angle research, and specific templates stay experimental until a real results read.

SEO

AI-powered search optimization, content tools, and technical automation

  • 🌿 seo/overview — AI-Era SEO overview: the zero-click shift, the fan-out→retrieval→citation gate chain, access control, and the articles-engine operating model
  • 🌳 seo/ai-seo-content — How to create AI-optimized content that gets cited
  • 🌿 seo/articles-engine — The six-stage articles engine (July 2026): the operational system most businesses are missing — map → produce → publish → interlink → measure → refresh. The measured freshness bias (25.7% by publish date, 13.1% by last-update, ChatGPT +458 days vs AI Overviews slightly older), Google’s value-not-method spam line (now explicitly covering AI Overviews/AI Mode citations), and the honest gap: cadence economics are unmeasured (the “200× at 12+/month” stat is refuted, do-not-cite).
  • 🌿 seo/agentic-search — How AI agents decide which brands get found
  • 🌿 seo/agentic-search-optimization — The full ASO discipline (the new SEO). May 2026 update: E-E-A-T binary filter, brand-mentions-3x-stronger-than-backlinks.
  • 🌳 seo/ai-visibility — Getting found in AI-generated answers. May 2026 update: 75M Google AI Mode daily users, E-E-A-T as binary filter, the four new load-bearing signals.
  • 🌿 seo/ai-crawler-access — The access-control flip side of AI visibility: the training / retrieval / user-fetch bot taxonomy, why a WAF enforces where robots.txt only asks, and current (2026) UA strings for OpenAI/Anthropic/Google/Perplexity/Meta. Cloudflare default-blocks AI crawlers + HTTP-402 pay-per-crawl. Anthropic now publishes IP ranges; Perplexity stealth-crawl incident Aug 2025.
  • 🌿 seo/geo-aeo-benchmarks-2026 — Hard numbers on AI search impact (2026 data). Refreshed May 18 with AI Mode adoption + E-E-A-T off-site validation findings.
  • 🌿 seo/new-site-ranking — How to rank without a big budget (long-tail strategy)
  • 🌳 seo/zero-click-strategy — The strategic operating model for a 64.82%-zero-click world (92-94% in AI Mode). Brand-and-visibility-first; the dual mandate; PR strategy as SEO strategy. Now source-calibrated: Pew Research’s July 2025 clickstream (primary) anchors the AI-Overview CTR collapse — ~47% measured; the 64.82%/96%/3× figures are flagged vendor estimates.
  • 🌿 seo/ahrefs-ai-search-studies-2026 — What 1B+ data points say about AI search: a calibrated synthesis of 14 Ahrefs 2026 studies. Six findings — AI surfaces are separate discovery layers (86% same answer, 13.7% citation overlap; only 38% of AIO citations rank top-10); CTR collapse accelerating (−58%, up from −34.5% in 8 months); YouTube mentions the top visibility correlate (0.737); schema markup has no measured effect on AI citations (the one quasi-experiment); “Best X” listicles 43.8% of ChatGPT citations; answers churn every 2.15 days but stay 0.95 semantically stable. Tier 2 single-vendor evidence, graded.

AI creative reverse-engineering cluster (published from drafts, June 2026 — 🌱 seedlings):

Competitor Analysis

The operational methodology for competitive intelligence — five layers tied to specific decisions

  • 🌿 competitor-analysis/overview — PILLAR — Five-layer operational methodology for 2026: win-loss analysis + continuous monitoring + battlecards + share of model + creative reverse engineering. Why SWOT/Porter fail as operational tools. AI compresses execution; strategy stays human. All five layers now have dedicated glossary entries.
  • 🌿 competitor-analysis/dna-beauty-paid-social-whitespace — Worked market note (June 2026): US paid-social DNA-skincare has one live paid funnel; the formulation+subscription model is unmarketed — and three attempts at it died. The skepticism headwind, the $50–70/mo churn line (anchor vs treatments, not serums), the whitespace angle map, and what an indirect ad sweep catches vs misses.
  • 🌿 competitor-analysis/ad-transparency-surfaces — The cross-platform ad-library map (July 2026): what Google, TikTok, LinkedIn, X, Microsoft, and Meta publicly expose about competitor ads. The DSA Article 39 asymmetry (EU-targeted ad data is structurally richer everywhere), the no-spend-data rule, ~1-year retention windows, Google’s July 2026 AI-ad labels (My Ad Center, not the Transparency Center — and unreliable as a CI signal), TikTok’s programmatic targeting/reach API, LinkedIn’s B2B targeting disclosures, the X €120M DSA fine.

