How to Get Cited by ChatGPT, Based on the Evidence

Here's how to get cited by ChatGPT, according to what the evidence supports. One lever is documented: letting OpenAI's search crawler in. Two more have indirect evidence, which are being mentioned on sites other than your own and putting specific, sourced answers high on your pages.

The rest is weaker, and no source promises a citation.

So which tips are real? Which ones come from a single vendor's chart? And how would you know if one worked?

Here's each lever with the study behind it, what that study covers and a way to test your own result.

Key takeaways

• OpenAI documents one lever: letting OAI-SearchBot crawl your site.
• The strongest observed pattern is mentions on other sites, not more content on yours.
• Most on-page tactics were tested on other engines, so they're likely helpful, not proven for ChatGPT.
• ChatGPT's answers change from ask to ask, so judge a change with counts over many prompts and several runs.

Which ChatGPT citation levers have evidence?

A ChatGPT citation is a link to a source inside an answer that used web search. OpenAI says ChatGPT may search automatically when a question would benefit from current information, so some answers carry no citations at all.

OpenAI's help article also says it ranks search results using multiple factors and that placement isn't guaranteed. It doesn't list those factors. Every tip below is therefore documented, tested, observed or inferred, and the table says which.

Does ChatGPT's crawler need access?

Yes, ChatGPT's crawler needs access to your site. OpenAI says sites opted out of OAI-SearchBot won't be shown in ChatGPT search answers, though they can still appear as navigational links.

OpenAI's crawler documentation recommends allowing OAI-SearchBot in robots.txt and allowing requests from its published IP ranges. That second part matters if a CDN or firewall filters bots. A robots.txt change can take about 24 hours to take effect for search.

GPTBot is a separate setting, used for training OpenAI's models. Each setting is independent, so you can allow search and still refuse training.

A third agent, ChatGPT-User, visits pages when a person's request triggers it. OpenAI says robots.txt rules may not apply to those visits and that ChatGPT-User isn't used to decide whether content appears in search.

Check three things:

1. Open your robots.txt and look for rules that name OAI-SearchBot.
2. Check your CDN or firewall for bot blocks that catch OpenAI's traffic.
3. Allow the IP ranges OpenAI publishes.

A page OAI-SearchBot can't crawl is unlikely to appear in ChatGPT search answers, whatever the copy says. Zyppy's review scores URL accessibility highest of 23 factors, though those scores are one author's judgment. OpenAI frames crawling as eligibility and promises no citation.

Do other sites matter more than yours?

Other sites appear to matter more than your own: one preprint finds ChatGPT's cited sources are mostly third-party, and Ahrefs finds brand mentions elsewhere track AI visibility. Neither proves cause.

A September 2025 preprint by Chen and colleagues compared AI search with Google. It found a systematic and overwhelming bias towards earned media, meaning third-party authoritative sources, over brand-owned and social content. The authors collected the data in August 2025 by querying up to four AI engines through their APIs.

Most questions were about products and rankings, so the result may not carry over to other topics. It also isn't the ChatGPT web app.

Ahrefs analyzed 75,000 brands and measured something different: brand mentions in answers, not cited sources. Branded web mentions correlated 0.66 to 0.71 with AI visibility across ChatGPT, AI Mode and AI Overviews, and YouTube mentions correlated higher, at about 0.74. The brands all had a Domain Rating above 40, and Ahrefs notes that correlation isn't causation.

The two studies don't agree on video. In one Chen experiment, ChatGPT's cited sources included no social platforms, a class that includes YouTube. API versus web app may explain the gap, and the studies don't settle it.

So the practical question is where ChatGPT looks for your category. Firemark's Sources screen lists the sites ChatGPT cited for your prompts, split between answers that name you and answers that name a competitor and not you. Review and comparison sites that it already cites are the safest places to be named, given the earned-media finding.

How to get cited by ChatGPT from your own pages

Write specific, sourced facts in self-contained sentences near the top of each page. The strongest controlled test of those tactics wasn't run on ChatGPT, so treat them as likely, not proven.

A preprint from Aggarwal and colleagues, GEO: Generative Engine Optimization, first posted in 2023 and published at KDD 2024, tested rewrites on a benchmark it built and on Perplexity.ai. On its own benchmark engine, adding citations, quotations and statistics gave relative gains of 30 to 40% on a metric that measures how much of an answer draws on your text and where. On Perplexity.ai, tested on a 200-query subset with source text supplied as uploaded files, the top gain on that metric was 22%. Keyword stuffing gave little to no improvement.

