How to get cited by ChatGPT and other AI engines

How do AI engines decide which sources to cite?

An engine can only cite what its search step hands the model. The model doesn't browse the open web. It sends one or more queries to an index, reads a shortlist of results, writes the answer and attaches links to the passages it used. That loop is retrieval-augmented generation, and tying the answer to fetched pages is called grounding.

So there are three gates, and you can fail any of them. Your page has to be in that engine's index, which depends on which crawler you let in and whether a CDN rule quietly blocks it. It has to match the query the engine actually sends, which is often a rewritten or split version of what the user typed. And it has to contain a passage the model finds worth quoting.

The engines differ on the first gate more than people expect. That's why a page can be cited by Perplexity every day and never once by ChatGPT.

Where does ChatGPT get its sources?

ChatGPT search pulls sources through OpenAI's own crawler, OAI-SearchBot, combined with third-party search providers and content supplied by publishing partners. OpenAI's launch post for ChatGPT search on 31 October 2024 described it that way2, and Microsoft had announced at Build in May 2023 that Bing would be ChatGPT's default search experience4. OpenAI has never said how much of today's results come from each provider.

OpenAI's crawler documentation is explicit about the controls. Sites that opt out of OAI-SearchBot "will not be shown in ChatGPT search answers," though they can still appear as navigational links; GPTBot is a separate training crawler with an independent setting; and ChatGPT-User handles visits a user triggers and is not used to decide what appears in search1. The same page says robots.txt changes take about 24 hours to reach search results1. The publisher FAQ adds that ChatGPT tags referral links with utm_source=chatgpt.com3.

ChatGPT shows few links. In Slate's August 2026 test of 1,000 B2B software questions it returned 3.10 links per answer, and 372 of the 1,000 answers had no link at all17. Those answers named no source, so whatever they said about a brand most likely came from training data, which is where GPTBot still matters. When it does cite, it favors reference sites: Profound's analysis of 680 million citations from August 2024 to June 2025 found Wikipedia was 7.8% of all ChatGPT citations and 47.9% of its top 10 cited domains15.

If you want ChatGPT to cite you

1. Grep your access logs for OAI-SearchBot and check the IPs against OpenAI's published searchbot.json. No verified hits in 30 days usually means a firewall or bot-management rule is blocking it, even when robots.txt says allow.
2. Verify the site in Bing Webmaster Tools and submit your sitemap. OpenAI uses unnamed third-party providers and Bing was the announced default, so a page missing from Bing is a page you are betting OpenAI found on its own.
3. Because only about three links make the cut, run your ten most valuable prompts and list every domain ChatGPT cites. Those domains, often Wikipedia, review sites and a couple of trade publications, are your outreach list. Getting your product named on them beats rewriting your own page for the fourth time.
4. Leave GPTBot allowed if you sell software. 37% of ChatGPT's answers in the Slate test carried no link, so what the model learned in training is the only thing those answers can say about you.
5. Build a GA4 exploration filtered to session source = chatgpt.com and watch which landing pages get the clicks. Those are the pages ChatGPT already trusts; extend them before you write new ones.

Where does Perplexity get its sources?

Perplexity cites from its own index, built by PerplexityBot. When it opened that index to developers as the Search API in September 2025, it described hundreds of billions of webpages, the same infrastructure behind its answer engine6. Its crawler documentation says PerplexityBot surfaces and links websites and is not used to train foundation models5. A second agent, Perplexity-User, fetches pages when a user's question needs them and "generally ignores robots.txt rules," per the same docs5.

It is the most citation-heavy engine in every recent test: 19.08 links per answer in the Slate study, with a link on all 1,000 questions17. The mix tilts toward community content. Reddit was 6.6% of all Perplexity citations in Profound's dataset and 46.7% of its top 10 cited domains15. Perplexity also ranks passages inside pages, which its launch post called sub-document precision, so a single strong paragraph on a long page can win the citation6.

If you want Perplexity to cite you

1. Allow PerplexityBot in robots.txt and at the firewall. Perplexity's docs tell Cloudflare users to create a rule with the action set to Allow for its published IP ranges, so it bypasses other security rules5.
2. Make each H2 section stand on its own: the question as the heading, the answer in the first sentence, the number and its source in the next two. Passage-level ranking means a section that needs the previous section to make sense is a weak candidate.
3. Search your category on Reddit, find the three or four threads Perplexity already cites for your prompts, and contribute where you have something specific to add. With Reddit at nearly half of Perplexity's top 10 domains, that thread is often the real competition.
4. Open the Sources panel for each tracked prompt and note which of your competitors' pages appear. Perplexity shows enough links per answer that you can usually see exactly which page format it prefers for a query.

How do Gemini, AI Overviews and AI Mode pick sources?

Google's AI features cite pages from the Google Search index, found through what Google calls query fan-out. Its AI features documentation says AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources, and keep finding supporting pages while the response is generated7. The May 2025 AI Mode announcement said Deep Search can issue hundreds of searches8. Our glossary entry on query fan-out covers the mechanics.

