Build a four-signal model
The first signal is a referral session: a browser request arrives with a recognized answer-engine referrer and can be analyzed like other acquisition traffic. The second is a verified crawler request used for search discovery or indexing. The third is a user-triggered fetch by an agent. The fourth is a citation or mention reported by an external platform, whether or not it produced a click.
Do not roll the four into “AI visitors.” A crawler is not a prospect, a fetch may not render the page, and a referral does not reveal the answer that prompted it. Keep source labels, verification state, landing page, outcome, and observation method visible.
Measure referral sessions conservatively
Create a maintained mapping of known referrer hostnames to a neutral channel such as AI referral. Preserve the raw referrer in a controlled field so classification can be corrected later. Report sessions, landing pages, meaningful outcomes, and conversion uncertainty; small volumes make percentage swings look dramatic.
Some applications remove referrers or open links through intermediaries, so AI-influenced visits may appear direct. UTM tags can help only when the platform adds them. Call the result “observed AI referrals,” not total AI influence, and avoid assigning all unexplained direct traffic to assistants.
Verify automated traffic
User-agent strings are easy to copy. Follow each operator’s current documentation for published IP ranges, forward-confirmed reverse DNS, or another verification method. Cache verification results carefully, rate-limit unknown automation, and store crawler analytics separately from human product analytics.
Crawler names also express different purposes. OpenAI documents OAI-SearchBot, GPTBot, and ChatGPT-User separately; Anthropic documents search, user-directed, and training-related agents; Perplexity distinguishes its crawler from user fetches. Robots directives should reflect the organization’s intended search and training posture.
Create a citation evidence loop
Track which pages are eligible to be cited: indexable, canonical, accessible, clearly sourced, and answer-first. Monitor answer-engine referrals by landing page and query context when available. Bing Webmaster Tools introduced AI Performance reporting in 2026, providing another platform view beyond site referrals.
Keep a reproducible manual sample for important buyer questions. Record date, platform, model or surface, location, prompt, cited domains, and screenshot. That sample cannot estimate total visibility, but it can reveal factual gaps and whether the page states an answer clearly enough to be selected.
Questions this guide answers
How can a site track traffic from ChatGPT?
Classify recognized referral hostnames, then report those observed sessions and outcomes separately from bot requests and unknown direct traffic.
Can website analytics see every AI citation?
No. A citation without a click creates no website request, so ordinary analytics cannot observe it.
Are AI crawlers the same as AI referrals?
No. Crawlers are automated requests; referrals are browser visits that arrive after a person follows a link.
Primary-source ledger
- OpenAI crawlers and user agentsOpenAI · accessed 30 July 2026
- Anthropic web crawler controlsAnthropic · accessed 30 July 2026
- Perplexity crawler documentationPerplexity · accessed 30 July 2026
Scope note: Referrer mappings and crawler documentation change. Treat attribution as observed evidence, verify operators, and never estimate unseen citations from clicks alone.
Found an outdated fact or a material omission?Read the corrections and recheck policy.