Automation

Workflow automation, integrations, no-code/low-code AI solutions

AI Implementation Case Studies by Industry (from Google Cloud 2026 dataset):

Tools

Reviews and guides for AI tools

  • 🌿 tools/ai-visibility-audit — Claude skill for GEO/AEO audits (0-100 score)
  • 🌿 tools/claude-connectors — 200+ ready-to-use integrations (Blender, Adobe, Spotify, Uber)
  • 🌿 tools/claude-skills — Reusable instruction packages for Claude workflows
  • 🌿 tools/claude-managed-agents — Anthropic’s ready-made agent infrastructure
  • 🌿 tools/claude-cowork — Desktop agent for autonomous knowledge work
  • 🌿 tools/gemini-omni — Google’s any-to-any multimodal model (May 19, 2026 launch). Unifies video, image, audio, text generation under one architecture with Gemini’s reasoning + world-model physics (Project Genie) + Veo + Nano Banana. Ships in Gemini app + Google Flow + YouTube Shorts; Vertex AI API in coming weeks. Publisher’s tool framing vs. Sora 2’s artist’s tool; prompt adherence + text rendering are load-bearing for ads.
  • 🌿 tools/mcp — Model Context Protocol for connecting AI to systems
  • 🌱 tools/obsidian — Markdown-based knowledge base app
  • 🌿 tools/pdf-streamer — Large PDF to markdown (page-by-page, resumable) (Primores)
  • 🌿 tools/product-article-generator — AI content tool for e-commerce (Primores)
  • 🌿 tools/reddit-thread-analyzer — Substance-based Reddit content extraction (Primores)
  • 🌿 tools/niche-hunter — Super-niche discovery & article mapping (Primores)
  • 🌿 tools/target-audience-research — Research-to-units compiler for paid social (Primores): brand inputs → evidence-graded unit table [TA × JTBD × angle] with stable IDs, three gates, verbatim language harvesting. Dogfooded to sign-off on a real engagement.
  • 🌿 tools/ad-template-foundry — Deconstructs a static ad into a brand-agnostic, catalogued template + a compositor (Primores skill 18). Dedupe-aware: 10 source ads → 4 templates. Builds the reusable library; distinct from ad-alchemy’s one-shot casting.
  • 🌿 tools/ad-template-apply — Assembles a finished ad: angle × template × product → rendered creative in all aspect ratios (Primores skill 19). Pulls copy from the angle research — an assembler, not a copywriter. The inverse of the foundry.
  • 🌿 tools/scenario-compiler — Unit-to-production-package compiler (Primores): 3+3 cards × 3 hook mechanisms = 18 named variants per unit, slot tables, asset-source routing, deduped shotlist. Dogfooded on a 20-unit / 360-variant batch.
  • 🌿 tools/ai-video-production-stack — Capability → job map for AI product-video tools (June 2026 snapshot): plate generation, real-product placement, environment-only inpaint, first-last-frame keyframing, chaining. Covers higgsfield as a hub (Soul, Nano Banana Placement, Start & End Frames) + the model stack (Kling 3.0, Seedance 2.0, Nano Banana Pro, Veo, Luma Ray3, Midjourney). Key finding: higgsfield’s MCP is generation-only — placement/inpaint/keyframe stay UI-only. Companion to marketing/ai-product-video-fidelity.
  • 🌿 tools/ai-email-production-stack — Capability → job map for AI email production (June 2026 snapshot): brand design system + generative-image MCP (hero) + channel-ready HTML assembly = single-session loop. Same collapsed-loop shape as the video stack, different channel. The MCP’s generation-only surface fits email’s one automatable visual step. Honest limits: copy/voice, brand fidelity, and deliverability stay human — AI-looking imagery costs trust ~4:1 and image-heavy emails hurt deliverability.

