How to Get Cited by ChatGPT

When a prompt triggers search, ChatGPT issues one or more queries against its own retrieval layer, receives a shortlist of candidate URLs, fetches a subset of them, and composes an answer from the passages it extracted. Citation happens at the passage level: the model attaches a source to a specific claim, not to the answer as a whole. That means a page can be fetched, read, and still not appear in the citation list because nothing in it was quotable enough to anchor a claim.

The shortlist stage looks a lot like classical search. Query-document relevance, freshness and site-level signals decide who gets fetched at all. This is why sites with no organic presence rarely appear in ChatGPT answers: they never enter the candidate pool. Everything an SEO would do to become retrievable — unique content, clean canonicals, a working sitemap, internal links from pages that already rank — is the entry fee.

The composition stage is where AEO takes over. Among fetched pages, the model prefers passages that state something definite and survive removal from their context. A sentence reading 'the answer depends on your situation' anchors nothing. A sentence reading 'a mid-size support team replacing 40 percent of tier-one tickets typically breaks even in four to seven months' anchors a claim, and the URL attached to it gets the citation.

OpenAI operates three documented user agents with different jobs. GPTBot collects content that may be used for model training. OAI-SearchBot builds the index that powers ChatGPT search. ChatGPT-User fetches a page in real time because a user's prompt referenced it or the model decided to open a link mid-conversation. They obey robots.txt independently, and their rules are read literally by user-agent name.

The common mistake is a blanket rule. Publishers who wanted to opt out of training frequently blocked all three, then discovered months later that they had also removed themselves from every ChatGPT answer. If your goal is visibility, allow OAI-SearchBot and ChatGPT-User at minimum. The training decision for GPTBot is a separate business question and can be answered either way without touching your citation prospects.

Write the rules explicitly rather than relying on a wildcard. Several AI agents ignore or mis-handle wildcard groups when a named group exists elsewhere in the file, and an explicit allow block removes any ambiguity for the operator and for you. Run our readiness audit on your own URL if you want the file checked automatically against every agent name we track.

The pattern is consistent across thousands of observed citations. Near the top of the page sits a direct answer, roughly forty to sixty words, phrased as a complete statement that reads correctly with no preceding sentence. Below it sits a short list of standalone claims. Then the body, broken by question-form headings that match how people actually phrase prompts, with numbers exposed in tables or labelled lists rather than buried mid-paragraph.

Length is not the driver, but there is a floor. Pages under roughly three hundred words of substantive text rarely accumulate citations, because they offer too few distinct claims to anchor anything. Past about fifteen hundred words the return flattens: what matters is the density of quotable statements, not the word count around them.

One structural habit pays out repeatedly: answer the question in the heading, immediately, in the first sentence after it. Models segment on headings. If the paragraph following an H2 opens with context rather than the answer, the extracted passage begins with context, and a passage that begins with context is much less likely to be selected.

It affects selection in the categories where the model is most cautious. For prompts touching money, employment, health, law or safety, passages attached to a named author with a resolvable profile are preferred over anonymous equivalents. The mechanism is not mystical: the retrieval and ranking layers consume the same provenance signals classical search does, and the model is tuned to prefer sources it can attribute.

Making authorship legible costs one afternoon. Name a real person on the page, link that name to a profile URL that actually resolves, state their relevant experience in plain sentences, and mirror the same person in Article or WebPage schema as the author, with the organisation as publisher. Add a stated review date, and update it only when the page genuinely changes.

Avoid the anti-pattern of a fabricated expert. Engines increasingly cross-reference author entities against the wider web, and a byline with no independent footprint contributes nothing while creating a trust liability if a reader checks. A real person with a modest footprint beats an impressive-sounding invention.

Search Console will not show it. Use three direct measurements instead. First, run your target prompts against ChatGPT on a schedule and record whether your domain appears in the citation list — this is tedious and irreplaceable. Second, watch analytics for referrals from chatgpt.com, which do appear as a distinct source when a user clicks a citation. Third, watch server logs for OAI-SearchBot and ChatGPT-User hits, which tell you retrieval is working even before citations appear.

Keep a written map of prompt to page. One question, one intended page. When a check fails, the map tells you which of two very different problems you have: the engine ignored your site, or the engine cited the wrong page of yours. The first calls for authority and structure work, the second for consolidation and internal linking.

Expect volatility. Citation sets are re-selected per query and change as competitors publish. A monthly cadence gives you a trend; a daily one gives you noise and a strong urge to make changes that were never warranted.