Which Sites Does AI Cite? How to Map the Third-Party Pages Behind Your Answers

A citation map is a ranked list of the domains each AI engine links to when it answers your buyer questions. It shows where answers come from today. It does not show why an engine picked those pages, and nobody outside the engine can see that. Our guide on how AI search picks sources explains the stages between a crawl and a citation.

Three facts shape how you read the map:

• Search does not always run. ChatGPT searches at the model's discretion, and Claude.ai search is a user toggle. An answer with no citations may come from training data alone, so there is nothing to map.
• Engines rewrite your prompt. OpenAI's help centre says ChatGPT search "typically rewrites your query into one or more targeted queries." Your prompt is a proxy for the search that actually ran.
• Citations are different units. Claude's web search tool cites short text spans, while Perplexity cites whole sources by number. See whether engines cite whole pages or specific sentences.

Why bother: when engines cite other sites instead of yours, being named on those sites is the route into the answer. The evidence shows a strong association, not a guaranteed effect. Citations also concentrate. An arXiv study of AI Search Arena data from March to May 2025 found the top 20 news sources accounted for 67.3% of citations for OpenAI models. A short list of domains often carries most of the weight, and the map finds yours.

Answers change with who asks, where, and when. If you change conditions between runs, you can't tell an engine shift from a setup shift.

• Log out, or turn memory off. OpenAI says ChatGPT may use saved memories when rewriting a search query. Personalised answers aren't reproducible.
• Hold location and language steady. Google AI Overviews don't appear on every query, and they vary by location and language.
• Treat each surface as its own engine. Google states that AI Overviews and AI Mode "may use different models and techniques." Log them separately from the Gemini app.
• Record the date and model where visible. As of October 2026, consumer apps change their default models often, and the model behind an answer is not always shown.
• If you use APIs, switch search on. Perplexity's Agent API makes web search opt-in, so a call without it returns an answer with no sources. API answers are also not identical to what the consumer app shows.

1. Write 20 to 30 buyer questions. Use the questions a prospect asks before buying: "best [category] for [use case]", "[your brand] vs [competitor]", "how much does [category] cost". Include a few that name you and a few that don't.
2. Pick your engines. ChatGPT, Perplexity, Google AI Overviews, AI Mode, Gemini, Claude and Copilot are the usual set. Start with the ones your buyers use.
3. Run each prompt on each engine twice. Answers vary between runs. Two runs show which citations are stable and which are noise.
4. Log every cited URL. Copy the full URL, not only the domain. Use one row per citation.
5. Flag answers with no citations. Record them as "no search", not as zero. They tell you the engine answered from memory.
6. Note who is named. For each answer, record whether your brand appears and which competitors appear.
7. Normalise domains. Strip "www", keep meaningful subdomains (a forum subdomain is not the main site), and keep the full URL alongside.
8. Tag each domain by type. Your site, competitor, review platform, directory, forum, publisher, encyclopedia, video.

A log with these columns is enough:

Say you run a dental software company. After one run, your sheet might show the same two review platforms and one industry blog cited on most cost questions. That pattern is what the next step makes visible.

Rank domains by prompt coverage: the number of prompts where a domain is cited at least once. Raw citation counts reward a single long page cited five times in one answer. Coverage rewards the domains engines reach for across your market.

Build one ranking per engine before you build any combined view. A blended list hides the most useful finding. In Otterly's 2026 experiment, Microsoft Copilot produced 2,718 of the 3,934 citations tracked for one test site, while ChatGPT, Gemini and Google's AI surfaces almost never cited it. That split only shows up per engine. Our note on why Perplexity can cite you while ChatGPT doesn't covers the same effect.

Then, for your top ten domains per engine, open the cited pages and answer three questions:

• Are you named on the page?
• Are competitors named?
• Is what it says about you correct?

Remember your log is a sample. In the same Otterly test, Microsoft's own report counted 42,817 citations for the site against 3,934 from the prompt tracker. Where an engine publishes its own data, use it. Bing Webmaster Tools' AI Performance report shows grounding queries, the queries the AI actually ran, for Copilot and Bing answers. As of August 2026 it had no API, so you export it by hand.

A good map

• Each engine has its own ranked list, and you can see where they differ.
• The top domains hold steady across both runs and across months.
• Every top domain carries a note: you named, not named, or named wrongly.
• The output is a short list of pages where competitors appear and you don't. That list feeds getting listed on pages AI engines already cite.

A bad map

• One blended score across engines.
• One run, with no repeat to separate signal from noise.
• Uncited answers counted as zero instead of "no search".
• Domains logged without the URL, so you can't find the page that matters.
• The map read as a ranking you can climb. No published data shows position inside AI answers, and there is no ground truth to check one against.

Two findings need special handling. If a cited page states wrong facts about you, that is a correction task, not a listing task. See what to do when a competitor's comparison page misrepresents you. If a cited page contains hidden instructions aimed at AI systems, flag it. Injected content is a known way to manipulate AI agents, it is illegitimate, and it can misrepresent your brand.

• Same prompt set as last run (changes logged and dated)
• Logged out or memory off on every engine
• Same location and language settings
• Each prompt run twice per engine
• Full URL captured for every citation
• Uncited answers marked "no search"
• Your brand and competitors recorded per answer
• Domains normalised and tagged by type
• Ranking built per engine by prompt coverage
• Top ten domains per engine opened and checked for your name and accuracy
• Wrong facts and suspicious content flagged separately
• Date and visible model recorded

Run the full map monthly. Citation sources move as engines update models and indexes, and a monthly cadence catches that without drowning you in data.

Run it again sooner when:

• An engine changes its default model or launches a new search mode.
• You get listed on, or corrected on, a page from your top list. Check whether the map changes, without expecting it to.
• A competitor launches a comparison page or a review campaign.

Keep the prompt set fixed between runs. Review it once a quarter, and date any change, or your trend line breaks. For a quick snapshot between full runs, the free checker on this site shows whether engines name you for up to three buyer questions. It is a snapshot, not tracking.

The map tells you where to work. The next step is the work itself: our guide to third-party mentions and AI answers covers what being named on those pages involves.