Learn how to optimize your website to be cited by ChatGPT, Google Gemini, Perplexity, and other AI search engines. Discover practical strategies to improve your AI visibility with SEO, AEO, and GEO.
Quick answer: Each assistant finds sources differently. ChatGPT search relies on OpenAI's OAI-SearchBot, Google AI Overviews and Gemini use Google's index, Perplexity uses PerplexityBot, Claude uses Claude-SearchBot, and Copilot uses Bing. Allow those crawlers, serve answers in plain HTML, and earn mentions on the sites assistants cite most, such as Reddit, Wikipedia and YouTube.
Most advice on getting cited by AI assistants treats ChatGPT, Gemini, Perplexity, Claude and Copilot as one machine. They are not. Each one retrieves pages through a different crawler or search index, follows different robots.txt tokens and leans on a different mix of source domains. If one of those pipes is blocked or misread, no amount of good writing will get you cited on that platform. This guide goes platform by platform, using the vendors' own documentation and the largest citation studies published in 2025 and 2026.
No. Each assistant reaches the web through its own retrieval route, so a page can be visible in one and invisible in another. The practical consequence is that "AI visibility" is really five separate access checks before it is a content question.
The table below summarises what each vendor documents about how it finds pages. It is the first thing we check on any site that says it never appears in AI answers.
| Platform | How it finds pages | Robots.txt token that controls citation | Separate training token | Where you can see citations |
|---|---|---|---|---|
| ChatGPT search | OpenAI's OAI-SearchBot surfaces websites in ChatGPT search; ChatGPT-User handles user-triggered fetches | OAI-SearchBot | GPTBot | No publisher dashboard; referral visits in your analytics |
| Google AI Overviews and AI Mode | Googlebot and the normal Google index; page must be indexed and snippet-eligible | Googlebot | Google-Extended (does not affect Search) | Search Console generative AI performance report |
| Gemini app | Grounding on the Google Search index at prompt time | Google-Extended governs grounding in Gemini Apps | Google-Extended | None specific |
| Perplexity | PerplexityBot surfaces and links sites; Perplexity-User fetches pages for live questions | PerplexityBot | Not used for foundation models, per Perplexity | Referral visits only |
| Claude | Claude-SearchBot and Claude-User; web search reportedly powered by Brave | Claude-SearchBot, Claude-User | ClaudeBot | Referral visits only |
| Microsoft Copilot | Bing's index and AI systems | Bingbot | n/a | AI Performance in Bing Webmaster Tools |
ChatGPT search cites pages that OpenAI's OAI-SearchBot is allowed to reach. OpenAI states that sites which block OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links, and that robots.txt changes take about 24 hours to take effect.
The detail many sites get wrong is that OpenAI runs three separate agents. OAI-SearchBot is for search. ChatGPT-User makes fetches when a user or a custom GPT asks for a page. GPTBot collects data to make OpenAI's foundation models more useful and safe. You can block GPTBot to opt out of training and still allow OAI-SearchBot, which keeps you eligible for citations. A blanket "block all AI bots" rule from a security plugin or CDN removes you from ChatGPT search entirely.
ChatGPT's citations are concentrated. In Ahrefs' monthly tracker for US queries, reddit.com held 16.8% of citation share and Wikipedia 7.0% in September 2026, followed by Consumer Reports and Forbes. The pattern is consistent: ChatGPT favours encyclopaedic references, big publishers and large community sites.
Google's AI features draw on the same index as classic Search. Google's documentation says a page must be indexed and eligible to be shown in Google Search with a snippet, and that there are no extra technical requirements, special schema or "AI text files" needed.
What differs is how many searches happen behind one question. Both AI Overviews and AI Mode may use a "query fan-out" technique, issuing related searches across subtopics and showing a wider set of links than a classic results page. That is one reason ranking and citation have drifted apart. Ahrefs found in March 2026 that only 38% of pages cited in AI Overviews also rank in the top 10 for the same query, down from about 76% in its July 2025 study, and that YouTube was the most-cited domain.
The Gemini app works differently again. Google describes Google-Extended as the control for training Gemini models and for grounding (providing content from the Google Search index to the model at prompt time) in Gemini Apps. Google also states that Google-Extended does not impact a site's inclusion in Google Search. So blocking Google-Extended can remove you from Gemini app answers while leaving AI Overviews untouched; blocking Googlebot removes you from both.
Perplexity cites pages that its own crawler, PerplexityBot, can index. Perplexity says PerplexityBot is designed to surface and link websites in search results and is not used to crawl content for AI foundation models. Its second agent, Perplexity-User, fetches pages in response to a live question and generally ignores robots.txt because the request comes from a user.
That second agent is contested. In August 2025 Cloudflare reported that Perplexity used stealth, undeclared crawlers with a generic browser user agent when its declared crawler was blocked. Treat robots.txt as a statement of preference to Perplexity, not a guaranteed lock.
Ahrefs' June 2025 sample found YouTube at 16.1% and Wikipedia at 12.5% of Perplexity's top-domain mention share. Perplexity was also the most accurate of eight tools in the Columbia Journalism Review's Tow Center test, yet still answered 37 percent of queries incorrectly when asked to identify the source of a quoted article; across all tools the failure rate was over 60 percent.
