How to rank in AI Overviews and get cited by ChatGPT
What actually determines whether Google AI Overviews, ChatGPT and Perplexity cite your site. Crawler access, passage structure, entity clarity, and brand mentions.
Answer engines changed the unit of value. A ranked list of links rewards being on the page; a synthesised answer with three citations rewards being the passage worth quoting. Being adjacent to a good answer is now worth nothing.
This is the work variously sold as LLM SEO, ChatGPT SEO, or generative engine optimisation. The labels matter less than the fact that most reasons a site does not get cited are concrete and fixable, rather than the mystical "brand authority" explanations the discourse reaches for. Here they are, in the order worth fixing them.
1. Confirm the crawlers can actually reach you
This is the cheapest fix and the most commonly missed. Each answer engine uses named crawlers, and many sites block them without realising. Sometimes through a blanket robots.txt rule, sometimes through a CDN bot-protection setting nobody reviewed.
The ones that matter today are GPTBot and OAI-SearchBot (OpenAI), ChatGPT-User (fetches when a user asks), ClaudeBot (Anthropic), PerplexityBot, and Google-Extended (which controls Gemini training and grounding, separately from Googlebot).
Worth understanding: blocking Google-Extended does not remove you from AI Overviews, because those are grounded through normal Googlebot indexing. Blocking Googlebot does. These are frequently confused.
2. Make sure your content exists without JavaScript
Most AI crawlers do not execute JavaScript. If your content only appears after a client-side fetch, a large share of them see an empty shell. And unlike Googlebot, they will not come back later with a renderer.
The test takes ten seconds: fetch your page with curl and read what comes back. If your key content is missing from the raw HTML, no amount of content optimisation will help until that is fixed.
The specific trap
A component that fetches its content from an internal API route on mount, where that route is disallowed in robots.txt. The content is then invisible to non-rendering crawlers AND at risk of failing to load for the ones that do render.
What to do instead
Server-render the content, or pass it as props from a server component. If it matters for citation, it belongs in the initial HTML response.
3. Write passages that can be quoted rather than paraphrased
This is where most sites lose. Language models lift self-contained passages and paraphrase everything else. And a paraphrase usually drops the attribution, which is the entire value you were competing for.
A citable passage answers one question completely, without requiring the surrounding page for context. In practice that means roughly 130 to 170 words: long enough to contain a claim, the mechanism behind it, and a qualifier; short enough to lift whole.
The common failure is a forty-word FAQ answer. It reads well, it says something true, and it is too thin to quote. So the model absorbs the fact and cites someone more thorough.
Lead with the direct answer
First sentence answers the question outright. Models and readers both reward this, and it survives being extracted out of context.
Then add the mechanism
Two to four sentences of how or why, with specifics. Numbers, thresholds, decision rules. This is the part that makes a passage worth quoting over a competitor’s.
Include the qualifier
When it does not apply, or what it depends on. Models cite hedged, accurate passages over confident, absolute ones, because the hedge is information.
4. Make your structured data describe one entity
Most sites re-declare their organisation independently on every page, with no shared identifier and inconsistent properties. To a system trying to build an entity graph, that looks like a dozen loosely related businesses rather than one.
The fix is a single canonical Organization node with a stable @id, declared once, with every other node (services, articles, breadcrumbs, the person who founded it) referencing that @id rather than restating the details. Add a WebSite node as the anchor and link pages to it.
This does not produce a visible rich result. It changes whether a model can confidently resolve "who is this" into a single entity it can reason about and attribute to.
5. Get mentioned where these systems actually retrieve from
The slowest lever and the most durable one. Answer engines lean heavily on sources with strong topical consensus (established publications, community discussion, and review directories) rather than on raw backlink counts.
For a services business the practical list is short: relevant industry directories with real reviews, genuine participation in communities where your buyers already are, and being quoted in publications your field reads. Video is disproportionately effective and disproportionately neglected.
This is also the part with no shortcut. Anything promising it in a month is describing a paid placement, not a citation.
What about llms.txt?
Worth being blunt: the evidence does not support the hype. Across large-scale monitoring of AI crawler traffic, requests for /llms.txt are a rounding error against requests for ordinary HTML, and Google has said on the record that it does not use the file.
It costs almost nothing to publish one and it does no harm, so keep it if you have it. But do not build a strategy on it, and be sceptical of anyone selling it as the AI-search unlock. The things above are what actually move citation rates.
Common questions
Does blocking Google-Extended remove me from AI Overviews?
No, and this is the most consistently misunderstood point in the whole area. Google-Extended is a control over whether your content is used for Gemini model training and for grounding in some Gemini surfaces. AI Overviews in Search are grounded on Google’s ordinary search index, which is built by Googlebot. So blocking Google-Extended leaves you eligible to appear in AI Overviews, while blocking Googlebot removes you from AI Overviews and from ordinary search results together. If your goal is to appear in AI Overviews but not to contribute to model training, blocking Google-Extended while allowing Googlebot is the configuration that expresses that. If your goal is to disappear from AI answers entirely, there is no clean way to do it without also losing the organic search traffic that comes with the same index.
How long should a passage be to get quoted?
Roughly 130 to 170 words is the range that gets lifted intact most reliably. Below about 100 words a passage usually lacks enough substance to stand alone, so the model absorbs the fact and paraphrases it, which means no citation for you. Above about 200 words it tends to get summarised rather than quoted, which also weakens attribution. The shape matters as much as the length: answer the question directly in the first sentence, then give two to four sentences of mechanism with specific numbers or decision rules, then a qualifier about when it does not apply. That structure survives extraction, because it does not depend on the surrounding page to make sense. Applying this to an existing FAQ section is one of the highest-return content edits available, because the pages already exist and already rank.
Can I track whether AI engines are citing me?
Partially, and less precisely than conventional rank tracking. AI visibility tracking is a real and growing category, but there is no equivalent of Search Console for these systems. None of them publish impression or citation data. What you can do is monitor AI crawler hits in your server logs, which tells you whether GPTBot, ClaudeBot and PerplexityBot are actually fetching your pages and which ones. You can also run a fixed set of buyer questions across the major assistants on a schedule and record whether you are named, which gives you a rough but genuine trend line over months. Several commercial tools now automate that second approach. Treat the output as directional rather than precise: results vary by phrasing, region and user history, so a single check tells you almost nothing while a consistent monthly sample tells you something real.
Services related to this guide
Tell us what you sell
A written audit covering AI crawler access, whether your key pages are structurally citable, how your brand currently surfaces across ChatGPT, Perplexity and Google AI Overviews, and the specific changes that would move it.