How to measure AI visibility

The server log is the only place that shows whether anyone reads the site and its files. Every request leaves a line with the address and the client name.

grep -E 'GPTBot|OAI-SearchBot|ChatGPT-User|ClaudeBot|Claude-SearchBot|Claude-User|PerplexityBot|Perplexity-User' access.log \
  | awk '{print $7}' | sort | uniq -c | sort -rn | head -n 20

The command shows which addresses the bots of AI companies requested. Look at /llms.txt separately:

grep ' /llms.txt' access.log | awk -F'"' '{print $6}' | sort | uniq -c | sort -rn

Keep in mind:

• Bots come in three classes. A training bot, a search bot and an agent acting for a user are different events. A training bot’s visit says nothing about appearing in answers. Classes and names: AI crawlers.
• A client name can be faked. Large owners publish the address ranges of their bots: verify against them.
• Requested does not mean used. A bot downloaded a file; that is a fact. Whether the text reached an answer, the log does not know.

of llms.txt files were requested by nobody in a month

97%

Ahrefs, 137K domains, May 2026

of AI agent visits on one site went to llms.txt

0.1%

OtterlyAI, 90 days of logs

When a person clicks a link in an AI answer, the site receives an ordinary visit with a referrer. Analytics shows it in the sources report.

Sources worth putting into a separate group: chatgpt.com, perplexity.ai, claude.ai, gemini.google.com, copilot.microsoft.com.

Limits:

• Some visits arrive without a referrer. Apps and some browsers do not pass it, and such a visit lands in “direct”.
• Google AI Overviews are not a separate source. Their clicks are counted in overall search statistics; Search Console has no separate breakdown.
• Referrals are few by nature. A person got the answer and often goes nowhere. No referrals does not mean no mentions.

A citation is a link to the site in an AI answer. There is one way to count it: ask the question and read the answer.

• Write down the questions the site should be a source for. Not the brand name, but real customer questions.
• Ask them in several systems with search enabled.
• Record: whether the site is named, whether there is a link, whether the meaning is conveyed correctly.

There is a lot of noise. The answer depends on the wording, the region, the conversation history and chance. One run means nothing: repeats are needed.

Share of voice is the share of answers that name the site or the brand over a fixed set of questions. It is a comparative metric: it is read over time and against competitors.

For the number to be trusted:

• The question set is fixed and does not change between measurements.
• Each question is asked several times, and a share is counted, not a single result.
• The system, mode and date are recorded. Models are updated, and a jump may have nothing to do with the site.
• The comparison is with the previous measurement on the same set, not with someone else’s figures.

Paid services do the same automatically. Each has its own method, and numbers from different services are not comparable.

That is enough to tell a change from noise and not to pay for a number that cannot be verified.

• GEO and AEO - where the terms came from and what stands behind them
• Does llms.txt work - studies with their methods and caveats
• Glossary - short definitions
• An AI-ready site - what to prepare before measuring