Implementation · Proof

How to measure AI visibility

There's no Search Console for answer engines. There are three methods that triangulate well enough to make decisions with.

The short answer

Measure AI visibility with three independent methods: a repeatable prompt panel scoring mention, citation and recommendation separately; server log analysis of AI crawler activity; and an analytics segment for referral traffic from AI domains.

The panel shows what engines say, the logs show what they fetch, and referrals show what humans do next. Any one alone will mislead you.

01The prompt panel

The core instrument. A fixed list of buyer questions, run on a schedule, scored consistently.

Fix the prompts. 30–60 questions in customers' own words. Once set, don't change them — comparability is the whole point. Add new prompts to a second list.
Fix the conditions. Fresh session, no chat history, logged out where possible, same country setting. Personalisation will otherwise flatter you.
Run each prompt three times per engine and record the proportion of runs.
Score three things separately. This is where most trackers oversimplify.

The three metrics, and what a gap between them tells you
Metric	Definition	If it's low
Mention rate	Brand named anywhere in the answer	Entity association is weak — the model doesn't connect you to the category
Citation rate	Your URL appears as a linked source	Content is not retrievable or not quotable — go back to layers 1 and 2
Recommendation rate	You're named as the suggested option	Entity strength or third-party validation is behind competitors

High citation with low recommendation is the classic pattern: your content is good enough to source but your brand isn't established enough to name. That's an entity problem, addressed in get cited faster.

02Log file analysis

Logs are the only place you see what machines actually did, as opposed to what they said. Filter for AI user agents and track four things weekly:

Hit volume by user agent — is anything crawling you at all?
Status codes — 403s and 429s mean a WAF or rate limiter is blocking you.
Which URLs get fetched — a good proxy for what engines consider important.
Real-time fetchers specifically — ChatGPT-User, Perplexity-User and Claude-User hits mean a human asked a question and an engine came to read your page to answer it. That's the closest thing to a live signal.

# AI crawler hits by user agent, last 7 days
grep -iE "GPTBot|OAI-SearchBot|ChatGPT-User|ClaudeBot|Claude-User|PerplexityBot|Google-Extended" access.log \
  | awk '{print $9}' | sort | uniq -c | sort -rn

03Referral traffic

Create a channel group in your analytics platform matching AI referrers: chatgpt.com, chat.openai.com, perplexity.ai, claude.ai, copilot.microsoft.com, gemini.google.com. Note that AI Overviews traffic generally arrives attributed as ordinary Google organic, so this segment understates the total effect — treat it as a floor.

Watch conversion rate and pages per session, not just volume. Assistant-referred visitors frequently convert better than organic search because the assistant already did the qualifying.

04What not to measure

A single run of a single prompt. Non-determinism makes it noise.
Screenshots as evidence. Useful for a slide, useless as a trend.
Your own logged-in sessions. Chat history contaminates results badly.
Total AI referral volume in isolation. It undercounts by design.

Reporting cadence that worksWeekly: crawler hits and status codes. Monthly: full prompt panel with all three rates. Quarterly: recommendation-rate trend and entity audit. Anything more frequent measures noise.

Next: the boundaries we hold to — white-hat rules.

Questions engines askFrequently asked questions

Why do I get different answers each time I run the same prompt?

Generative engines are non-deterministic, personalise on session context, and change retrieval results as the web changes. That's why the method is a panel — a fixed set of prompts run repeatedly, scored as rates rather than as individual results. Run each prompt three times in a fresh session and record the proportion of runs you appeared in.

Is AI referral traffic worth tracking if the volume is tiny?

Yes, because the composition matters more than the volume. Visitors arriving from an assistant have typically already had their question answered and are further along in a decision, so conversion rates often run well above organic search. Track it as a separate channel from day one so you have a trend line when the volume grows.

Do I need a paid AI visibility tracking tool?

Not to start. A spreadsheet, three browser sessions and your server logs will give you a defensible baseline. Paid tools earn their cost when you need scale — hundreds of prompts across multiple engines and markets, tracked continuously — or when you need reporting you didn't build yourself. Do the manual version first so you understand what the tool is counting.