Retrieval
The engine searches the live web and pulls candidate pages. If your page is not retrieved — blocked, unindexed, or rendered only in JavaScript — nothing else about it matters.
katman / What is GEO
REFERENCE · WRITTEN 4 SEPTEMBER 2026
What is GEO (Generative Engine Optimization)?
GEO — Generative Engine Optimization — is the practice of making a brand findable, extractable and corroborated so that generative answer engines like ChatGPT, Claude, Gemini, Perplexity, Copilot and Google AI Overviews name it when a buyer asks a question. It is not ranking. A model does not put you in position three; it either has a sentence about you that it can lift and stand behind, or it writes an answer without you in it.
GEO can also mean something else
Outside digital marketing, "geo" is short for geography, and AEO in customs and logistics means Authorised Economic Operator — a trade authorisation with no relation to search. This page covers Generative Engine Optimization for AI answer engines. The disambiguation is here because ambiguous acronyms are how a page ends up answering the wrong question.
GEO, AEO, LLMO, AI SEO: are they different?
Not meaningfully. They name the same work from different angles, and the category has not settled on one label. What does differ is the outcome each one is measured against, and those are genuinely three different things.
| Term | What it emphasises | Measured by |
|---|---|---|
| SEO | A position in a ranked list of links | Rank for a keyword |
| GEO | Being usable by a generative engine writing an answer | Appearance rate across a question set |
| AEO | Being the answer rather than a result | Citation and recommendation rate |
| LLMO | The model's internal representation of the brand | Recall when the brand is named |
The distinction worth carrying is not between the acronyms. It is between being mentioned, being cited and being recommended. A page can be cited as the evidence for a competitor's superiority. Collapsing the three into one score hides which one you are missing, which is why we report them separately.
How a model actually decides what to recommend
Three steps, in order. A failure at any one of them makes the next two irrelevant.
Extraction
The model lifts a span of text out of what it retrieved. A sentence that only makes sense inside the paragraph around it is hard to lift. A sentence that names the subject, the category and the claim survives being cut out.
Corroboration
The model prefers a claim that appears on a source it did not get it from. This is why third-party presence beats a longer page on your own domain, and why a brand nobody else describes stays invisible however good its own site is.
What we measured, and what it showed
On 4 September 2026 we asked ChatGPT, signed out, in a temporary chat, which sites a brand could hire so that AI would recommend it. It named six. We crawled all six — 102 pages — and ran the same measurements against our own site.
Six of six publish a page defining the category. Six of six state somewhere that no one can guarantee a placement in an AI answer. Five of five whose robots.txt we could read name every AI crawler explicitly rather than relying on a wildcard. Six of six are either listed in a public agency directory or operate one — and two of the six results were directories, because an assistant asked "which sites do this" reaches for a list, and a directory is a list.
| Measured on 102 pages | The six sites ChatGPT named | katman.pro, before this page |
|---|---|---|
| Words per page, median | 902 – 1,695 | 410 |
| Subheadings per page | 13 – 17 | 5.5 |
| Pages carrying FAQ schema | 35% – 86% | 21% |
| Distinct outbound domains | 11 – 12+ | 1 |
| AI crawlers named in robots.txt | all of them | none |
We are publishing our own worst numbers because a page about evidence that hides its own is not evidence. The full dataset is in the repository, and this page is the first correction to it.
The two ways an AI visibility check lies to you
Both are common, both feel like proof, and both are how a brand concludes it is visible when it is not.
- You typed your own brand name. A model handed a brand almost always recognises it. We measured a site that scored high on branded recall and appeared in zero of thirty discovery answers across five model families and four languages, in the same run. Branded questions are excluded from every number we report, on purpose.
- You were signed in. A founder showed us a screenshot of ChatGPT recommending his product second. Repeated in a signed-out temporary chat, the product did not appear at all — matching all thirty API observations. A signed-in account carries memory and history; your buyer's account does not. Checking while signed in measures your account.
What is worth doing, in order
- Let the crawlers in, by name. The tokens are not interchangeable: GPTBot governs training, OAI-SearchBot governs ChatGPT's search index, ChatGPT-User is the live fetch for one person's question, and Google-Extended governs Gemini's grounding independently of Googlebot. Blocking the wrong one costs visibility you meant to keep.
- Publish the page that answers the question. One question per page, answered in the first forty words, in a sentence that survives being cut out of the page.
- Give the page entry points. Thirteen or more subheadings, lists and tables. Each heading is a place a retrieval system can enter.
- Cite sources you do not own. A claim with an external source attached is a claim a model can corroborate.
- Get described somewhere you do not control. Directories, review sites, community threads. This is the half that cannot be bought with better copy.
- Measure with discovery questions, signed out, in every language you sell in, repeatedly. Anything else is not a measurement.
