What is generative engine optimization? Where the term came from and what the evidence says works

By Andrej Ruckij · November 1, 2026

TL;DR: Generative engine optimization (GEO) is the work of getting your business mentioned and cited in answers written by AI, in ChatGPT, Google’s AI Overviews and AI Mode, Perplexity, Claude and Copilot. The name comes from a 2023 research paper that rewrote web pages and found that adding quotations and statistics raised their visibility in AI answers by up to about 40%. That result only held for pages the AI had already picked up. A July 2026 review of 45 studies found no technique with a stable effect across platforms, and several studies found formatting changes on their own did little. My read: GEO is mostly SEO plus being talked about elsewhere. Get your pages found by the search systems these assistants use, put specific facts on them (prices, dates, numbers), earn mentions on the sites AI reads, and measure with a fixed list of buyer questions. Be wary of anyone selling page rewrites as the main lever.

Primores helps brands get found and recommended in AI answers: see how we work with brands.

Where the term comes from

The term was coined in a paper called “GEO: Generative Engine Optimization” by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. It was first posted in November 2023 and accepted to KDD 2024, a major data-science conference. The authors called it “the first novel paradigm to aid content creators” in AI search.

What they actually tested is worth knowing, because in my reading much of today’s GEO advice still leans on it.

The setup. For each question, they took the top 5 Google results, had GPT-3.5 write an answer from them, and then rewrote one of the five pages with one of nine methods: sounding more authoritative, adding statistics, keyword stuffing, citing sources, adding quotations, simpler language, better fluency, unique words, and technical terms.
The benchmark. GEO-bench, 10,000 queries from nine sources, with results reported on a 1,000-query test set.
The measure. How many words from the page ended up in the answer, weighted by how early it was cited, plus a “subjective impression” score that GPT-3.5 itself graded.

What they found: the abstract reports gains of up to about 40%; in the paper’s Table 1, the best methods reach up to 41% on the word-count measure and 28% on impression. Adding quotations and adding statistics worked best. Keyword stuffing did worse than doing nothing (17.8 against a 19.5 baseline). On a 200-question sample in Perplexity, adding quotations raised the word-count measure by about 22%, and keyword stuffing came out 10% worse. Citing sources lifted the fifth-ranked page by 115.1%, while top-ranked pages dropped 30.3% on average under the same method. The authors read that as a chance for smaller sites to compete.

The limits. The pages were already among the five sources the AI read. So the paper shows how to get more of your page into an answer once the AI has already picked it up as a source; it says nothing about how to get picked up in the first place. The authors say they didn’t evaluate search ranking effects, tested only two engines, and that methods “may need to adapt over time as GEs evolve”. GPT-3.5 is long gone from these products.

A July 2026 survey by Olivier Martinez (an arXiv preprint; PPC Land gives his affiliation as Sciences Po) makes the same point more sharply: the paper’s gains hold only when a source is already in the model’s context, so “it does not mean 40 percent more readers will click”, as PPC Land’s report on the survey puts it.

How do AI answers pick their sources?

In the prompts studied so far, the assistants almost always search the web before answering. In a dataset of 602 prompts studied by Zhang, He and Yao (April 2026 preprint), search was triggered 98.64% of the time in ChatGPT, 99.67% in Google and 100% in Perplexity. The answer is then written from what that search returned. So the first question is always: which search does each assistant use?

Assistant	Where it finds pages (from primary or named sources)	What that means for you
Google AI Overviews and AI Mode	Google’s own index. Google says both may use “query fan-out”, “issuing multiple related searches across subtopics and data sources”. A page must be “indexed and eligible to be shown in Google Search with a snippet”.	Normal Google SEO is the entry ticket. Google says there are no “special optimizations necessary”.
ChatGPT	OpenAI’s crawler, OAI-SearchBot, which OpenAI says “is used to surface websites in search results in ChatGPT’s search features”. Peec (a tracking vendor) found in September 2026 that ChatGPT also pulls from at least eight outside providers, including scraped Google results, Microsoft’s Web IQ (which Peec describes as Microsoft’s grounding platform, the service that also powers Copilot), Yelp and TripAdvisor.	Allow OAI-SearchBot. Peec’s advice: “stop assuming Bing visibility is a proxy for ChatGPT visibility.” Since ChatGPT still queries Google, my inference is that a drop in Google could show up in ChatGPT too.
Perplexity	Its own crawler. Perplexity: “To ensure your site appears in search results, we recommend allowing `PerplexityBot`”.	Allow PerplexityBot. Perplexity cites the most sources per answer (16.35 on average in the Zhang study, against 6.88 for ChatGPT).
Claude	Brave Search, by strong evidence rather than an announcement: Anthropic added Brave Search to its subprocessor list on March 19, 2025, and Simon Willison found Claude’s citations matched a Brave search exactly.	Being findable in Brave matters for Claude.
Microsoft Copilot	Peec reports Microsoft confirms Web IQ powers Copilot as well as ChatGPT.	The same Microsoft grounding layer as ChatGPT’s Web IQ results; I found no public detail on how it ranks sources.

