How to get cited by ChatGPT: what the evidence supports

No one can promise that ChatGPT will cite a particular brand, and anyone who does is promising something they cannot control. What published studies do offer is a reasonable picture of where ChatGPT's citations tend to come from, which popular tactics have little evidence behind them, and how to tell whether anything you changed made a measurable difference. This guide sets that out, with every figure linked to its source and the limits of each study stated.

The short answer

The evidence points in a fairly unglamorous direction:

• Be in the general search index. Most of what ChatGPT cites comes from pages a search engine already indexes.
• Be on the third-party pages that answer engines read. Comparison pages and "best of" lists and, in the UK, news sites make up a large share of what gets cited.
• Keep the crawlers able to reach you, and treat a crawler visit as a fetch, not a citation.
• Do not lead with llms.txt or schema. Neither has shown a clear effect on citations in the studies available (llms.txt, schema).
• Measure over repeated runs. A single answer is not a result; a rate across many runs is steadier.

The rest of this guide explains each point and where the evidence is thin.

ChatGPT mostly cites pages the search index already holds

In an analysis sponsored by Ahrefs and published by Search Engine Journal, 88.5% of ChatGPT's citations came from the general search index. Ahrefs sells an AI visibility product, so treat the figure as vendor research rather than an independent audit. Its direction matches the rest of the evidence, though: when ChatGPT searches the web, it largely draws on pages that conventional search engines already know about.

That makes the first step ordinary technical SEO. A page that is not indexed, is blocked, or does not answer the buyer's question in plain words is unlikely to be the one an answer engine finds and cites. For each question that matters to a client, check three things:

• a page exists that answers the question directly, in the words a buyer would use;
• that page is indexed and not blocked to crawlers;
• the claims on it are ones the client will stand behind if an answer repeats them.

Third-party pages carry much of the weight

Answer engines do not only cite the brand's own site. In the same Ahrefs-sponsored analysis, "best of" list pages accounted for 43.8% of the pages cited. A study of UK banking questions published by Genezio found that the ChatGPT app leaned on comparison sites far more than the API did, 32.2% against 11.0%, which is one reason measurements taken through the API and in the app can disagree. ChatGPT API vs the ChatGPT app covers that difference in detail.

News coverage matters more in the UK than in the US. Semrush's 2026 AI Visibility Index, as reported by MarTech Edge, found that news sites made up 29.5% of ChatGPT's top-100 citations in the UK, against 19.6% in the US. Semrush sells AI visibility tools, so this is vendor research too.

For an agency, the practical reading is that earned media and accurate listings are part of the job:

• Find the comparison and "best of" pages that appear for the client's category, and check what they say about the client, if anything.
• Correct factual errors on pages the client can influence, and make sure the client's own pages give those authors something accurate to cite.
• For UK clients, treat trade and national press coverage as part of AI visibility work rather than a separate PR silo.

None of this guarantees a citation. It raises the odds that, when an engine does look, it finds accurate material about the client.

What the evidence says about llms.txt and schema

Two tactics are sold widely as ways to "optimise for AI": an llms.txt file and structured data (schema markup). The evidence for both is weak so far.

• llms.txt. An analysis of 300,000 domains found no clear effect of an llms.txt file on AI citations.
• Schema. In the Ahrefs-sponsored analysis, the measured changes in citations against controls were +2.2%, +2.4% and -4.6%: small, and not consistently in one direction.

That is not proof that either is harmful, and structured data has other uses in ordinary search. But neither should be sold to a client as the lever that gets them cited. If an audit or proposal leads with them, ask what evidence sits behind the expected effect.

Let the crawlers in, but do not mistake a fetch for a citation

AI companies run their own crawlers, and a robots.txt file can allow or block them. Check that the client's robots.txt rules for GPTBot, PerplexityBot, ClaudeBot and Google-Extended say what the client actually intends, rather than what a template or a security plug-in decided years ago.

Server and CDN logs show whether those crawlers reach the pages that matter. Be careful with the conclusion: a crawler fetching a page shows that the page was retrieved, not that any answer used it, quoted it or cited it. AnswerLift can import a summary of a client's own logs for the AI crawlers (up to 5,000 rows per import) and labels every crawler hit as a fetch, never as a citation, for exactly this reason.

How to tell whether anything changed

The hardest part of getting cited is proving that it happened, because AI answers vary from one run to the next even when nothing about the client has changed.

• In a study by SparkToro, co-run with Gumshoe (which sells an AI visibility tool), the same list of brands came back less than once in 100 runs, and in the same order less than 0.1% of the time, across 2,961 runs of 12 prompts.
• An academic study that asked 80 questions 20 times each found that 44.5% of the pages ChatGPT cited appeared in only one of the 20 runs.

SparkToro's own write-up concludes that rates measured over many runs are steadier than positions. Ahrefs, which also sells a tracking product, puts it this way: "Think of AI tracking less like rank tracking and more like polling". A review of how AI visibility tools collect data makes a related point: whether a tool uses the API or the consumer interface matters less than whether it discloses which, and one run a week is a poll.

So a defensible before-and-after needs:

1. The same question, worded the same way, recorded with its version, so you are comparing like with like.
2. The same engine and mode, because an answer from a model with web search switched on is a different thing from one answering from model knowledge.
3. Repeated runs, reported as a count: named in k of n runs, with the dates. A single answer that named the client proves only that it can happen.
4. The cited domains, not just whether the brand was named, so you can see whether the pages you worked on are among them.
5. No causal claim. If the client was named more often after a change, say that, and say that answers vary on their own. Do not say the change caused it.

Why AI answers change every time covers the variance evidence, and how to report it to a client, in more detail.

Where AnswerLift fits

AnswerLift does not publish content or change client sites; it records what the answer engines say. A Visibility Sprint asks the client's buyer questions of the models behind ChatGPT, Claude, Gemini and Perplexity through their providers' APIs, reached through the OpenRouter gateway: the OpenAI and Anthropic models with their own web search, Perplexity's Sonar model (which always searches), Gemini without search grounding, and five comparison models. Cited sources are recorded only from the engines that searched; an engine answering from model knowledge cites nothing, and nothing is invented for it.

Each question is asked three times of each engine by default, and results are reported as "named in k of n runs", never as a score. A named reviewer reads and signs off every record before it can appear in a client report. The full method, including what it cannot tell you, is on the methodology page. To see it on one of your own clients first, request a free Snapshot.