Glossary

Plain-English definitions of AI concepts

  • 🌿 glossary/ai-agent — AI systems that take actions
  • 🌿 glossary/ai-crawler — Automated bot that fetches your pages for an AI system; the three types (training / retrieval / user-fetch) decide your access policy. Anchors the AI-crawler-access cluster.
  • 🌿 glossary/gptbot — OpenAI’s training crawler; block to opt out of training (separate from ChatGPT search, which is OAI-SearchBot)
  • 🌿 glossary/ccbot — Common Crawl’s bot; one block reduces your reach into many models’ training data
  • 🌿 glossary/bytespider — ByteDance’s crawler; the canonical robots.txt-ignoring scraper
  • 🌿 glossary/waf — Web application firewall; the enforcement layer that actually blocks AI bots (robots.txt only asks)
  • 🌿 glossary/llms-txt — Advisory markdown map of your site for AI; helps comprehension, not access control
  • 🌿 glossary/pay-per-crawl — Cloudflare’s HTTP-402 model letting sites charge AI crawlers for access
  • 🌿 glossary/ai-creative-reverse-engineering — Deconstructing a winning ad’s formula into a reusable template for your product (the cluster’s canonical definition)
  • 🌿 glossary/ai-ugc-ads — AI-generated UGC-style ads: what works, and what reads fake
  • 🌿 glossary/astroturfing — Fake grassroots marketing patterns
  • 🌿 glossary/awareness-levels — Schwartz’s Five Levels of Awareness + Market Sophistication
  • 🌿 glossary/ai-agent-behavior — How AI agents make decisions and their biases — now with the underlying Allouah et al. (Columbia/Yale Dec 2025) study (n=1,000 × 8 categories), Cialdini-on-AI 28K-prompt finding, and model-improvement curves
  • 🌿 glossary/agent-adoption-frictions — Wharton 2026 Blueprint: three psychological barriers (perceived competence, trust, delegation of control) blocking agent adoption. User-side counterpart to ai-agent-behavior.
  • 🌿 glossary/ai-competitive-analysis — Can you hand competitive/strategic analysis to an LLM? Yes, with three disciplines: aggregate many runs (a single LLM is biased — Doshi et al. 2025, SMJ), keep humans on the judgment dimensions where LLMs score worst (Csaszar et al. 2024, Strategy Science), and use AI to triage signal not decide. Plus: AI for both framing and ideation cuts strategic quality −15pp via anchoring (Wu et al. 2025, INSEAD).
  • 🌿 glossary/appropriate-reliance — The goal with AI is calibrated reliance, not maximal. AI labeling triggers costly over-reliance (Klingbeil 2024) and suppresses critical thinking (Lee 2025, CHI) — yet experts under-rely, and disclosing AI use erodes trust (Schilke & Reimann 2025, 13 experiments). Reconciled via expertise × stakes moderation.
  • 🌿 glossary/agent-engineering — Karpathy at Sequoia AI Ascent 2026: the professional discipline of coordinating agents reliably. “Vibe coding raises the floor. Agentic engineering raises the ceiling.” Includes the Software 1.0/2.0/3.0 framing.
  • 🌳 glossary/automation-eats-execution — Cross-domain framework: AI compresses execution work, strategy stays human-leveraged (paid media + influencer marketing + software anchors)
  • 🌿 glossary/cognitive-automation — AI that makes decisions in workflows
  • 🌿 glossary/context-engineering — Designing information flow for AI agents
  • 🌿 glossary/customer-perception-moments — Primores framework consolidating the behavioral-evidence research: perception crystallizes at discrete moments of judgment (decision, review-writing, failure-recovery); every headline finding has a context-dependent moderator that can flip it; honest-assessment is the unifying mechanism. The hub tying together weekend-review-effect + ai-humor-forgiveness + honest-assessment.
  • 🌿 glossary/distinctive-assets — Brand-specific cues (colors, logos, tone) that build mental availability — Sharp
  • 🌿 glossary/double-jeopardy-law — Smaller brands get hit twice (fewer buyers + less loyalty); penetration is the lever
  • 🌿 glossary/dual-process-thinking — Kahneman’s System 1 / System 2 + the cognitive substrate underneath mental availability, scarcity, and information scent