Two things matter more than the details. First, retrieval and citation are separate steps. In the Zhang study, Perplexity cited about 16 pages per answer and ChatGPT about 7, but the pages ChatGPT used had far more influence on its answer. Our wiki covers the gap between being read and being credited in glossary/retrieval-vs-citation.

Second, each assistant is a different target. One OpenAI block in robots.txt (OAI-SearchBot) removes you from ChatGPT search answers, though OpenAI says you “can still appear as navigational links”, and the change takes “~24 hours”. Blocking GPTBot only opts you out of training. new-seo-playbook-ai-search has the data on how little the citations overlap between Google’s AI surfaces and ChatGPT.

Generative engine optimization vs SEO: what actually differs

Semrush’s definition is a fair one: GEO is “optimizing your presence and content to appear in responses generated by AI-powered search systems”, and in GEO “you’re competing to be part of the final output” rather than for the top of the results page.

In practice the overlap is large. Microsoft’s AEO and GEO guide, which is written for retailers and their product catalogs, says that “SEO and catalog investments built a foundation” and “the base remains the same: up-to-date product feeds and clear, crawlable, structured content.” Google goes further and says you need nothing extra.

What does change:

The unit of success. SEO counts rankings and clicks. GEO counts whether you’re mentioned, cited and described correctly in an answer. Semrush lists “AI visibility, AI mentions, AI citations, and AI share of voice.”
Who else talks about you. An answer is written from several sources at once, and many are not yours. In Amit Sharma’s study of 112 Product Hunt startups, ChatGPT recognised 99.4% of them by name but recommended only 3.3% when asked for the best tool in their category. Reddit mentions (r = +0.40) and links from other sites (r = +0.32) predicted being recommended. The GEO score of the startups’ own pages didn’t (r = −0.10, not significant). That study tested GPT-4 without web search, so read it as a direction. how-can-chatgpt-recommend-my-business goes through it in detail.
The search engines involved. For ChatGPT, Claude and Copilot, your Google position is only part of the picture.
Volatility. Martinez’s survey notes that commercial audits find low overlap between sources and high run-to-run variation. The same question asked twice can cite different pages, so a single check means little.

GEO vs AEO, and what is AI search optimization called?

They’re mostly two names for the same work. You’ll also see “AI SEO” and “AI visibility”. Our glossary entry treats them as one: glossary/geo-aeo.

Microsoft is one of the few to draw a line between them. Its guide defines AEO (answer or agentic engine optimization) as optimizing content “for AI agents and assistants (like Copilot or ChatGPT) so they can find, understand, and present answers effectively”, and GEO as optimizing “for generative AI search environments” to make content “discoverable, trustworthy, and authoritative”. Its example for a rain jacket is the clearest explanation I’ve seen:

SEO: “Waterproof rain jacket”.
AEO: “Lightweight, packable waterproof rain jacket with stuff pocket, ventilated seams and reflective piping”.
GEO: “Best-rated waterproof jacket by Outdoor magazine, no-hassle returns allowed for 180 days, three year warranty, 4.8 star rating”.

So in Microsoft’s split, AEO is complete, specific facts and GEO is outside proof. You need both, and the label you hire under matters less than whether the work covers both.

What the evidence says works, and what doesn’t

Here’s what I found, from studies that tested things rather than vendor blog posts.