Zyppy's review covers ChatGPT, Gemini and Perplexity. It scores an answer near the top of the page, self-contained passages and factual specificity highly, and it stresses these are correlations, not confirmed ranking factors. The main named evidence for the top-of-page effect is a finding about Gemini.

Put that together and the page-level work is plain:

1. Answer the question in the first sentence under each heading.
2. Put each figure and its source in the same sentence.
3. Name the subject in every sentence, so the sentence still makes sense when lifted out.
4. Keep key text in visible HTML, not behind tabs or scripts.

Nothing here is exotic. It's clear writing, and the correlational evidence points the same way. Whether it moves ChatGPT specifically is untested.

What does the evidence not support?

Three popular levers have little support: llms.txt files, link volume and page count.

Google's guidance on generative AI features says Google Search itself doesn't use llms.txt files, and that site owners can ignore creating unnecessary AI text files like llms.txt. That covers Google's AI features, so it says nothing directly about ChatGPT.

For ChatGPT and other engines, the clearest evidence is about use. Ahrefs studied about 137,000 domains using its web analytics and found that 97% of the llms.txt files it identified received zero traffic in May 2026. It concludes that for showing up in ChatGPT, Perplexity or AI Overviews, the file is largely decoration.

A fetch log shows only that a file was requested, not that it changed an answer.

Zyppy's review also says it couldn't find any credible evidence or experiment showing llms.txt files influence AI citations, and it notes most reviewed studies may not have tested them.

The exception is coding agents reading developer documentation. Ahrefs found coding agents among the top requesters of the files, ahead of every AI search and assistant bot. Tenpoint Labs' guide to llms.txt covers who documents reading the file and when it is worth publishing.

Ahrefs found very weak correlations for backlinks and URL rating, and a weak one for Domain Rating (0.266 for ChatGPT). It also found almost no relationship between the number of pages on a site and AI visibility. A weak correlation isn't proof of no effect, but publishing at volume has no case behind it either.

Can you rank in ChatGPT?

You can't rank in ChatGPT the way you rank in Google, because ChatGPT answers have no fixed positions. A better question than "how to rank in ChatGPT" is how often it names you.

SparkToro and Gumshoe had 600 volunteers run 12 prompts through ChatGPT, Claude and Google's AI, 2,961 runs in November and December 2025. Their study found less than a 1 in 100 chance that ChatGPT gives the same list of brands in any two responses, and less than 1 in 1,000 for the same list in the same order. A rank report for ChatGPT is a horoscope with a spreadsheet.

The authors say newer models may behave differently. They call percent visibility across many runs probably a reasonable measure, suggest asking each prompt at least 60 to 100 times, and don't claim it's the best metric. The study covers lists of brands and products, not citations. If you track by tool, Firemark's MCP server lets Claude or ChatGPT read your counts, so you can ask which prompts name you.

How do you know a change worked?

Count the answers that name or link you across a fixed set of prompts, before and after the change, over several runs. One answer proves nothing, and one run proves little.

Ask the same prompts from the same place each time. OpenAI says ChatGPT uses an approximate location from your IP address, so a run from another city can return different sources.

Firemark's method page sets the rules it uses, and you can copy them into a sheet:

1. Fix at least 20 prompts, because Firemark shows a 95% range only from 20.
2. Count named and linked separately. A link alone isn't a mention.
3. Record three comparable runs as your baseline.
4. Change one thing.
5. Run again and call a movement only after three comparable runs and a change of 5 points or more.

The size of the set decides what you can see. This arithmetic assumes a 40% share on one engine and is not data from a real brand.

A 5-point move inside a 36-point range is noise. Firemark's 5-point rule is a floor, and a small set can't resolve it. SparkToro's 60 to 100 asks per prompt is a higher bar than 20 prompts and three runs, so treat those as a minimum.

Counts show what answers say, not what people do next. Pair them with referral sessions in your analytics, because Firemark doesn't measure clicks, visits or sales, and an analytics suite fits better for that.

Firemark measures and doesn't change what ChatGPT says. Start with 20 buyer questions and a sheet, or open the demo report to see counts on sample data. Test the levers one at a time.