  • 🌿 glossary/e-e-a-t — Experience, Expertise, Authoritativeness, Trustworthiness. In 2026 it acts less like a soft ranking signal and more like a near-binary gate on AI citation. The 96%/3×/<4%-DA figures are vendor estimates; the direction (earned third-party authority beats brand-owned) has preprint support. The load-bearing concept under the SEO/GEO cluster.
  • 🌿 glossary/creative-formula-vs-creative-skin — Reusable formula vs. swappable skin in ad creative
  • 🌿 glossary/creative-is-new-targeting — Why performance-marketing leverage moved from audience to creative (post-ATT, Advantage+/PMax/Smart+)
  • 🌿 glossary/app-tracking-transparency — Apple’s iOS 14.5 privacy change (Apr 2021): most users declined tracking, collapsing ad signal. The root cause behind modeled attribution, the MMM revival, AI broad targeting, and “creative is the new targeting.” The ~$10B Meta hit was a Feb-2022 company estimate. Actually landed on schedule — unlike Google’s walked-back cookie deprecation
  • 🌿 glossary/meta-andromeda — Meta’s ML ad-retrieval engine (detailed Dec 2024): selects a few thousand ads from tens of millions of candidates, built for Advantage+ creative volume. +6% recall / +8% ad quality (on selected segments), 10,000× model capacity, NVIDIA Grace Hopper. Careful attribution: Meta built the engine; “creative is the new targeting” is the agency synthesis on top. Vendor lift numbers flagged unverified.
  • 🌿 glossary/advantage-plus — Meta’s umbrella AI ad-automation suite: end-to-end campaign types (Sales/App/Leads) + single-step features (Audience/Creative/Placements). By 2026 the enforced default — Advantage+ Shopping renamed to Sales (Feb 2025), legacy ASC/AAC APIs deprecated (v24.0 Oct 2025 → v25.0 Q1 2026). All Meta perf figures self-reported + stale (2022); the “32% ROAS” only holds with its ASC+BAU footnote. Core critique: attributed return ≠ incremental lift — paired with Measured’s counterpoint (64% new-customer incrementality). Control is hybrid, not absolute.
  • 🌿 glossary/performance-max — Google’s goal-based AI campaign: assets + a goal in, ML-run bidding/placement/creative across all Google inventory (Search/Display/YouTube/Gmail/Discover/Maps) out. The Google twin of Advantage+. GA Nov 2021; force-migrated Smart Shopping & Local (2022). Launched as a black box, walked back transparency 2023–25. Two critiques: branded-search cannibalization (Optmyzr, 503 accounts, 91% overlap) and attributed ≠ incremental. Google’s “13% incremental” is self-reported, attribution-based.
  • 🌿 glossary/advantage-plus-creative — Meta’s single-step creative-automation feature (formerly Dynamic Experiences): standard enhancements + generative AI (background/full-image generation, AI video, voice dubbing). On by default at ad level; per-enhancement opt-ins since API v22.0 (Jan 2025) — but disabled enhancements reportedly re-activate. Brand-safety risk is documented, not hypothetical: the “AI granny” ad swap, distorted imagery, brand-color/copy drift. Meta’s own +7% conversions / +22% ROAS figures are attribution-based. The concrete case for human-anchored creative.
  • 🌿 glossary/focal-hierarchy — Ordering the eye’s path through an ad
  • 🌿 glossary/framing-archetype — 8 reusable ways to stage a product in an ad
  • 🌿 glossary/lighting-recipe — Highest-leverage transferable element in ad creative
  • 🌳 glossary/geo-aeo — Optimizing content for AI search engines
  • 🌿 glossary/geo-anchor — First-sentence citation optimization
  • 🌿 glossary/honest-assessment — AI trust signal through admitting weaknesses
  • 🌿 glossary/human-anchored-ai-multiplication — The market uses AI as a creator; the data says it’s an amplifier. Fully AI-generated ads underperform (Ipsos −14%/−17%); AI-alone output measurably converges (3 evidence layers); the trust penalty concentrates on fabricated people and recognizable AI. Multiply the human shoot — the non-promptable input. Counter-findings carried (pure-AI CTR win; the perceived-humanness reconciliation).
  • 🌿 glossary/ftc-ai-marketing-enforcement — The deceptive-AI-claims risk (vs disclosure law). The FTC’s May 2026 Cox Media “Active Listening” settlement: $930K across three firms over a marketed AI ad-targeting service that used no voice data and resold marked-up broker email lists. Claiming an AI capability you don’t have is itself deceptive under FTC §5 — the enforcement edge of “AI-washing.” Distinct from glossary/content-provenance (labeling AI content).
