SeeWAIS
2026-11-15
SeeWAIS Research · Independent research on how AI answer engines name and cite businesses.

GEO definition: what generative engine optimisation actually means

A plain definition of GEO — generative engine optimization — what counts as a generative engine, the three parts any working definition needs, and how to test one against your own business.

GEO stands for generative engine optimization — generative engine optimisation, in British spelling. It is the work of getting a business named or cited inside an AI-generated answer, rather than ranked in a list of links. That is the whole definition. Most of what is written about GEO beyond that sentence is either scope, method, or a sales claim wearing a definition's clothes.

GEO definition: generative engine optimization is the practice of making a business visible inside answers produced by AI assistants — named when a buyer asks a relevant question, and cited among the sources the answer was built from.

Two words in that sentence carry the weight. Named means the business appears in the answer a real buyer would read, not that a tool found its domain somewhere in a crawl. Sources separates being mentioned from being the material the answer actually rests on. A business can be named in passing and cited nowhere; the second is the stronger position. The vocabulary is set out at greater length on what GEO means, and the five labels the market uses for the same work are compared in our piece on LLM visibility, GEO, AEO and AI SEO.

What counts as a generative engine

Search engine Generative engine
Output A ranked list of links A written answer with sources
Where the reader looks The list The answer
Unit of competition Position Mention or citation
What decides it Ranking systems Retrieval, the model, and the sources it trusts

A generative engine is a system that answers a question in prose, with sources attached, instead of returning a list of links for the reader to sort. Google's AI Overviews and AI Mode, ChatGPT with web search, Gemini with grounding and Claude answering from memory all qualify: the output is an answer, and there is no page two.

The engines are not one thing. ChatGPT and Gemini search the live web; Claude in our panel answers from what the model holds in memory; Google's AI surfaces are built on the same index as ordinary search. The pairing that matters most for a diagnosis is memory against search. If an assistant names a business only when it can look it up, the model does not know the business — it found it. Those two states need different work, and our method is built to tell them apart.

Google's own guidance is the least sentimental definition on offer. Its AI optimization guide says that, from Search's perspective, optimising for generative AI search is still SEO (Google Search Central, accessed 2026-10-27). Its page on AI features states there are no additional requirements, "nor other special optimizations necessary" beyond being indexable and eligible for a snippet (Google Search Central, accessed 2026-10-27). Both statements describe Google's surfaces specifically. They do not describe what ChatGPT does when it retrieves through its own crawler, which is why a site can be perfectly healthy in Google and absent from ChatGPT on the same day.

The three parts of a working definition

Any definition of GEO should survive three tests.

1. The unit is a mention, not a position. A page can rank eighth and be cited, or rank first and be absent. Positions and mentions are different objects counted on different surfaces, and neither number converts into the other. The comparison of what changes sets out the four differences that matter.

2. The sources are wider than your own site. In a ranked list, your page is the thing being judged. In a generated answer, your page is one candidate among the sources the model retrieved — map listings, directories, review sites, trade press and competitors' pages included. In the local market we published, the overviews leaned first on Google's own map listings, then on competing businesses' pages, then on the review and directory sites the answers depend on; the measured business's own site was not among them (how those answers get built).

3. The measurement needs repetition. The same question asked twice can produce two different lists, and grounded answers also depend on what the live web returned that second. A single run is an anecdote, not a measurement. Our method asks 30 real buyer questions across five live AI surfaces — 25 of them never mention the brand — records 146 answers (96 graded, 50 mechanically checked) and 242 API calls, and asks the three most important questions three times each so a finding can be labelled stable or not. METHOD v1.3 is published and frozen.

A definition that cannot be tested is a slogan. GEO is testable: ask real buyer questions that never mention the brand, on several surfaces, and record whether the business is named and cited in each answer.

