Ask ChatGPT, Perplexity, or Google's AI Overviews a question and the answer comes back with a short list of sources named underneath it. Being one of those named sources is the visibility that matters now. A blue-link ranking still helps, but a growing share of research starts and ends in the answer box, and the pages it cites are the ones that get seen. Our AI search citation benchmark shows which sources tend to win those mentions.
This is a guide to earning that citation. It covers the concrete, mostly technical work that decides whether an engine quotes your page or a competitor's. It is the companion to GEO citation tracking, which covers how to measure whether the work is paying off. Earning the citation comes first, so start here.
What is an AI citation?
Answer engines do not read your page the way a person does. They retrieve passages, meaning short self-contained chunks of text, and assemble an answer from several sources at once. When a passage from your page is clear enough to quote and trustworthy enough to rely on, the engine names you as the source. When it is not, the engine reaches for a page that is.
So the unit of GEO is not the page. It is the passage. A 3,000-word article can go uncited while a single well-formed paragraph on a thinner page gets the mention, because the paragraph answered the question cleanly and the article buried the answer.
How do I get cited in AI search results?
Get cited by clearing five signals in order: crawler access, an answer stated up front, specific facts, entity schema, and no snippet or freshness block. The order is not cosmetic. A crawler block or a JavaScript wall keeps you out before a word of your writing counts, so access comes first, and schema is worth nothing on a page an engine cannot read.
Once access is clear, the writing does the work. Answer engines pull passages rather than whole pages, so two clean sentences that stand on their own beat a long article burying the same point. Here is each signal in turn.
1. The crawler has to reach the page
If the AI crawlers cannot fetch your content, none of the rest matters. GPTBot, PerplexityBot, ClaudeBot, Google-Extended, and the others each need to be allowed in robots.txt, and your server and CDN have to serve them a real 200 response rather than a bot challenge or a JavaScript wall. Each publisher documents its own token: OpenAI lists GPTBot and its siblings in its crawler reference, and Google documents Google-Extended as a robots.txt control token with no user-agent string of its own, so a rule written against a user agent will not catch it. We regularly find sites blocking the exact crawlers they are trying to reach, usually through a security rule nobody remembered setting. Check your robots file and your edge rules before you touch anything else.
2. Lead with the answer
Engines lift the passage that answers the question, so put the answer first. Word order inside that passage may matter too: Google Research has found that models answer questions worse when the question reverses the subject-object order of the sentence the fact was written in, which we look at in subject-object entity order and what AI answers say. Open the section that targets a question with a direct two-sentence answer, then expand underneath it. A heading phrased as the question, followed immediately by a plain-language answer, is the single most useful structural change most pages can make. Answers buried three paragraphs down get skipped for pages that state them up front.
3. Make the facts trustworthy and specific
Models prefer sources they can rely on. Concrete numbers, dates, named methods, and claims that agree with what other reputable sources say all raise the odds of a citation. Vague or hedged copy gets passed over, because the engine has no reason to trust it over an alternative. If you make a specific claim, make it checkable.
4. Resolve who you are with schema and entities
Structured data helps an engine understand what your page is about and whether to trust it. Article, Organization, and where relevant LocalBusiness markup, with sameAs pointing at your canonical profiles, does more for AI grounding than it does for a traditional rich result. It ties the page to a known entity the model can place. Be clear about what that does and does not buy you: Google's own guidance on AI features and your website states that there is no special schema.org markup you need to add to appear in AI features, and no new AI text file either. Standard Article and Organization markup is not a shortcut into the answer; it is how the engine works out who is speaking. We go deeper on which types still earn their keep in schema markup that still matters.
5. Clear the freshness and snippet blockers
A restrictive max-snippet directive can quietly stop engines from quoting enough of your text to cite it. Google's robots meta tag specification defines that directive alongside nosnippet, data-nosnippet and noindex, and its AI features documentation names those same four as the controls that limit what its AI surfaces can show from a page. Set one without meaning to and you have opted out of the quote while still ranking for the query. Missing or stale published and modified dates get a page discounted. An llms.txt file gives AI systems a clean map of what to read first. None of these earns a citation on its own. Together they remove the reasons an engine would skip a page that otherwise deserves the mention.
How do I check whether my page can be cited?
Run every important page through the same short checklist. Each item is one of the five signals above paired with the check that tells you whether the page passes it, and most pages can be worked through in a few minutes:
- Crawler access. Fetch the page with a GPTBot user agent and confirm it returns a 200 with the real content in the HTML, not after a client-side JavaScript render. Read
robots.txtfor accidental AI-crawler blocks while you are in there. - Answer-first passages. Read the first two sentences under each heading. Do they answer the heading's implied question, or set up an answer that arrives three paragraphs later?
- Facts worth grounding on. Look for the numbers, dates and named methods a model could quote, and confirm each one is checkable against a source you can name.
- Entity resolution. Search your served HTML for Article and Organization schema, plus LocalBusiness where it applies, and a populated
sameAspointing at your canonical profiles. Establish who or what the page is about at first mention so the engine can place you. - Nothing blocking the quote. Check your meta robots for a restrictive
max-snippet, confirm the page carries an accuratedatePublishedanddateModified, and publish anllms.txtthat maps what to read first.
Anything that fails is work that is very likely holding back citations right now. If you would rather have the checklist graded for you, paste a passage into our free tool and it will score a page's citability against these same signals. Once the list is clear, watch which engines actually name your domain in GEO citation tracking rather than guessing from analytics.
How do I know if it worked?
AI citations are hard to see in standard analytics, so measure them directly. Monitor a set of real queries across the engines you care about and record whether your domain is named, which page earned it, and how that moves week over week. That is the subject of GEO citation tracking, and it is how you tell a real gain from a lucky screenshot.
This work and traditional SEO are not a trade-off. The foundations overlap almost entirely, and a page built this way tends to do better on Google too. We mapped where the two diverge in GEO vs SEO: what changed in 2026.
Where this leaves you
Most of the list above is a few hours of focused technical work per page, and it compounds once it is done. The reason it often does not get done is the same reason a lot of good SEO never ships: it sits in a report instead of going to production. We run this as part of a weekly loop on every domain we manage, on one flat plan you can cancel any month. The pricing page lays out exactly what that covers, and the comparison against an agency or an in-house hire covers where each model makes sense.