  • 🌿 glossary/content-provenance — Four mechanisms people constantly conflate: C2PA/Content Credentials (signed metadata, easily stripped) · SynthID (in-pixel watermark, robust but Google-family only) · platform labels (TikTok/Meta/YouTube — labeling doesn’t cut reach where stated) · disclosure LAW (enforceable). The law is converging on 2 Aug 2026 (EU AI Act Art. 50 + California); NY synthetic-performer law live 9 Jun 2026; FTC §5 + Operation AI Comply (Endorsement Guides last revised 2023; penalties attach to order/rule violations). Absence of a credential proves nothing. Disclose conspicuously; don’t lean on platform labels for legal compliance.
  • 🌿 glossary/information-foraging — Pirolli & Card’s theory of attention as rate-optimization (scent, patches, Charnov’s MVT)
  • 🌿 glossary/llm — Large Language Models explained
  • 🌿 glossary/llm-evals — Evaluation systems for AI products
  • 🌿 glossary/llm-nudges — How AI guides user decisions
  • 🌿 glossary/llm-wiki-pattern — Compounding knowledge bases with AI
  • 🌿 glossary/mental-availability — Sharp’s central thesis — propensity to be thought of in buying situations
  • 🌿 glossary/persuasion-principles — Cialdini’s six (reciprocation, commitment, social proof, liking, authority, scarcity) + click-whirr meta-framework
  • 🌿 glossary/prompt-engineering — Getting better AI outputs
  • 🌿 glossary/rag — Retrieval-Augmented Generation
  • 🌿 glossary/rumpelstiltskin-effect — Why naming customer problems drives sales
  • 🌿 glossary/advisor-strategy — Anthropic April 2026 model-pairing pattern: cheap executor (Sonnet/Haiku) + Opus advisor on hard decisions. Haiku+Opus = 2× BrowseComp at 85% cheaper than Sonnet alone
  • 🌿 glossary/agent-outcomes — Goal-oriented agent work with graders
  • 🌿 glossary/ai-skill-leveling — Three studies (Brynjolfsson n=5,179, Noy-Zhang n=444, Dell’Acqua n=758): AI raises low-performer productivity disproportionately; the skill premium compresses
  • 🌿 glossary/ai-task-restructuring — Noy & Zhang 2023 Science: AI compresses rough-drafting; idea generation and editing become the new bottleneck
  • 🌿 glossary/fine-tuning — Customizing AI models for your tasks
  • 🌿 glossary/jagged-frontier — Dell’Acqua 2023, BCG × Harvard, n=758: AI is asymmetric — +12% inside the frontier, −19pp outside it. The frontier is invisible from a task description. Now also includes Karpathy’s “jagged intelligence” (Sequoia 2026) as the model-side cousin.
  • 🌿 glossary/recognition-primed-decision — Klein 1998 + Klein-Kahneman 2009: when pattern-matching judgment (human or AI) is reliable. Theoretical foundation for the jagged frontier
  • 🌿 glossary/skill — Reusable AI instruction packages
  • 🌿 glossary/smra — Social Media Recommendation Algorithms explained
  • 🌿 glossary/substance-ranking — Content quality over popularity metrics
  • 🌿 glossary/super-niche — Audience × Problem × Context territory selection
  • 🌿 glossary/topical-authority — Exhaustive interlinked coverage strategy
  • 🌿 glossary/tokens — How AI measures and charges for usage
  • 🌿 glossary/tpb — Theory of Planned Behaviour in AI adoption
  • 🌿 glossary/vibe-coding — Karpathy’s term for AI-assisted coding by intent — works for prototypes, risky for production. Now includes the May 2026 Sequoia update with the floor/ceiling distinction and Software 1.0/2.0/3.0 framing.
  • 🌿 glossary/weak-ties — Granovetter’s bridge argument; algorithmic feeds as synthetic weak-tie bridges (Primores extension)
  • 🌿 glossary/zettelkasten — Connected notes methodology
  • 🌿 glossary/hallucination — The signature failure mode of LLMs: plausible-sounding but false output. Why human verification is non-optional for AI-generated factual claims
  • 🌿 glossary/agentic-memory — How AI agents remember across sessions: 4 layers (working/episodic/semantic/procedural). Memory is engineered, not built-in. The wiki itself is one realization of semantic memory
  • 🌿 glossary/tool-use — The capability that turns chatbots into agents. Model decides when/which/how; application executes. The basic primitive under all 2026 agentic work