Holds up:

Being relevant to the actual question. In a controlled test by Vishwakarma, Kumar and Jamidar (May 2026 preprint, 252,000 trials across six AI models and 18 content factors), topical relevance and position in the list of sources were the strongest drivers of which page got cited first. In the Zhang study, relevance to the answer was the strongest single correlate of a page’s influence (r = 0.43); that study is observational, and its authors say it doesn’t show these features cause citation.
Specific facts. The same controlled test found that explicit price information and a recent timestamp also helped a page get cited. In the Zhang data, which is observational, pages with numbers, definitions, comparisons, how-to steps or code had more influence on answers.
Being mentioned elsewhere. Reddit mentions and links from other sites predicted being recommended in the Sharma study, while on-page tuning didn’t; those are correlations, not proof of cause. Peec names Yelp and TripAdvisor as licensed sources ChatGPT uses for local results.

Doesn’t hold up, or is unproven:

Formatting-only edits. The Vishwakarma test found they had little impact. In the Zhang data, Q&A-style pages had slightly less influence than other pages (0.0947 against 0.1005), and the authors concluded Q&A packaging alone doesn’t help without real evidence and a good fit to the question.
Rewriting the page body for AI. Martinez’s survey reports a test across 171,003 documents and 2,700 queries where rewriting only the page body cut top-10 presence after reranking by 16% (per PPC Land’s report). The mechanism: a rewrite can make a page more quotable once it’s picked up while making it less likely to be picked up at all.
Generic GEO recipes across domains. In one benchmark the survey covers, only 3 of 54 method and domain combinations were significantly positive; in ecommerce, 10 of 15 heuristics were neutral or negative (again via PPC Land).
Schema markup and llms.txt for AI. Covered in new-seo-playbook-ai-search (an Ahrefs experiment found adding schema didn’t change AI citations) and llms-txt-for-seo-geo. Google says you don’t need “AI text files, or markup” to appear in its AI features.
Keyword stuffing. Worse than nothing, in the original paper itself.

Martinez’s bottom line, from 45 studies published between November 2023 and July 2026: “no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability”. Its best-supported finding is narrower: “already-retrieved content can causally alter its citation or use.” PPC Land’s summary adds that “claims about return on investment from GEO clearly outstrip the academic evidence”.

A caveat on my sources: most are 2026 preprints without peer review, and I read Martinez’s detailed figures through PPC Land’s report rather than the full paper. I found no independent replication of the original paper’s 40% figure on a current assistant.

My read: GEO is search visibility across several search engines at once, plus clear, specific facts on your pages, plus third-party mentions. I see no secret levers in the evidence. That fits what we hold about AI search generally: being cited matters more than the click, and honest, specific content gets used more than praise. My read is that this may also be why the tactics in the original paper haven’t transferred well in later tests. They work on the last step, and the last step is the one you control least.

How to measure generative engine optimization

There’s no Search Console for ChatGPT. Google counts AI Overviews and AI Mode appearances inside your normal Search Console totals, in the Performance report under the “Web” search type, so you can’t separate them there.

What works is a fixed list of the questions your buyers ask, run on a schedule, recording whether you’re mentioned, cited, and how you’re described, against named competitors. You can do this by hand or with a tool. Prices as listed when I read the pages (late October and November 2026):

Tool	Price per month	Questions tracked	Assistants covered
Ahrefs Brand Radar (Custom Prompts)	from $50	5 on the lowest plan, 20 on Advanced	AI Overviews, AI Mode, ChatGPT, Perplexity, Gemini, Copilot; Claude available
Otterly	$29 / $189 / $489 (lower on annual billing)	15 / 100 / 400	ChatGPT, AI Overviews, Perplexity, Copilot; Claude, AI Mode and Gemini as paid add-ons
Peec	$95 / $245 / $495	50 / 150 / 350	ChatGPT, AI Mode, AI Overviews, Copilot, Gemini, Naver; Perplexity and Claude on Enterprise
Semrush AI Visibility Toolkit	$99	25	ChatGPT and AI Mode named for prompt tracking
Ahrefs Brand Radar (AI Visibility Index)	from $199	a pre-collected prompt set, any brand	all platforms listed above
Profound	Enterprise: custom; a self-serve Agency Growth plan (400 credits a month per client workspace) with no published price; free 7-day trial	trial: 50 questions run daily for 7 days; Enterprise: custom	trial: ChatGPT, Gemini, AI Overviews; Enterprise: up to 9, including Perplexity, AI Mode, Copilot, Claude

Because answers vary from run to run, judge your share across the whole list over weeks rather than any single answer on any single day. For most small businesses, 15 to 25 questions are enough. The cheapest plans above cover that, and a spreadsheet does too.

