Short answer: GEO, short for generative engine optimization, is the work of making a brand eligible to be named, cited and recommended in answers that AI assistants write, such as ChatGPT, Perplexity, Gemini, Claude and Grok, and of measuring honestly whether that is happening. It overlaps with SEO on the technical basics, but the output is different: an answer engine composes one reply from passages it selected, so there is no list of ten positions to climb, only whether you are in the answer, how, and on whose evidence.
Nothing in GEO guarantees an outcome. Assistants decide what to name answer by answer, and nobody outside the engine controls that. What GEO controls is eligibility, and the quality of the evidence about you that an engine can find.
The word "generative" is the important part. A search engine retrieves documents and lets you choose; a generative engine retrieves documents and then writes. Everything distinctive about GEO follows from that second step: the writing is done from passages rather than pages, it names a handful of brands rather than listing ten results, and the reader often never sees the sources at all.
What GEO is trying to change
"Being in AI answers" is really three separate outcomes, and they have different causes:
- Mentioned: your name appears in the answer. This depends mostly on the engine knowing clearly which entity you are.
- Cited: one of your URLs is marked as a source. This depends on the engine being able to fetch your page, read it and lift a passage from it.
- Recommended: the answer actively suggests you for the reader's situation. This depends on comparable facts about you that independent sources confirm.
Your page can be cited as a source while the answer recommends a competitor, and you can be recommended without a single link. Good GEO work, and any report about it, keeps the three apart. The split also tells you who to involve: citations are mostly a site and content problem, while mentions and recommendations depend heavily on what other sites say about you.
GEO and SEO: what changes and what does not
GEO does not replace SEO. Pages that cannot be crawled or indexed fail at both, and much of the technical groundwork is shared. What differs is the unit of output, the agents involved and what counts as success.
The table puts the same differences in rows:
| SEO | GEO | |
|---|---|---|
| What the user sees | A ranked list of links | One composed answer, sometimes with sources |
| Unit that gets used | A page | A passage lifted from a page |
| Who fetches | Mainly search engine crawlers | Several agents per vendor, split between training, search index and user-triggered fetches |
| JavaScript | Depends on the search engine | Most AI crawlers do not execute it |
| What you measure | Position, impressions, clicks | Presence, citation and recommendation, per engine |
| Traffic | The click is the goal | Most answers are read without a click |
The last row matters for how you judge results. Being the cited source is often a brand outcome rather than a traffic outcome, so it has to be measured on the answer side, not in web analytics. Our GEO vs SEO guide goes through the overlap in detail.
How answer engines decide what to say, as far as anyone can tell
No vendor publishes how its assistant chooses which brands to name, so be wary of anyone who claims to know the formula. What is documented, or measurable, is this:
- Access is split by purpose. OpenAI documents GPTBot for training, OAI-SearchBot for the index behind ChatGPT's search features, and ChatGPT-User for fetches a person asked for. Anthropic and Perplexity make similar splits. Google's equivalent control, Google-Extended, is not a crawler at all but a usage control over content Googlebot already fetched (OpenAI's crawler documentation, Google's).
- Engines behave differently. In our own benchmark one model named 3.5 providers per answer and another 2.1. A result on one engine tells you little about another.
- Categories differ. In our Q3 2026 benchmark of 12 categories across five models, the three categories with a clear default answer were all product categories. Every local-service category we measured had a leader below 13%.
- Answers vary between runs. The same question can name different brands on different days even when nothing changed, which is why single checks prove little.
Category shape changes the job. Where one brand already appears in a quarter or more of answers, being present is not enough: the realistic aim is to become the answer to a specific sub-question the incumbent answers badly, such as a budget, a constraint or a use case, rather than to compete on the head term. Where no brand dominates, as in the local-service categories we measured, the ground is open and cheaper to win. Some vendors also leave gaps you cannot work around: xAI does not publish a documented crawler user agent, so Grok cannot be allowed or blocked by name, and the only lever is that your pages are readable to general crawling.
The five areas of GEO work, in order
Each area is a precondition for the next, so the order matters. The first two are technical and usually belong to a developer; the third belongs to whoever writes the pages; the fourth reaches outside your site; the fifth is how you find out whether any of it worked.
- Crawler access. Allow the search and user-triggered agents you want, decide separately about training agents, and test with the real user agent, because a firewall rule returning 403 is a block whatever robots.txt says. See our AI crawler access guide.
- Readable HTML. Put the content that matters in the initial HTML. A page that is complete only after client-side rendering is an empty shell to most AI crawlers, and a 200 response with an empty body is a failure.
- Extractable writing. Answer first under each heading, name things in full, and state facts with units and dates, so a passage still makes sense when it is lifted. Writing for extraction shows how.
- A clear entity, corroborated elsewhere. Use one name and one description everywhere you control, add Organization markup with sameAs links, and make sure independent sources state the same facts. A model needs the same fact from independent places before it states it confidently. Third-party corroboration covers where to look.
- Honest measurement. A fixed set of buyer questions, run on each engine, reported separately before any average.
Measuring GEO without fooling yourself
Measurement is where GEO programmes most easily go wrong, because the easiest numbers to produce are the most flattering. These are rules we adopted after getting each one wrong in our own product:
- Exclude questions that contain your brand name. An LLM-generated question set once contained the brand name in 13 of 15 questions, and the visibility score read 13% when the honest value was zero.
- Count only sources marked in the answer text. Counting everything the engine retrieved overstated citation figures by two to four times.
- Count competitors the same way you count yourself. Asymmetric detection once showed a brand at 31% share of voice and rank two; measured symmetrically it was 0% and last.
- Never show unmeasured as zero, and never compare runs made with a different method. Enabling web search once moved a site from 3% to 13% while the site did nothing.
One more rule sits above the others: report each engine on its own before any average, and state the question set, the engines and the date range before any finding. A percentage without its denominator is not a finding. Our GEO measurement framework sets these out as a working method.
Common GEO misconceptions, and where to start
- "GEO is a file you add." llms.txt is cheap, and Shopify now serves /agents.md by default and mirrors it to /llms.txt. None of the crawler documentation cited here says the file is read when composing answers. Treat it as housekeeping.
- "Schema markup is the lever." Markup helps remove ambiguity about who you are. None of the crawler documentation cited here says it raises the chance of being cited.
- "You can rank first in ChatGPT." There is no first place in a composed answer. Anyone selling a guaranteed position is selling something the engines do not offer.
- "Block AI crawlers to protect content, it costs nothing." Blocking training agents is a defensible choice. Blocking search agents such as OAI-SearchBot removes you from those answers.
Where to start
Read your robots.txt for each agent by name, fetch your three most important pages without a browser and check the key sentences are there, then write ten questions a buyer would ask without knowing your brand exists and run them on each engine you care about. That baseline tells you which of the five areas to work on first. The glossary entry for GEO has the short definition, and if you would rather not run the checks by hand, InsightWonder automates the measurement side.
Sources
- OpenAI, overview of OpenAI crawlers.
- Google, Google's common crawlers, including Google-Extended.
- Shopify, llms.txt.liquid.
- InsightWonder Q3 2026 benchmark (12 categories, five models) and our measurement incident records, July to September 2026.