  • 🌿 glossary/guardrails — Production safety layer for AI systems. Paired with tool-use: every powerful tool needs a corresponding guardrail. Six basic categories per Pimenov’s playbook
  • 🌿 glossary/embeddings — Numerical representations of text that preserve semantic similarity. Foundational layer under semantic search, RAG, recommendation systems
  • 🌿 glossary/marketing-mix-modeling — Top-down statistical attribution; aggregate spend + outcome data, no cookies needed. Adoption +212% since 2023 due to cookie deprecation. Google Meridian + Meta Robyn democratized what was six-figure consulting
  • 🌿 glossary/incrementality-testing — The causal layer under attribution. Geo-holdout, audience-split, time-based test designs. 10–20% holdout standard; synthetic controls for geo; “lift” = truly-incremental conversions. When MMM and incrementality disagree, incrementality wins
  • 🌿 glossary/multi-touch-attribution — Fractional conversion credit across journey touchpoints (rules-based: last/first/linear/time-decay/U/W; data-driven: Shapley, Markov). ATT + cookie loss degraded it (reported 40–60% iOS signal loss) — now a tactical, in-platform signal, not ground truth. “MTA is dead” is vendor framing; it’s demoted, not dead. Triangulate with MMM (strategic) + incrementality (causal). Completes the attribution trio.
  • 🌿 glossary/engaged-view-conversions — Why YouTube ROAS reads too high: EVC credits a watch-then-buy with no click, and sits inside the main Conversions column (video-action/Demand Gen) — so conversions can exceed clicks. Decisive tell = conversions > clicks; fractional counts = modeled; restate absurd ROAS as an absurd CPA. Inflation scales with organic baseline (worst on high-traffic sites). Believe: backend order reconciliation → conversion-type segmentation → click-through floor → geo-holdout.
  • 🌿 glossary/cohort-analysis — The shape of the retention curve matters more than the M12 endpoint. M0–M3 onboarding cliff drives LTV; once-a-DTC-customer-buys-twice they retain 85–90%. 2026 LTV:CAC benchmarks: B2B SaaS 3.2:1, DTC subscription 4.1:1
  • 🌿 glossary/agent-payment-protocols — Four standards constitute the production agentic-commerce stack (Sep 2025 – May 2026): AP2 (Google authorization), x402 (Coinbase+Cloudflare HTTP-native stablecoin), UCP (Google+retailer-coalition commerce schema), Visa TAP (agent identity). AWS Bedrock AgentCore Payments shipped May 7. x402 hit $600M annualized by March 2026. Anthropic’s Project Deal documented the “agent quality gap.”
  • 🌿 glossary/win-loss-analysis — The CI layer practitioners credit most with moving win rate. Now academically calibrated: the 5–15pp magnitude is a practitioner figure, but the mechanism (structured retrospective review) is anchored in tier-1 AAR/debrief meta-analyses (Tannenbaum & Cerasoli 2013 d=.67; Keiser & Arthur 2021 d=.79) and “interview buyers not reps” in dyadic sales research (Endres 2023). Klue/Crayon 8-practice convergence on methodology.
  • 🌿 glossary/battlecards — The sales-enablement layer. 2026 evolution: static battlecards dying; living battlecards are modular, AI-assisted, role-specific (SDR/BDR/AE/CSM), with governance metadata. One-screen rule. Success metric: win-rate-against-named-competitor
  • 🌿 glossary/share-of-model — The AI-search competitive dimension that didn’t exist in 2024. ChatGPT holds ~79% of generative AI traffic; AI Overviews in 89% of brand searches. Sector concentration extreme (Apple 54.38% mention share in consumer electronics). For non-dominant brands, specialization is the only viable path
  • 🌿 glossary/prompt-caching — Production cost-optimization layer for LLM applications. Three mechanisms: prompt cache (vendor-level, Anthropic cache_control / OpenAI automatic), semantic cache (vector-similarity via Redis/GPTCache), KV cache (model-internal). Cache reads 10% of base price; 70-80% total cost reduction achievable with combined strategies
  • 🌿 glossary/continuous-monitoring — The 2026 CI discipline replacing the quarterly competitive-landscape deck. Always-on signal tracking across pricing, product, hiring, ads, reviews; Crayon/Klue track 100+ signal types per competitor. Bottleneck is signal-to-noise triage and decision routing, not collection. AI reads 200 competitor pages and surfaces the 3–5 that matter. Layer 2 of the five-layer competitor-analysis methodology.