The eligibility rule is short. A supporting link "must be indexed and eligible to be shown in Google Search with a snippet," and Google says there are no additional requirements or special optimizations7. Googlebot does the crawling. Google-Extended, the token many publishers block, has no effect on Search; it governs Gemini training and grounding in the Gemini app and Vertex AI9. The Gemini API's grounding documentation says the model decides whether a search would help, runs one or more queries, and returns citation annotations tied to text spans10.

Fan-out makes Google's AI results unstable. Ahrefs compared AI Overviews and AI Mode on the same September 2025 US queries and found only 13.7% of cited URLs overlapped (16.3% for the top three), though the answers averaged 86% semantic similarity16. Wikipedia was in 28.9% of AI Mode citations against 18.1% for AI Overviews16. In the Slate test, AI Overviews showed 9.42 links per answer, while Gemini 2.5 Flash through the API returned links on just 16 of 1,000 answers17.

If you want Google's AI features to cite you

1. Run your money pages through URL Inspection in Search Console. "URL is on Google" plus no nosnippet is the whole technical requirement; anything else is a blocker you can fix today.
2. Search your templates for data-nosnippet and max-snippet. Google won't use a passage it isn't allowed to show as a snippet, and these often get added for one widget and then copied site-wide.
3. Pull the "People also ask" questions and the related searches for your head term and give each one its own H2 or H3 with a one-sentence answer. That's the closest public proxy for the sub-queries fan-out sends.
4. Check that any structured data matches the visible text. Google's AI features page asks for exactly that and says no special schema is required7.
5. Don't wait for a separate AI report. Google counts AI Overviews and AI Mode traffic inside the normal Performance report under the Web search type7, so track the queries where your clicks fall while impressions hold.

Where does Claude get its sources?

Claude cites pages returned by Anthropic's web search tool, which searches when a request depends on current information and answers from training data when it doesn't11. Citations are always on for web search results, and Anthropic's help center says every response that uses web search includes them1122.

Anthropic doesn't name its search provider in product docs. Brave Search was added to its subprocessor list in March 2025, and Simon Willison found that Claude's cited results for a test query matched Brave's for the same search13. That matters for robots.txt, because Brave's crawler help page says it uses no distinct user agent and won't crawl a page Googlebot isn't allowed to crawl26. Anthropic also runs its own crawlers: Claude-SearchBot indexes content for search quality (blocking it may reduce visibility in Claude's results), Claude-User fetches pages for a user's question, and ClaudeBot collects training data12.

Claude cites more than its reputation suggests: 12.49 links per answer from Claude Sonnet 4.6 in the Slate test, with links on 978 of 1,000 answers17. Claude and Perplexity shared 13.6% of cited pages, the highest overlap of any pair in that study17.

If you want Claude to cite you

1. Search your target queries on search.brave.com. If your page isn't in Brave's top results, Claude is unlikely to see it. Brave's help page has a form to request a re-fetch of a page26.
2. Keep Googlebot allowed everywhere you want Claude to reach, since Brave follows your Googlebot rules26. A leftover Disallow for Googlebot on a docs or help subdomain would hide it from both.
3. Allow Claude-SearchBot and Claude-User, and confirm hits against Anthropic's IP list12.
4. Reuse your Perplexity work. With the highest overlap of any engine pair, the same answer-first, well-sourced sections tend to help in both.

Where does Microsoft Copilot get its sources?

Copilot grounds web answers in Bing. Microsoft's documentation for Microsoft Copilot and Copilot Chat says it generates a short search query from the prompt, sends it to Bing, and writes the response from what comes back14. Bingbot access and Bing indexing are the gate.

Bing has page-level controls for this. Content tagged NOARCHIVE is left out of chat answers and not used to train Microsoft's models, while NOCACHE allows only URLs, titles and snippets; both still appear in normal Bing results, per Bing's 2023 announcement21. In February 2026 Bing added an AI Performance report to Webmaster Tools in public preview. It shows total citations, which URLs were cited across Copilot, Bing's AI summaries and select partners, and a sample of the grounding queries behind them24.

If you want Copilot to cite you

1. Open the AI Performance report and export the grounding queries. They are the closest thing any engine publishes to the real search queries an assistant sent for your pages.
2. Set up IndexNow so Bing hears about new and updated URLs when you publish; Bing is one of the participating engines25.
3. Search your templates for NOCACHE and NOARCHIVE robots values. Either one limits what Copilot can show.
4. Verify bingbot hits with a reverse DNS lookup (the host should end in search.msn.com) before you rate-limit anything that claims to be Bing.

How do the engines compare?

The table sums up the retrieval source, the crawler you need to allow, and citation behavior from the studies above. Link counts come from one API-based test, so use them to compare engines and avoid quoting them as fixed figures.

A worked example: who is even eligible for "how to block GPTBot"?

We can't reproduce engine answers here without fabricating them, so this example checks the part you can verify from public files: which top-ranking pages each engine is allowed to retrieve. We ran the query how to block GPTBot on Brave Search (the index Claude appears to use) while researching this guide, took the top 10 organic results, and parsed each site's live robots.txt for six user agents.