What does not work, or is not known to
Publishing llms.txt is not a lever. Ahrefs measured 137,210 domains carrying the file and found 97% received no requests for it. Every site in our crawl publishes one, which makes it a correlate of a site doing everything else. Publish it because it costs nothing; do not pay anyone for it.
Citation multipliers quoted by tool vendors are correlations. Figures like "2.8× more citations" come from companies selling citation tracking. They may be directionally right. They are not causal findings and should never be repeated to a client as fact.
Programmatic pages are a liability when they are thin. A template with a city name swapped in adds pages a model has no reason to prefer, and dilutes the vocabulary the rest of your site is trying to own.
Can this be guaranteed?
No. No one sells placement in an organic AI answer, and model answers change by day, country, account and search provider. Anyone offering a guaranteed position is either selling something else or does not know what they are selling. What can be sold honestly is a dated measurement, the evidence behind it, and the work that follows.
Questions
What is GEO in one sentence?
GEO is making a brand findable, extractable and corroborated so that generative answer engines — ChatGPT, Claude, Gemini, Perplexity, Copilot and Google AI Overviews — name it when a buyer asks a question.
Is GEO the same as AEO, LLMO or AI SEO?
In practice, yes. GEO, AEO and LLMO name the same work from different angles, and the category has not settled on one label. The distinction that matters is not between the acronyms; it is between being cited as a source and being recommended as an option, which are different outcomes and are measured separately.
How is GEO different from SEO?
SEO competes for a position in a list of links. GEO competes to be the sentence the model writes. A page can rank first on Google and appear in no AI answer, because ranking is a comparison and extraction is a lookup: the model either has a usable sentence about you or it does not.
Can anyone guarantee that ChatGPT will recommend my brand?
No, and you should stop talking to anyone who says otherwise. No one sells placement in an organic AI answer. Model answers change by day, country, account and search provider. What can be sold honestly is measurement, evidence and executed work.
How do I know if my brand is visible in AI answers right now?
Ask in a signed-out temporary chat, using questions that never contain your brand name, in every language your buyers use, and repeat on separate days. If you type your own brand name, or ask while signed in, you are measuring your account's memory rather than your brand's visibility.
Why does typing my brand name into ChatGPT not count?
A model handed a brand name will nearly always recognise it. We measured a site that scored high on branded recall and appeared in zero of thirty discovery answers across five model families and four languages. The branded question and the buying question have different answers.
Does publishing an llms.txt file make AI engines cite me?
There is no evidence that it does. Ahrefs measured 137,210 domains carrying the file and found 97% received no requests for it at all. Every site AI assistants recommend in this category publishes one anyway, which makes it a correlate of a site doing everything else, not a lever. Publish it because it costs nothing; do not buy it as a service.
How long does it take to change?
Technical access changes take effect as soon as the crawlers return, which is days to weeks. Being recommended is slower, because it depends on corroboration from sources you do not control. Anyone quoting a fixed timeline is quoting a sales cycle, not a measurement.
What actually correlates with being recommended?
We crawled 102 pages across the six sites ChatGPT returns for this category. All six publish a definition page for the category, state that no one can guarantee placement, name every AI crawler explicitly in robots.txt, and are either listed in a public agency directory or operate one. Their pages run 900 to 1,700 words with 13 to 17 subheadings, against a median of 410 words and 5.5 subheadings on ours before this page existed.
Do AI crawlers run JavaScript?
Assume not. If your text only appears after a client-side render, treat it as invisible to retrieval, and check your server logs rather than your analytics — GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot and PerplexityBot never appear in a client-side analytics tool.
Should I block AI crawlers to protect my content?
That is a business decision, not a technical one, and the tokens are not interchangeable. Blocking GPTBot withholds your pages from training. Blocking OAI-SearchBot removes you from ChatGPT's search results. Blocking Google-Extended removes you from Gemini's grounding while leaving Google search untouched. Blocking the wrong one costs visibility you meant to keep.
What does katman actually sell?
A dated measurement and the work that follows from it. We put buyer questions to five model families with live web search, record every answer verbatim with its sources and model version, and turn the gaps into specific published work. We never guarantee a position in any answer engine.
Sources
- Aggarwal et al., "GEO: Generative Engine Optimization" (arXiv, 2023) — the paper the term comes from.
- OpenAI, crawler documentation — what GPTBot, OAI-SearchBot and ChatGPT-User each do.
- Google, common crawlers — Google-Extended and GoogleOther.
- Anthropic documentation — ClaudeBot and Claude-User.
- llmstxt.org — the llms.txt proposal itself.
- katman AEO and GEO glossary — the vocabulary on this page, defined.
- frutti.ai, 30 observations, published in full — the measurement referenced above.
- AI traffic study — first-party session and revenue share by channel.
Measure your own site. The audit is free and reads only public pages. It checks crawler access, sitemap, schema and language structure, then writes the buying questions from your pages rather than your domain name.