Which sources do AI assistants cite? A guide to source analysis

A cited source is a web page that an AI assistant links to as a basis for its answer. ChatGPT, Gemini, Claude, Perplexity and Grok cite sources when they search the web before answering, and those citations show which websites shaped what the assistant said about your category, your competitors and you. Source analysis means collecting citations across many answers and asking which sites appear, how often, and whether yours is among them.

When do AI assistants cite sources?

Only when they search the web. An answer written from the model's own knowledge involves no retrieval, so there is nothing to cite. If you ask such a model for sources anyway, it may write out web addresses from memory, and those can be outdated, wrong or invented.

Whether a search happens depends on the assistant, its settings and the question. Perplexity searches for every answer. Other assistants treat search as a tool they may or may not use: in our own test, Gemini searched in about 6 of 10 answers even with search enabled. That makes the share of answers that cite anything a metric in itself, and the first number to check in any source report.

Where do citations come from?

Assistants that search return citation data alongside the answer text, through their APIs, each in its own format. OpenAI's API, for example, returns URL citations attached to the answer, and Gemini returns grounding metadata that lists the pages it used. This data is separate from the text the model writes.

That distinction matters. A cited source, taken from the provider's citation data, shows a page the assistant actually retrieved. A web address the model merely wrote in its text is a claim. Source analysis should rely on the first; Geovium counts only the citations the provider returns.

What kinds of sites get cited?

Citations for a typical category question fall into a few groups:

• Your own website: product pages, pricing, documentation, blog posts.
• Competitors' websites: their product pages, comparison pages and guides.
• Review and comparison sites: software directories, product review sites, "best of" lists.
• Media and industry publications: news sites, trade magazines, analyst content.
• Marketplaces and directories: app stores, e-commerce platforms, local listings.
• Communities and forums: discussion platforms and question-and-answer sites.
• Reference and official sources: encyclopedias, standards bodies, public authorities.

Which groups dominate depends on the category, the question and the assistant. That is exactly why you should measure citations for your own questions instead of relying on general lists of "sites AI loves".

Why do citations of your own website matter?

• You control the facts. When your page is the source, the prices, features and positioning in the answer come from you rather than from a third party's summary.
• Your brand becomes part of the evidence, not only the text. A citation shows that the assistant used your page to form its answer.
• Readers get a path to you. Citations are usually shown as links the reader can open.
• It is scored. In Geovium's visibility score, having your website named or linked in the answer is worth 15 of 100 points, and the link rate tracks how often answers that mention you also cite your site.
• It shows which pages work. Topics where competitors or third parties are cited and you are not point to pages you are missing.

How do you analyze cited sources?

1. Start with the share of answers that cite any source, per assistant. If it is low, few answers searched the web; check that before drawing conclusions.
2. Rank sites by the share of answers that cite them, not by raw link counts. One answer can link the same site five times; counting distinct answers keeps a single long answer from distorting the picture.
3. Classify each site as yours, a competitor's or another site.
4. Compare assistants. Different assistants often rely on different sites for the same question.
5. Compare topics and questions. Sources for "best X" questions tend to differ from those for "how do I" questions.
6. Compare periods. Note the sites that are new this period and the sites that dropped out.
7. Open the pages. Look at what earns the citation: format, freshness, specificity, and whether the page answers the question in its first lines.
8. Turn the findings into three lists: your own pages to create or improve, third-party sites to earn a place on, and competitor pages to learn from.

How can you earn more citations?

Nobody can guarantee a citation, but these steps make your pages easier to find, read and quote:

• Let the search crawlers in. Allow the crawlers that assistants use for search, such as OAI-SearchBot (OpenAI), Claude-SearchBot (Anthropic), PerplexityBot and Googlebot, and make sure your text is in the HTML your server sends.
• Give each important question a page. Put the answer in the first paragraph, then the detail.
• Be specific. Prices, specifications, limits, supported integrations, delivery areas and dates are easier to cite than general claims.
• Keep it current. Update facts when they change, and show when a page was last updated.
• Structure it for quoting. Headings phrased as questions, short paragraphs, lists, tables and a brief FAQ.
• Publish original material. Your own documentation, data and methodology give others, and assistants, something to cite that only you have.
• Earn a place on the sites that are already cited. If a review site or an industry list is cited again and again for your questions, a genuine presence there can matter more than yet another page on your own site.

Avoid shortcuts: mass-produced thin pages for every question, hidden text and fake reviews damage the trust that citations depend on. Our article on practical steps to get recommended covers the wider playbook.

What can't source analysis tell you?

• Citations show which pages were used, not how much weight each one carried.
• The list of sources for the same question varies between answers, so read shares across many answers, not single answers.
• A citation is not an endorsement: a page can be cited for a fact while recommending someone else.
• Answers collected through APIs approximate, but do not reproduce, the personalized answers a signed-in user sees in a consumer app.

How does Geovium analyze cited sources?

Geovium's source report lists the websites the assistants cite for your tracked questions. For each site it shows the share of answers with sources that cite it, which assistants cite it, whether it is your site, a competitor's or another site, and how that changed from the previous period, including the sites that are new this period. A separate figure shows how often answers with sources cite your own website.

A few rules keep the numbers honest. Only citations returned by the provider count, not addresses the model wrote in its text. Each site is counted once per answer. Sites are grouped by their main domain, with subdomains still visible, and the redirect links some assistants return are resolved to the real site. If no answer searched the web in a period, the report says so instead of showing empty results. The website points in the score follow the same logic, as described in our methodology. For how web search changes the numbers, see answers with and without web search.

Frequently asked questions

What is a cited source in an AI answer?

A cited source is a web page that an AI assistant links to as a basis for its answer, returned by the provider together with the answer. Assistants cite sources when they search the web before answering.

Why is my brand mentioned but my website not cited?

The assistant may have taken its information about you from other sites, such as reviews or lists, or you may not have a page that answers that question directly. Check which pages are cited in those answers.

Do all AI assistants cite sources?

Only when they search the web. Perplexity searches for every answer, while ChatGPT, Gemini, Claude and Grok search depending on the question and on their settings.

Are the sources a model lists in its text reliable?

Not always. Without a web search they come from the model's memory and can be outdated or invented. Geovium does not count them as citations.

How do I get my website cited more often?

Allow search crawlers, answer each important question on a clear, specific and up-to-date page, and earn a presence on the third-party sites that are already cited for your questions.