Blog
Generative engine optimization (GEO): what you can control
Generative engine optimization (GEO) is the work of making sure search engines that answer with AI can find your content, understand it and cite it when they put an answer together. The term comes from a 2023 academic study presented at the KDD 2024 conference. In practice it rests on the same ground as SEO: crawlers can get in, the text is on the page, and it answers clearly.
What the search engines say
- Google says SEO best practices still apply to AI Overviews and AI Mode, and that there are no additional requirements or special optimizations needed to appear in them. It doesn’t ask for new files or special structured data either.
- Microsoft, in guidance from the Bing team from October 2025, says there is no secret sauce that guarantees selection in its AI answers. It recommends questions with direct answers, sentences that make sense out of context and descriptive headings, and warns that content hidden in tabs or expandable menus may not be read.
- OpenAI asks site owners not to block OAI-SearchBot if they want to appear in ChatGPT search answers.
What you control and what you don’t
| What | Do you control it? | How |
|---|---|---|
| Whether search crawlers get in | Yes | With robots.txt, by the name of each crawler: it’s an instruction these companies say they respect, not a technical block |
| Whether your content is used to train models | Yes, separately | Through the training crawlers, which have different names |
| Whether the answer is in the page text | Yes | Visible text in the HTML, not only in images, PDFs or tabs that open on click |
| Whether the text can be quoted | Partly | Passages that stand on their own, with your own figures and linked sources |
| Whether a given answer cites you | No | It depends on the query, the system and each person; the same question can get a different answer |
Search crawlers and training crawlers
The main companies separate the crawler that feeds their search from the one that collects content to train models. You can allow one and block the other.
- OpenAI: OAI-SearchBot surfaces sites in ChatGPT search; GPTBot collects content to train its models; ChatGPT-User makes visits that people ask for and doesn’t decide what appears in search. OpenAI notes that each setting is independent and that a robots.txt change can take about 24 hours to apply.
- Anthropic: Claude-SearchBot indexes to improve search, ClaudeBot collects content that could be used for training, and Claude-User handles requests made by people.
- Perplexity: PerplexityBot surfaces and links sites in its results and isn’t used to train models; Perplexity-User visits pages when someone asks a question.
- Google: Googlebot controls crawling for Search, including its AI features. Google-Extended decides whether your content is used to train Gemini and to ground answers in the Gemini app and on Vertex AI. According to Google, it doesn’t affect inclusion or ranking in Google Search.
That last detail changes the decision. Blocking Google-Extended doesn’t remove you from Google’s results, but it can stop the Gemini app from using your content to ground its answers. If you want to limit what Google shows from your pages in Search, the tools are different: nosnippet, data-nosnippet, max-snippet or noindex, which we explain in our guide to the robots meta tag.
A real example: HeyMark’s robots.txt allows all of these crawlers, including the training ones. The reason is written in the file itself: we sell an MCP integration with AI assistants, so blocking them would be inconsistent. A publisher that lives off its articles could reasonably decide the opposite. What doesn’t work is blocking a search crawler and expecting your content to be used in those answers: OpenAI notes that, in that case, the site can only appear as a navigational link.
Writing passages that can be quoted
A system that builds an answer takes fragments. If the information is split between a slogan, a closed tab and an image, there is nothing to take.
Fictional case. A pottery studio has this sentence on its course page:
Our courses adapt to you. Message us and we'll tell you everything!
Someone asks an assistant how long a beginner pottery course in their city lasts and what it costs. That page doesn’t answer. The studio rewrites it:
The beginner pottery course runs for eight two-hour sessions, includes materials
and kiln firings, and costs 180,000 Chilean pesos including VAT. Each group has
eight spots; the next one starts on December 3.
The sentence stands on its own, its figures can be checked on the same site, and it matches how people ask. If the page makes a claim about the world, such as the firing temperature of a glaze, link the source.
That advice matches the original study. On a set of 10,000 queries, its authors tested nine rewriting methods and found that adding citations to sources, quotations from relevant sources and statistics increased a source’s visibility in answers by over 40%. In a test on Perplexity, the improvement reached up to 37%. Repeating the words of the query, as in old-style keyword stuffing, didn’t work. These are lab results with 2023 models, and the authors themselves show the effect varies by topic, so treat them as a writing guideline, not a promise.
How to measure it
- Google: traffic from AI Overviews and AI Mode is counted in Search Console, under the “Web” search type, along with the rest of the results. It doesn’t appear as a separate report. Our explanation of impressions, clicks and CTR in Search Console helps you read it.
- ChatGPT: the publisher FAQ that OpenAI kept until October 2026 (archived copy) says ChatGPT adds
utm_source=chatgpt.comto the links in its answers. Check in your analytics tool whether those visits arrive with that parameter and separate them, like any other UTM parameter. - Crawlers: server logs show whether OAI-SearchBot, Claude-SearchBot or PerplexityBot visit your pages. If they don’t arrive, check robots.txt and your firewall before rewriting copy.
Being cited varies from one query to the next and from one day to the next. A manual test in ChatGPT shows which sources one answer used; it doesn’t measure your position. If you’re wondering whether you also need a file for language models, we cover it in llms.txt: what it is and when it’s worth creating.
Featured snippets are a different Google Search feature, older than the AI ones, with their own controls: see what a featured snippet is.
Plan your next post in HeyMark.
Keep the idea, review the draft with your team, and see how it performed in the accounts you connected.