Short answer: there are four independent signals, and they measure different things: manual prompt tests show whether you are cited at all, analytics referrers show whether citations produced visits, server logs show whether AI crawlers fetched your pages, and structured sampling shows frequency over time.
1. Manual prompt testing
Open ChatGPT, Perplexity, Gemini or Copilot and ask the questions your customers ask. Look for two separate things:
- Mention — is your brand named in the answer?
- Citation — does the answer link to your domain as a source?
In Perplexity and Copilot, citations are usually visible as numbered sources. In ChatGPT, sources appear more often when browsing or search grounding is enabled. Ask "which sources did you use" if links are not shown.
This is free and immediate, but it is a single sample on a single day. It tells you yes/no, not how often.
2. Analytics referrers
In GA4, inspect traffic where the referrer is an AI platform domain. The list changes over time; common entries include chatgpt.com, chat.openai.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai.
Practical notes:
- Referrals only capture clicks. A citation with no click is invisible here.
- Many AI-originated visits arrive as direct traffic or with no referrer, so referrer data is a lower bound.
- Some tools open links in an in-app browser that strips the referrer.
If you want to see this without configuring exploration reports by hand, Bidvix can connect a GA4 property and surface AI referral sessions, users and revenue directly.
3. Crawler logs
Server or CDN logs show whether GPTBot, OAI-SearchBot, ClaudeBot or PerplexityBot have fetched your pages, and which URLs they requested.
This is the most reliable signal of access, and it is diagnostic: if crawlers never fetch a page, that page cannot be cited. Check both status codes and paths — a 403 from a WAF looks identical to a 404 when you only count requests.
Caveats: log retention is often short, user-agent strings can be spoofed, and a crawl does not guarantee inclusion in an answer.
4. Structured sampling
The rigorous method: fix a prompt set, run it across engines on a schedule, store the raw answers, and count mentions and citations per prompt.
Rules that make this number defensible:
- Keep the prompt set stable for at least a quarter, and version it when you change it.
- State the sample size and the date range for every reported figure.
- Store the raw answer text, not only the derived counts.
- Record which engine version and region produced each answer.
- Sample enough to survive variance — one run per prompt per month is not a trend.
This is what a GEO monitoring product automates. Bidvix does it on a smaller, diagnosis-oriented scale: it generates the customer questions for your category, samples the configured engines, and reports mentions, citations and competitor share of voice, with every counted answer inspectable.
Combining the four
A complete picture looks like this:
| Question | Best signal |
|---|---|
| Can AI crawlers reach us? | Server logs + robots.txt check |
| Are we named in answers? | Structured sampling |
| Are we the cited source? | Structured sampling with citations |
| Did citations produce visits? | Analytics referrers |
| Did those visits convert? | Analytics conversion data |
If logs show no crawl traffic, the other signals will be zero regardless of content quality. Fix access first.