Generative engine optimization agency and tools: what they charge, and red flags

I found no independent survey of what GEO agencies charge. The nearest benchmark is Ahrefs’ survey of 439 SEO providers (updated August 2024): the most common monthly retainer was $501-1,000, 68.8% of all respondents charged $2,000 a month or less, and agencies averaged about $3,209 a month. GEO work that’s mostly SEO plus PR should sit in a similar range; treat that as my estimate, not a measured price.

Red flags, most of them straight from Google’s own guidance on hiring an SEO, which applies just as well here:

Guarantees. Google: “No one can guarantee a #1 ranking on Google.” Nobody can guarantee a ChatGPT mention either, given how much answers vary.
A “special relationship”. Google warns against SEOs who “allege a ‘special relationship’ with Google”. The same goes for anyone claiming a line into OpenAI or Perplexity.
Secret methods. “Be careful if a company is secretive or won’t clearly explain what they intend to do.”
Fabricated content. A 2024 GEO guide from Seer Interactive includes rewriting prompts that tell the user to add invented, “hypothetical” statistics and quotes. Google warns that deceptive or misleading content made on your behalf can get your site “removed entirely from Google’s index”. Ask to see the prompts any agency uses.
Page rewrites as the whole offer. Given the evidence above, a package that’s only “GEO-optimizing” your existing pages buys the weakest lever.
Reporting from one check. Ask how many times each question is run and how they handle variation.

Google’s suggested questions work for GEO too: “What kind of results do you expect to see, and in what timeframe? How do you measure your success?” A good answer names the questions they’ll track and in which assistants. It should also say which competitors you’re compared with and how often you’ll get a report.

Generative engine optimization for small business: who this is for

For: businesses whose buyers ask an assistant before they buy: local services people compare (restaurants, clinics, trades, hotels), online stores, B2B services and SaaS. Local businesses have a specific route, since ChatGPT pulls local results from Yelp, TripAdvisor and Google Maps data, according to Peec.

Not for: businesses whose customers don’t research online, and anyone whose own Google basics aren’t done. If you’re not indexed in Google, GEO work has nothing to build on.

What to do with this: how to implement generative engine optimization

A plan for a small business, about half a day to start and an hour a month after.

Check the doors are open (15 minutes). Open yourdomain.com/robots.txt. Make sure it doesn’t block OAI-SearchBot (ChatGPT search) or PerplexityBot. Blocking GPTBot is fine if you don’t want your content used for training. In Google Search Console, run URL Inspection on your three most important pages and confirm they’re indexed. Changes take about a day to reach ChatGPT.
Write your question list (30 minutes). Fifteen to twenty-five questions a buyer would type, such as “best [service] in [city] for [need]” and “[your category] vs [competitor]”. Use the wording customers use in calls and emails.
Take a baseline (1 hour). Run each question in ChatGPT, Perplexity and Google AI Mode, three times each. In a spreadsheet, note: mentioned (yes/no), cited (your URL or not), what it said about you, and who else appeared. If you’d rather pay, Otterly’s or Semrush’s entry plans cover this.
Put the specific facts on your pages (1-2 hours). Prices or price ranges, service area, opening hours, guarantees, a visible “updated” date, and your numbers (years, jobs done, ratings). In the controlled test, price information and recent dates also helped. Answer your top questions in plain sentences on the relevant page, without a separate “AI version”.
Fix the outside sources (1-2 hours). Local: complete your Google Business Profile, Yelp and, for hospitality, TripAdvisor listings, with the same facts as your site. Everyone: find the “best [category]” lists and directories that appeared in your baseline and ask to be included.
Earn mentions over time (ongoing). Answer questions in the communities your buyers use. how-can-chatgpt-recommend-my-business covers this in detail, with links to doing Reddit without getting banned.
Re-run monthly (1 hour). Same questions, same assistants, three runs each. Track your share of mentions across the list. Expect noise from month to month; look for the direction over a quarter.

The test for whether it’s working: more of your buyer questions mention you, and the descriptions match your facts. Visits from AI assistants will stay small for most businesses, so judge it on the answers.

More on the numbers behind AI search: seo-stats-vendor-vs-measured and new-seo-playbook-ai-search.

Sources

More of this every day on the channel: https://t.me/aidaily_business