What to count
LLM visibility is how often AI answers bring up your site for the questions your buyers ask. Tools wrap it in scores and shares of voice, but underneath there are three things you can check by hand, for each answer:
| Count | What it means | How to check |
|---|---|---|
| Mentioned | Your brand or domain appears in the answer text | Search the answer for your names |
| Cited | A URL on your site is among the answer’s sources | Read the citations the API returns |
| Found | Your page was in the search results the model read, cited or not | Read the search results the API returns |
Record each one as “x of 3” for a question asked three times, and keep the three apart. A mention without a citation gives the reader no link to you; a page that is found but never cited tells you something different from one that is never found.
Our test: 300 answers from Claude
On 19 October 2026 we asked Claude 100 questions written for three websites we run, three times each:
- Model and tool: Claude Sonnet 5.5 through the Claude API, with the web search tool (
web_search_20250305, at most 5 searches per question), no system prompt and default settings. - Questions: written the way a buyer would type them. 91 are general questions; 9 are about a site itself, 8 of them by name and 1 describing what the site does without naming it. By kind: 33 how-to, 31 what-is or how-much, 23 asking for recommendations, 4 comparisons and the 9 about a site itself.
- Cost: about $15.26 by our estimate from the token counts, or 5 cents an answer (see What it costs).
For the 91 general questions:
| Site | Questions | Cited at least once | Cited every run |
|---|---|---|---|
| Site A | 17 | 9 | 5 |
| Site B | 28 | 0 | 0 |
| Site C | 46 | 0 | 0 |
Of the 9 questions about the sites themselves, Site A was cited in every run for all 3 of its questions, Site B for none of its 2, and Site C for 2 of its 4. Four lessons came out of the data.
1. Ask every question more than once
Site A was cited for 8 of its 17 questions in the first run, 5 in the second and 8 in the third; across the three runs it was cited at least once for 9. Only 5 questions cited it every time. One run would have reported anything from 5 to 8.
The other sites in the answers moved even more. Of the 73 questions where Claude searched in all three runs, only 6 cited exactly the same set of websites each time. On average, of all the sites cited in any of a question’s three runs, 39% were cited in all three.
So treat a single answer as a sample. Three runs per question is our minimum; record “cited in 2 of 3”, not “cited”.
2. Count searched and unsearched answers apart
76 of the 300 answers (25%) did not search at all. Claude answered from what the model already knew, and an answer without a search has no sources to cite. Anthropic’s documentation says Claude decides when to search based on the prompt: it searches for current or changing information, and answers directly from stable knowledge such as established facts and concepts.
In our data the decision depended on the question, not on chance. 73 questions searched in all three runs, 24 never searched, and only 3 were mixed. The kind of question mattered most:
| Question kind | Questions | Answers without a search |
|---|---|---|
| Asking for recommendations | 23 | 0 of 69 |
| About the site itself | 9 | 0 of 27 |
| Comparisons | 4 | 3 of 12 |
| What is it, how much | 31 | 27 of 93 |
| How to | 33 | 46 of 99 |
For a question the model answers from memory, no page you publish this week can be cited, however good it is. Report those questions separately, or leave them out of your tracking set, so they do not dilute the citation rate of the questions where your pages can compete.
3. Count questions about you separately
When a question names your site, the answer repeats the name: in the 8 questions that named one of our sites, all 24 answers did. So “mentioned” tells you nothing there. Look at whether the answer cited you.
Even asked by name, Site B was never cited. Site C was cited in every run for 2 of its 3 named questions and never for the third. For Site B, Claude searched every time, but none of the search results were the site’s own pages.
The ninth question described what Site C does without naming it. None of its three answers mentioned or cited Site C. A question like that is the better test of whether an AI assistant connects the topic with you, because the answer cannot simply repeat your name.
4. Found was enough to be cited
In every one of the 36 answers where one of our sites appeared in the search results Claude read, the answer cited it. In our data the hard part was getting into the search results, not being picked from them.
That is one model, three sites and one day, so treat it as a hint. But it points to the first thing to check when you are never cited: whether AI search can reach and index your pages at all. Anthropic’s search crawler is Claude-SearchBot; see the AI crawler list for it and the others, and how to verify Googlebot and AI crawlers reach your site for whether they really get through.
How to run your own test
- Write the questions. 20 to 100 questions the way your buyers would type them, without your name. Tag each by kind, add a few that name you and a few that describe what you do without your name, and keep those apart. Avoid questions tied to one day’s news, so that one week can be compared with the next.
- Use the official API with web search on. Do not script the chat apps. Anthropic’s consumer terms forbid using its services “through automated or non-human means” except through an API key, and OpenAI’s Terms of Use for ChatGPT forbid extracting output automatically or programmatically.
- Fix the settings and write them down. Model, search tool and version, search limit, system prompt (or none) and date. Change one and you have a new baseline.
- Ask each question at least three times.
- Save everything for each answer: the full text, the cited URLs, the search queries and the search result URLs. With Claude’s web search tool, citations are always on, and the response lists every search result with its URL.
- Count per question: mentioned, cited and found, each as “x of 3”, with searched and unsearched answers marked.
- Repeat on a schedule. Run the same set every week and compare question by question. A change that shows up in one week only is not yet a result.
- Compare with the official reports. Bing’s AI Performance report counts citations in Microsoft Copilot and Bing’s AI summaries, and Search Console counts impressions in Google’s AI Overviews and AI Mode. See how to read Bing AI Performance and Google’s AI reports.
What it costs
Claude’s web search costs $10 per 1,000 searches plus tokens, and the search results Claude reads are billed as input tokens. Our 300 answers made 294 searches ($2.94) and used 4.0 million input tokens ($8.01 at Sonnet 5.5’s $2 per million) and 0.43 million output tokens ($4.31 at $10 per million). So most of the money went on reading search results, not on the searches themselves.
What this does not tell you
- Other AI assistants. Our numbers are for Claude through the API. ChatGPT, Gemini, Perplexity and Copilot each need their own runs; do not assume results from one apply to another.
- What people see in the apps. claude.ai and Claude’s mobile apps send Claude a system prompt that the API does not, so the same question can get a different answer there.
- Mentions with certainty. Matching names in text misses spelling variants and can match lookalikes; check a sample of answers by hand.
- Traffic. A citation is not a click. For visits, look at the referrers in your analytics and the clicks in Search Console.
Citeloop’s free GEO audit checks the access side from outside: whether your robots.txt lets Claude-SearchBot, OAI-SearchBot, PerplexityBot, Googlebot and Bingbot in, and whether your pages carry noindex or nosnippet. It does not ask AI assistants your questions; that is the test above.