  • 🌿 glossary/creative-reverse-engineering — Systematic discipline of deconstructing competitor ads to extract reusable structural patterns (composition, lighting, focal hierarchy, copy skeleton). The 2026 vision-LLM stack (Claude for copy + GPT-4o for visuals + Foreplay/Atria/Motion for collection) compressed what used to be an art-director consulting engagement into an afternoon. Pattern extraction across volume is load-bearing; single-ad deconstruction underperforms. Layer 5 of the five-layer competitor-analysis methodology.
  • 🌿 glossary/ai-humor-forgiveness — Self-deprecating humor in AI service-failure responses produces +47.8% forgiveness uplift vs no humor (Xie et al. 2025, JBR, n=1,919). Outperforms positive humor by ~10pp consistently. Two boundary conditions: severity gate (works for low-severity; vanishes for high-severity) and focal-customer gate (Honora et al. 2025, JBE counter-finding — humor directed at burned customer reads as sarcasm). The Hosanagar forgiveness-asymmetry quote: AI errors are weighted heavier than human errors. Adjacent tactics: thanks-not-sorry (gratitude beats apology for rejection failures) + sparse interjections. One of the few empirically-supported tactics for the AI-error perception gap.
  • 🌿 glossary/weekend-review-effect — Online reviews submitted on weekends average 3% lower 5-star share and 6% higher 1-3 star share vs weekdays (Bayerl et al. 2026, JMR, n=400M reviews / 33 platforms). The Eleanor Rigby thesis: weekend reviewers self-select into a more socially-isolated population. Counter-finding: reverses for hedonic products (2023 study, n=588K). Industry response-rate data conflicts (Saturdays are high-volume per Bazaarvoice). The wiki’s four-lever review-ops cluster integrates the timing finding with first-review anchoring, incentive-positivity transfer (up to +83.4% per Woolley & Sharif 2021), and display-order effects.
  • 🌿 glossary/review-response-strategy — How to reply to reviews, backed by two ISR studies: responses lift future review volume 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 ones (Ravichandran & Deng 2023). The fifth review-ops lever.
  • 🌿 glossary/reference-image-conditioning — Show-don’t-tell for AI aesthetics: control generation by feeding reference images (composition / palette / structure / style / character) instead of prose. Routes around the description bottleneck for non-visual users and for looks you can recognize but can’t articulate. Tools: Midjourney sref/cref/cw, Flux Kontext, Soul guided generation, Nano Banana Pro. The high-leverage use is brand coherence — feed the client’s own glossary/distinctive-assets so output reads as them. Carries the N=1 “coherence via structure, not literal copy” pattern.
  • 🌿 glossary/best-x-listicle — The single most-cited content format in AI chatbots (43.8% of ChatGPT’s cited page types per Ahrefs 2026). Top-third placement + freshness drive AI recommendation; AI favors third-party comparative judgment over brand-owned copy. Strategy: own the list in your category AND get placed highly in third-party lists (digital PR, not content marketing).
  • 🌿 glossary/retrieval-vs-citation — Being fetched by an AI engine ≠ being cited. ChatGPT cites only ~50% of retrieved URLs; the rest feed the answer as uncredited background. Predictors of citation: title↔prompt semantic match (0.602 vs 0.484 cosine) and readable URL slugs (89.8% vs 81.1%). Reframes GEO as two gates: get retrieved AND get cited.
  • 🌿 glossary/query-fan-out — AI search doesn’t search what the user typed: engines rewrite the query into multiple synthetic sub-queries and answer from the combined pool (Google-documented for AI Mode, primary source; ChatGPT fans out on ~2/3 of prompts vs ~30% Perplexity, vendor-measured). Rank #1 on the typed phrase ≠ visibility in the answer. Models inject “best” + the current year — the structural reason listicles dominate. Upstream mechanism of retrieval-vs-citation: the query your title must match is the model’s rewrite.

