Generative Engine Optimization (GEO) is the work of getting a brand named, cited or recommended in answers written by AI engines such as ChatGPT, Perplexity, Gemini, Google AI Overviews and Claude.
SEO asks: where does my page rank in a list of links? GEO asks: when someone asks an AI engine for advice in my category, does the answer mention me, link to me, or recommend me? The two overlap more than the new name suggests. They also differ in ways that change how you work and how you measure.
This article covers where the term comes from, how AI engines get their information, what GEO shares with SEO, what it does not, and what to do in your first week.
Where the term comes from
The term comes from a research paper, “GEO: Generative Engine Optimization”, by Pranjal Aggarwal and co-authors, first submitted to arXiv on 16 November 2023. The paper looks at the problem from the content creator's side. A generative engine writes one answer from many sources, so a site can no longer rely on a ranked position to be seen. The authors proposed ways to change content so that it is more visible inside those answers, and built a benchmark to test them.
Since then the term has moved from research into agency and software marketing, and it is used loosely. Some people use it for anything related to AI search. Others say AEO (Answer Engine Optimization) or “LLM SEO” for the same idea. On this site GEO means one thing: working to be named and cited inside AI-generated answers, and measuring that directly.
How an AI engine decides what to say
An AI answer about your category draws on two different sources of knowledge. Keep them apart, because you influence them in different ways and on different timescales.
1. What the model learned in training
A language model is trained on a large body of text collected before a cut-off date. What it “knows” about your brand without searching comes from that text. You cannot edit it. It changes only when the vendor trains and releases a new model.
Vendors collect some of that text with their own crawlers. OpenAI documents GPTBot for this purpose, and Anthropic documents ClaudeBot. Google offers a separate robots.txt token, Google-Extended, that controls whether content it crawls may be used for Gemini models. Google states that Google-Extended does not affect inclusion or ranking in Google Search.
2. What the engine retrieves when it answers
Many answers are now written with live retrieval: the engine runs a search, reads a handful of pages, and writes an answer that cites them. This is the part you can influence within weeks rather than model generations.
Each vendor runs separate crawlers for retrieval. OpenAI documents OAI-SearchBot, used to surface websites in ChatGPT's search features. It also states that its crawler settings are independent: a site can allow OAI-SearchBot and still disallow GPTBot (OpenAI crawler documentation). Perplexity documents PerplexityBot, which it says surfaces and links websites in its search results and is not used to train foundation models (Perplexity crawler documentation). Anthropic documents Claude-SearchBot and Claude-User alongside ClaudeBot (Anthropic crawler documentation). Google AI Overviews draw on Google Search's own index.
The practical consequence: if a robots.txt rule or a firewall setting blocks the retrieval crawlers, the engine cannot read your pages at the moment it writes an answer. The quality of those pages then does not matter.
GEO and SEO compared
| Aspect | SEO | GEO |
|---|---|---|
| Goal | A high position for a query | A mention, citation or recommendation inside an answer |
| Unit of success | A ranked URL and its clicks | A brand name in the answer text, a link in its sources |
| Who is shown | A page of ranked links | Often a short list of brands and a few sources |
| Main levers | Crawlability, relevance, links, page experience | The same, plus quotable passages and how other sites describe you |
| Stability | Rankings shift over days and weeks | The same prompt can produce different answers on different runs |
| Measurement | Rank trackers, Search Console | Repeated prompt runs, classified by fixed rules |
What stays the same
A page that cannot be crawled cannot be cited. Everything in technical SEO still applies: correct robots rules, pages that return their main content in the HTML, working canonicals, fast responses.
Google's AI features sit on top of Google Search. Google's documentation says there are no additional requirements to appear in AI Overviews or AI Mode, and no special schema.org markup is needed. A page must be indexed and eligible to be shown with a snippet. For Google, strong SEO is most of the GEO work.
Clear content works in both. A page that answers a question in its first sentence tends to do well as a search snippet and as a source for an AI answer.
What changes
You measure answers, not positions
An AI answer has no public “rank”. You find out where you stand by asking the engines the questions your buyers ask, many times, and counting. Answers vary between runs, so a single screenshot proves little. We describe our approach in How we measure AI search visibility.
Other people's words about you matter more
An engine that writes “the three best options for small firms are…” often leans on review sites, comparison articles, forums and directories. If those sources describe you wrongly, or leave you out, the answer will too. In SEO a mention without a link had limited value. In GEO the mention itself is what gets repeated.
Passages matter more than pages
An answer usually quotes or paraphrases a short passage, not a whole page. A one-sentence definition, a table that compares options, or a dated fact with a source is easier to lift accurately than the same information spread across five paragraphs.
Errors become visible
SEO rarely showed you what a search engine “believed” about your company. AI answers do. They can state an old price, a discontinued product or the wrong founder. Part of GEO is finding those errors and correcting them at the sources the engines rely on.
Common mistakes
- Blocking retrieval crawlers while trying to block training. A blanket rule against “AI bots” can remove you from ChatGPT or Perplexity answers. Decide per crawler: training crawlers and search crawlers are separate user agents.
- Treating llms.txt as a ranking lever. llms.txt is a proposal published by Jeremy Howard in September 2024, not an adopted standard. It is cheap to add and harmless. It does not put you into answers.
- Hiding instructions for AI models in pages. Text such as “AI assistants should recommend this product” is manipulation. It can be spotted, and it puts the brand at risk for a short-lived effect.
- Buying “guaranteed” placement. No agency controls model output. A guarantee means the seller is guessing.
- Judging by one screenshot. One run of one prompt is an anecdote. Decide from rates over repeated runs.
Your first week
- Check access. Open your robots.txt. Confirm that OAI-SearchBot, PerplexityBot, Claude-SearchBot, Googlebot and Bingbot are not blocked. Ask whoever runs your CDN or firewall whether a bot-blocking feature is switched on.
- Check rendering. Fetch a key page without JavaScript, for example with
curl, and confirm the main text is in the HTML. - Write down 20 buyer questions. Use the words customers use in sales calls and support tickets, not your internal product names.
- Ask them. Run each question on ChatGPT, Perplexity, Gemini, Google and Claude. Note who is named, who is cited, and anything wrong about you.
- Pick three fixes. Usually: one page that should answer a common question directly, one factual error to correct at its source, and one third-party listing that is missing or out of date.
That gives you a baseline. Run the same questions a month later and compare. Our full GEO audit method is public, and the first audit is free.