• Seven of the ten (Vercel's knowledge base, Mersel AI, Search Logistics, Krasamo, cside, iSocialWeb and Otterly.AI) allowed all six: GPTBot, OAI-SearchBot, Claude-SearchBot, PerplexityBot, Googlebot and bingbot.
• Raptive blocked GPTBot only. Its page stays eligible for every AI search engine while opting out of training, which is the split OpenAI's docs describe.
• Mashable names OAI-SearchBot, Claude-SearchBot and PerplexityBot in disallow groups but lets Googlebot and bingbot in. Its article can support an AI Overview or a Copilot answer, and per OpenAI's docs it won't be shown in ChatGPT search answers. Brave still ranked it ninth, because Brave follows Googlebot rules, so whether Claude cites it depends on how Anthropic treats Brave results for a site that blocks Claude-SearchBot. Anthropic hasn't documented that.
• The Reddit thread in tenth place sits on a domain whose robots.txt, as served to our request, is two lines: User-agent: * and Disallow: /. Reddit made that change in mid-2024, and Microsoft confirmed Bing stopped crawling it27. Reddit content still reaches Google through a data licensing deal28, a reminder that robots.txt isn't the only door into an index.

For anyone writing on this topic, 8 of the 10 competing pages are open to every AI search crawler, so your edge on this query has to come from passage quality and third-party mentions. News queries are different. In our robots.txt study of 261 major websites, 33 of 48 news sites blocked at least one AI search crawler, against 1 of 95 B2B SaaS sites, so on a news query simply being retrievable can put you in ChatGPT's three links.

To test real engine output, run the same prompt set in each engine on the same day, logged out or in a fresh session, and record what you see in a sheet like this:

prompt | engine | date run | answered with search? (y/n) | cited URLs (in order) | our domain cited? | our brand named in text? | competitors named | robots.txt allows engine's bot? (y/n)

Run each prompt three times per engine, because answers vary run to run, and count a citation only if it shows up in at least two of three. Twenty prompts across five engines is 300 runs, about a day of work by hand. Our guide to measuring AI visibility covers sampling and share of voice in more detail, and the free Arobis AI visibility checker runs a buyer-prompt set across engines if you'd rather not do it by hand. We compare it with nine other free options in our roundup of free AI visibility checkers, including which ones show actual citations and which only score readiness.

How little do the engines agree with each other?

Overlap is low in every study that measures it. Slate's same-page overlap ranged from 0.5% for ChatGPT and Claude to 13.6% for Perplexity and Claude, and 74.1% of pages cited by the three link-showing engines weren't in Google's top 10 or AI Overview data for the same questions17. The Digital Bloom's 2025 report puts the share of domains cited by both ChatGPT and Perplexity at 11%, though it doesn't publish its method18.

Watch how source shares get quoted. Profound reports two different metrics: Wikipedia's 47.9% for ChatGPT and Reddit's 46.7% for Perplexity are shares of each platform's top 10 domains, while their shares of all citations were 7.8% and 6.6%15. Plenty of articles mix the two up and end up claiming half of all ChatGPT citations are Wikipedia.

Referral traffic is shifting as well. Goodie's report of 21 May 2026 found ChatGPT's share of AI referral sessions on a mostly B2B panel fell from 89.1% in May to August 2025 to 62.6% in March and April 2026. Claude rose from 1.4% to 18.5%, Gemini from 2.4% to 10.6%, Perplexity from 3.1% to 7.3% and Copilot from 3.2% to 4.0%19. For B2B sites, optimizing for ChatGPT alone now ignores more than a third of AI referrals.

Citations aren't always right, either. The Tow Center at Columbia tested eight AI search tools on 1,600 queries in March 2025 and found they misattributed the source of news excerpts more than 60% of the time; Perplexity was wrong on 37% and Grok 3 on 94%20. Some tools also answered correctly about publishers whose robots.txt should have kept them out20. More figures like these are on our AI search statistics page.

Why one page wins in Perplexity and never shows up in ChatGPT

Usually it's the first gate. A page indexed by Google but blocked for OAI-SearchBot (directly, or by a CDN rule that challenges unknown bots) can sit in AI Overviews for months and never appear in ChatGPT search. A page Bing never indexed can't ground Copilot, however well it ranks on Google.

When access is fine, the difference is usually link budget. Perplexity shows around 19 links, so a decent page on a niche question gets in. ChatGPT shows about three, so it takes the most established sources, and for many commercial queries that means Wikipedia, a big review site and a trade publication. If your page wins in Perplexity and not ChatGPT, look at the pages ChatGPT does cite for that prompt and work on getting named there before you touch your own copy again. The original GEO paper found that adding statistics, quotations and cited sources lifted visibility in generative answers by up to 40%23, which helps once you're in the shortlist. Our generative engine optimization guide covers those on-page tactics, our AI crawler user agent list has the robots.txt rules, and the citability grader scores a page's answer-first structure and sourcing.

Frequently asked questions

Sources