Comparisons

X vs Y analyses

Experiments

Tests, trials, and their results

Case Studies

Real-world implementations and lessons learned

Questions

Open explorations and things we’re figuring out

  • 🌿 questions/ai-as-personal-advisor — How can AI serve as a personal business advisor? Now synthesized into a reliability framework: validity boundary × expertise-stakes calibration (glossary/appropriate-reliance) × selective escalation (glossary/advisor-strategy). The deliverable is the reliance-calibration discipline, not the tool stack.
  • 🌿 questions/automation-eats-execution-next-domains — Which marketing functions are next on the automation-eats-execution curve? Now with a candidate-scoring matrix — and the sharpened rule: AI eats execution that’s high-volume/structured, not execution that’s cumulative (brand) or relational (B2B sales).
  • 🌿 questions/managed-agents-break-even — When does DIY beat Managed Agents on cost? (parked July 2026 — answered as of May pricing: break-even ~2,000–5,000 sessions/mo; reopen on pricing change)
  • 🌿 questions/what-ai-tools-actually-deliver-roi — What AI tools actually deliver ROI for small businesses? Now with the two-axis model (frontier position × error cost) + function-by-function map + free-vs-paid reasoning.

Ad reverse-engineering & Meta Ad Library Q&A (published from drafts, June 2026 — 🌱 seedlings):

Meta

About this wiki

  • 🌳 about — Who we are and what we do
  • 🌳 contributing — How to use and grow this wiki
  • 🌿 getting-started — Orientation for new readers — who this wiki is for, how it’s organized, quick paths into the content
  • 🌳 maintenance — Lint and maintenance protocol (health checks, cadences)
  • 🌳 methodology — How this wiki is built
  • 🌿 strategist-pattern — Turn this wiki into a thinking partner (worked Primores example)

Stats

MetricCount
Total pages222
Glossary entries92
Tool reviews18
Comparisons4
Domain pages78
Case studies8
Experiments4
Open questions3 (1 parked July 2026)
Google Cloud AI cases232 (of 1,048 ingested; ingestion idle since May — resume or close deliberately)
Academic foundations cited10 (Ajzen, Pirolli & Card, Granovetter, Sharp, Cialdini, Kahneman 2011, Klein 1998, Brynjolfsson 2023, Noy-Zhang 2023, Dell’Acqua 2023)

About

This wiki is maintained by Primores.org — practical AI consulting for businesses.

Questions? Ideas? Get in touch