The ChatGPT funnel: what you can measure and what you cannot

Start with separate records

A web server sees requests to your site. It does not see every ChatGPT answer, every person who reads it or every source the model considered. A useful report keeps that boundary visible instead of constructing an impression count from crawler traffic.

OpenAI documents different purposes for its agents. OAI-SearchBot supports search discovery. GPTBot collects content that may be used for training. ChatGPT-User is used for certain user-initiated actions and may fetch a page. None of those descriptions turns a request into a human click or a visible brand recommendation.

Keep the unit with each measurement

Record	What it supports	What it does not establish
Verified agent request	A particular URL was requested at a particular time	A unique person, answer impression or purchase
Captured AI answer	A citation or brand mention in that captured response	How many customers saw similar answers
Human referral session	A visit attributed to an AI source under your analytics rules	Every AI-influenced visit or an incremental customer
Completed conversion	An order or qualified action tied to the measured visit	Revenue caused solely by AI discovery

Set up the server-log view

Keep timestamp, requested path, response status, user agent and the minimum network information needed to verify the bot. Compare the source address with the provider's published verification information. A user-agent string alone can be copied by another client.

Group the verified requests by agent and URL. Track failed fetches and blocked responses so the technical team can fix access problems. Label the chart “verified agent requests”, not “ChatGPT impressions”. Repeated requests can come from retries, different actions or the same session.

Build the human-referral view separately

In your analytics tool, inspect session source and medium for recognised AI domains. Keep the source list versioned and review unfamiliar referrers rather than automatically grouping every bot or tool domain as an AI customer.

Use GA4's Traffic acquisition report to inspect session-level acquisition. Report completed orders or defined key events alongside sessions. Preserve the distinction between first-user acquisition and session acquisition.

A missing referrer does not prove that AI played no role. A person can copy a URL or return later by another route. Conversely, do not assign all direct traffic to AI. A short discovery question on a qualified-lead form can add context, provided you keep it separate from observed referral attribution.

Use the same definitions across site platforms

On WordPress, Shopify or a custom site, start from the data actually available to you. Analytics can record human sessions; CDN or hosting logs may record agents. A plugin does not create access to ChatGPT's internal impression count merely by recognising its user agent.

Do not join a bot request to a later human visitor solely because the timestamps are close. There is no reliable person-level join in that observation. Minimise personal data and follow the consent and retention rules applicable to your analytics setup.

A monthly report that can be audited

Show verified fetches and failures, sampled citation coverage, named recommendations, AI-referred human sessions and completed outcomes as distinct series. Include dates, sample sizes and definitions. A rise in fetches can help diagnose discovery or access, but it is not a conversion-rate denominator.

For a captured prompt set, repeat the same questions and preserve model and market context. For revenue, compare completed outcomes over a defined period and document other campaigns or site changes. Call an association an association unless the design supports an incremental effect.

The fan-out query guide explains how to inspect the searches behind a response. The metrics guide explains which answer-level rates to calculate.