How to measure AI bot access without confusing it with customer traffic

Measure access first

A bot request can tell you that an agent reached a page or encountered an error. It does not tell you whether a person saw your brand in an answer. Start with access diagnostics before adding customer-performance labels.

OpenAI's agent documentation separates search crawling, potential training collection and certain user-triggered page visits. Keep those purposes distinct in your report. Check current published agent identities rather than relying on a hard-coded list from an old article.

Keep a minimal, useful log

For each verified agent, record the requested URL, timestamp and status. Where your infrastructure exposes it, also record response size and the blocking rule that caused a failure. Use the provider's published network information to validate identity; the user-agent label by itself is not authentication.

Compare successful fetches with blocked, redirected and missing-page responses. Investigate important pages that fail repeatedly. A healthy request count is not a reason to ignore an inaccessible pricing, location or product page.

Then look for answer evidence

Capture relevant questions with the visible model, date and market. Save the final answer and the cited URLs. Score your brand's appearance separately from your domain's citation: an assistant can cite your research without recommending your product.

Repeat the captures. In Aiso's variance study, identical prompts did not consistently produce identical brand lists. That finding is about a fixed test set, not how many people saw each recommendation.

Add human outcomes without inventing the missing join

Measure AI-referred sessions in analytics and qualified enquiries or completed orders in your conversion system. Keep the source and attribution model visible. Do not divide human visits by agent requests and call the result an AI click-through rate.

An impression-based click-through rate needs an actual exposure denominator. Site logs do not provide all ChatGPT answer exposures, and one agent request need not correspond to one visible answer. Report a referral conversion rate only when both conversions and sessions use the same population and period.

Choose the action from the failed stage

Fix access failures with the technical team. Fix inaccurate product facts at the cited source. Investigate conversion friction on the landing page when qualified human sessions fail to complete the intended action. Those are different problems, even when they appear in the same dashboard.

For implementation details, use the ChatGPT measurement setup. For query metadata, use the fan-out extraction tutorial.