Claude uses its own search agents, and Copilot uses Bing. Anthropic lists three agents: ClaudeBot for training, Claude-User for user-requested fetches and Claude-SearchBot to improve search result quality for users. Anthropic warns that disabling Claude-SearchBot may reduce your site's visibility and accuracy in user search results. For the underlying index, TechCrunch reported in March 2025 that Anthropic had added Brave Search to its subprocessor list and that at least one search in Claude and Brave returned identical citations. Anthropic has not published a full description, so treat Brave as a likely but unconfirmed input.
Copilot is the only assistant with a proper publisher report. Since 10 February 2026, Bing Webmaster Tools has shown total citations, cited pages and the "grounding queries" behind citations across Microsoft Copilot, Bing's AI summaries and selected partners. Microsoft's advice in the same announcement is to use IndexNow so AI systems reference the current version of a page, and to structure content with clear headings, tables and FAQ sections.
Across every study, a small group of reference, community and video sites takes a large share, and brands that are talked about elsewhere on the web get mentioned more. The table summarises the main 2025 and 2026 findings.
| Study | Platform | Finding |
|---|---|---|
| Ahrefs, Sept 2026 | ChatGPT (US) | Reddit 16.8%, Wikipedia 7.0% of citation share |
| Ahrefs, June 2025 | Perplexity | YouTube 16.1%, Wikipedia 12.5% of top-domain mention share |
| Pew Research Center, March 2025 data | Google AI summaries | Wikipedia, YouTube and Reddit were 15% of cited sources; government sites 6% versus 2% in standard results |
| Ahrefs, May 2025, 75,000 brands | Google AI Overviews | Brand web mentions correlate 0.664 with AI mentions; backlinks 0.218; 26% of brands had zero mentions |
The Ahrefs brand study is the most useful for a business. Off-site signals, branded web mentions (0.664), branded anchors (0.527) and branded search volume (0.392), correlated far more strongly with AI Overview mentions than link counts did. Ahrefs is careful to say correlation is not causation, but the pattern matches what the domain tables show: assistants repeat what the wider web already says about you.
Fix access first, then make your pages easy to extract, then build the off-site evidence that assistants repeat. In order:
Bot protection can hide you from assistants without anyone noticing. A client's host served our own scanner a bot-challenge page ("One moment, please...", returned as HTTP 200) instead of the real site. If an AI crawler meets the same wall, it has nothing to read or cite, so test your key pages with crawler user agents, not just a browser.
The most common reason a site is missing from AI answers is not authority. It is that the crawler never saw the content.
Our own FAQ page is the clearest example. Its 12 questions were built with JavaScript after the page loaded, so the raw HTML that a crawler receives contained about 55 words. To Googlebot, which renders JavaScript, the page looked fine. To crawlers that read only the HTML, it was close to empty. We moved the questions into the HTML, taking the raw page to 595 words, and added FAQ schema. Many AI crawlers do not run JavaScript, which is exactly what the Vercel data above shows, so this one change mattered more for assistants than for Google.
The second lesson concerns reliability. During testing, our own repeated verification requests against a client's live e-commerce site, on top of real traffic, pushed it over its own API rate limiter, and the errors looked exactly like a broken deploy. AI crawlers add their own bursts of automated traffic. If your server or firewall answers them with errors or rate-limit responses, those pages cannot be read or cited, so check your logs for 429 and 5xx responses to known crawler user agents.
When we run these checks for clients through our SEO and AI search work, we start with the raw HTML and the robots and CDN rules before touching a word of copy.
Stop chasing a single "AI ranking factor". The evidence says citation is a pipeline: access, extraction, then reputation. The access layer is binary and cheap to fix, yet it is where we most often find the problem, usually a training-bot block that also killed search bots, or content that only exists after JavaScript runs. Fix that in a day. Then spend your real effort on the reputation layer, because every large study shows assistants citing Reddit, YouTube, Wikipedia and well-known publishers, and the Ahrefs brand data shows mentions beating links. For a small business, that means being discussed accurately in the places your buyers already read, not publishing more pages. If you want the definitions behind these terms, see our guide to AEO vs SEO vs GEO; for a full site checklist, use our AI search optimisation guide.
Not from ChatGPT search. OpenAI documents GPTBot as its training crawler and OAI-SearchBot as the crawler that surfaces websites in ChatGPT search. You can disallow GPTBot and allow OAI-SearchBot, which opts you out of model training while keeping you eligible to be cited. Blocking OAI-SearchBot is what removes you from search answers, apart from navigational links.
No major vendor documents using it for citation. Google states you do not need new machine-readable files, AI text files or special schema to appear in AI Overviews or AI Mode. An llms.txt file is cheap and harmless, but it will not fix a blocked crawler, JavaScript-only content or a lack of third-party mentions, which are the issues that actually stop most sites being cited.
It depends on how quickly each platform recrawls. OpenAI says robots.txt changes take about 24 hours to apply, while Google notes recrawling can take from several days to several months. Using IndexNow for Bing and requesting indexing in Search Console speeds up discovery. Citation itself also depends on whether the page answers the question better than sources the assistant already trusts.
Partly. Bing Webmaster Tools shows citation counts and cited pages for Copilot and Bing's AI summaries, and Google Search Console's generative AI report shows impressions in AI Overviews and AI Mode, without clicks. ChatGPT, Perplexity and Claude offer no publisher dashboard, so track their referral visits in analytics and run a fixed set of test prompts each month.
No. In the Tow Center's 2025 test, eight AI search tools gave incorrect answers to more than 60 percent of queries asking them to identify a quoted news article's source; even the best performer, Perplexity, was wrong 37 percent of the time. Check how assistants describe your business, and correct errors at the source pages they draw from.