A data point that complicates the acronym

The flagship deliverable of many GEO packages is an llms.txt file, a plain-text introduction to a business served at its domain root. Our scan of 34 head-of-market websites found 22 serve one. The split is the useful part: 13 of 16 software companies, against 4 of 8 direct-to-consumer brands. Companies whose customers are developers publish developer files, and the file tracks the industry a company sells into rather than whether assistants name it. The full file scan and the benchmark record show the rest of the pattern, including the median readable homepage serving 1,059 words before any JavaScript ran — with four serving under 300. A page that renders beautifully can arrive at an AI crawler almost empty.

That is a reminder about definitions in this field: the thing being sold is often not the thing being measured. Having a file and having written one are different states, and neither is proof of being named.

What the definition should exclude

  • Geography. The geo- in GEO is an acronym, not a prefix about maps, localities or geospatial software.
  • A guarantee. No vendor controls what an assistant says, and no honest provider promises a mention, a rank or a timeline.
  • A single score. A number from one run, across an undisclosed question set, measures that run. It does not describe a standing position.
  • Paid placement. There is no paid slot in a recommendation list. If one appears, it will be labelled as an advertisement.
  • A file, a tag or a technique. No schema markup, keyword or llms.txt file makes an assistant recommend a business. The prerequisites are plainer: a crawlable site, clear facts and corroboration elsewhere.

How to test the definition on your own business

Open a private browser window and ask an assistant three questions a buyer would ask — "best [your category] in [your city]", the same question with a constraint such as "open on Saturdays", and a comparison against a named competitor. None of the three should mention your business. Record the answers word for word: whether the name appears, whether anything said about the business is wrong, and which sources the answer cites. The free check shows the question set a full measurement would use.

If the name is missing, that is the finding the definition is about — and a DIY evening is enough to see it may be there. What three questions cannot tell you is how often, on which surfaces, and whether the result holds when the same question is asked again. That needs a larger question set and repeats. Our report measures 30 real buyer questions across five live AI surfaces, producing 146 recorded answers (96 graded, 50 mechanically checked) and 242 API calls, and costs $4.99 once, with no subscription. A complete real run is public, and the cost guide compares a one-time report with human audits and monthly retainers.

The practical test is not whether a vendor's definition sounds sophisticated. It is whether the definition names something you can observe and re-observe. If it does, you can buy the observation once, act on it, and check again later.

Common questions

What does GEO stand for? GEO stands for generative engine optimization (optimisation in British spelling): the work of getting a business named or cited inside an AI-generated answer rather than ranked in a list of links.

Is GEO the same as SEO? No. SEO competes for position on a list of links; GEO competes to be named inside an answer that mostly has no list. The two share foundations — crawlable pages, accurate facts, consistent listings — but they are measured on different scoreboards.

Does GEO mean geography? No. Here GEO is an acronym for generative engine optimization. The geo- in geography, geospatial data and geo-location is a different word, and search results about maps or location software have nothing to do with it.

Can GEO guarantee that an AI assistant names my business? No. Answers vary between runs, models change, and no vendor controls what an assistant says. Google's own documentation states its AI features need no special optimisations beyond being indexable, and anyone promising a guaranteed mention is selling certainty that does not exist.

Sources

  • Google Search Central, "AI optimization guide" — https://developers.google.com/search/docs/fundamentals/ai-optimization-guide (accessed 2026-10-27)
  • Google Search Central, "AI features and your website" — https://developers.google.com/search/docs/appearance/ai-features (accessed 2026-10-27)
  • llmstxt.org, the llms.txt proposal — https://llmstxt.org/
  • SeeWAIS, "Should you add an llms.txt file? What 34 measured websites actually do" — https://seewhataisays.com/articles/llms-txt-guide
  • SeeWAIS benchmark record — https://seewhataisays.com/benchmark
  • SeeWAIS METHOD v1.3 — https://seewhataisays.com/method
  • SeeWAIS published sample report — https://seewhataisays.com/sample

What does AI say about your business?

The free check reads your site and shows you the exact questions we would ask — before you enter any card details. The full report is $4.99 once.