AI Citations vs Mentions: What Each One Measures and When to Track It

A mention is your brand's name in the answer text. If an engine says "for small clinics, consider Brand A or Brand B," both brands are mentioned. A citation is a source the engine attaches to the answer: a link, a footnote number or a source card. The cited page can be yours, or a review site, forum thread or article that talks about you.

The two overlap less than most dashboards imply. Four combinations happen:

• Mentioned and cited. The engine names you and links your page.
• Mentioned, not cited. The model names you from memory, or names you while citing someone else's page.
• Cited, not mentioned. Your page supports a general claim, but your brand never appears in the text.
• Neither. You are absent.

Each combination tells you something different, so count them separately.

Answers come from two places. With retrieval off, the model answers from parametric knowledge, which is frozen at a training cutoff. OpenAI documents a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path (chat-latest). These answers can mention you but cannot cite anything, because no pages were fetched. Nothing you publish today reaches them.

With retrieval on, a citation becomes possible, but only after three gates:

1. Access. A crawler must be allowed and able to fetch the page. OpenAI says sites opted out of OAI-SearchBot "will not be shown in ChatGPT search answers, though can still appear as navigational links."
2. Retrieval. The page must enter the model's context for that query. For Google AI Overviews and AI Mode, Google says a page must be indexed and eligible to show with a snippet.
3. Generation. The model decides which retrieved pages to cite.

Retrieval is per query and discretionary, as of October 2026. ChatGPT decides when to search, Claude.ai search is a user toggle, and AI Overviews often do not trigger. Published estimates disagree sharply: Profound reports about 18% of ChatGPT conversations trigger a web search, while cloro reports 80.5% globally on a 100-prompt set. Treat retrieval state as unknown unless you see citations.

What it means: a mention with no citation of your site is often a parametric answer, or a retrieval answer built from other people's pages. That second case is why third-party mentions matter.

The access row matters most. Content that appears only after JavaScript runs may be invisible to non-Google AI crawlers, and a firewall can block a crawler that robots.txt allows. Both stop citations of your site before any content question arises.

It is tempting to treat citations as the hard metric. The evidence says they are noisy, and the noise is not random.

• Engines read more than they cite. OpenAI's API documentation says the number of sources consulted is often greater than the number of citations. Pages that shaped the answer can go uncredited.
• Citation output is mechanically separable from sources. A 2026 interpretability study on Llama-3.1-8B found that amplifying specific attention heads repaired over 90% of missed citations and eliminated 69% of spurious ones without changing the sources or harming answer accuracy.
• Units differ. Anthropic's web search tool cites text spans of up to 150 characters, while Perplexity cites whole sources by number. Summing them adds unlike things.
• Brand size dominates. A 2025 study of AI Search Arena data found the top 20 news sources account for 67.3% of all citations for OpenAI models. Compare yourself with like-sized brands.
• Prompt samples miss volume. In Otterly's 2026 experiment, Microsoft's AI Performance report counted 42,817 citations for one test site, against 3,934 logged by Otterly's prompt tracker. The windows may not match, but a prompt sample is a sample.

What it means: a citation shows the engine fetched and credited a page. It does not measure how much that page shaped the answer. For what the studies say about trying to earn citations, see LLM citation optimization.

Mentions

Ask the same buyer questions on each engine, logged out with memory off, so results are repeatable. Record whether you are named, which facts are stated, and whether those facts are right. Note whether citations appeared, since that is your only signal of retrieval state. A reproducible ChatGPT tracking method covers the setup.

Citations

Record every cited URL and its domain. Split your own domain from third-party domains. For traffic, OpenAI says ChatGPT adds utm_source=chatgpt.com to referral links, so analytics can see some of it. Google AI Overviews send no referral tag, and Search Console's generative AI report shows impressions only.

Keep every engine separate. Otterly's test found Microsoft Copilot produced 2,718 of one site's 3,934 tracked citations while ChatGPT produced almost none. A blended score would have hidden that. Be wary of a single share of voice number for the same reason.

For a quick baseline, the free checker on this site shows whether AI engines name you for up to three buyer questions. It is a snapshot, not continuous tracking.

Pick the metric that matches the decision you need to make.

• Awareness and consideration: track mentions. Being named when buyers ask for options is the outcome. A link is a bonus.
• Reputation and accuracy: track what mentions say. A wrong price hurts whether or not anything is cited. Fix the fact on a page engines can fetch, and see whether tools can prevent hallucinations.
• Referral traffic: track citations of your own domain, paired with analytics. Mentions without links rarely show up as visits.
• Technical health: use citations of your site as an access signal. If no engine ever cites your pages, check access first: robots.txt rules for OAI-SearchBot, PerplexityBot and Claude-SearchBot, firewall 403s, and whether key content is in the HTML before JavaScript runs.
• PR and partnerships: track which third-party domains get cited in answers about your category. That list shows where being named is the route into the answer. Compare your own site vs third-party mentions.

Neither metric is a promise. You control access and the accuracy of your own facts. You do not control what the models choose to say or cite.

Can ChatGPT mention my brand without citing my website? Yes. When ChatGPT answers without searching, it draws on training data with a fixed cutoff and attaches no sources, so any mention arrives with no link. When it does search, it may name you while citing a review site or article instead of your own pages. Both are normal. Check which case you are seeing by looking for citations in the answer.

Is a citation worth more than a mention? It depends on the goal. A citation of your own domain can send traffic and confirms an engine fetched your page. A mention in a recommendation shapes buyer consideration even with no link. For a brand most buyers have not heard of, being named accurately often matters more than being linked. Track both, separately, and judge each against the goal it serves.

Why do my citations differ so much between engines? Engines use different indexes, crawlers and citation formats. In Otterly's 2026 experiment, Microsoft Copilot, which grounds on Bing, cited one test site thousands of times after Google deindexed it, while ChatGPT almost never cited it. Claude cites text spans and Perplexity cites whole sources, so even the unit differs. Compare each engine with itself over time, not with the others.

Does adding schema markup get my pages cited more? The evidence does not support that. Google says structured data is not required for generative AI search. Ahrefs' matched-control study of 1,885 pages found AI Overview citations fell 4.6% after schema was added. Existing markup does no harm, but it is not a citation lever. Crawler access, server-side rendering and clear facts on fetchable pages have stronger support.