Three different things are often called AI traffic
An assistant mentioning your website, a crawler requesting a page and a person opening a link are different events. Only the last is a potential visit to your product. Even then, a recorded request is not automatically a qualified prospect or a payment. Start an AI traffic report by keeping these categories separate.
This matters when deciding whether to invest in content. A page might receive many automated fetches and few customers. Another might receive only a handful of identifiable assistant referrals but attract relevant buyers. Neither citation counts nor bot request counts should be added to the human visitor denominator.
Use the source evidence you actually receive
Inspect the source and referring hostname attached to recorded visits. For a clearly labeled AI-source segment, maintain a reviewed list of assistant domains observed in your data. If a URL contains a campaign source naming an assistant, treat it as a submitted label, not cryptographic proof of origin. People can share tagged links outside the original context.
Prefer explicit hostname comparisons to a broad text rule such as “contains ai.” That substring can match unrelated sites. Record the mapping version and inspect unknown sources before adding them. Sources and browser behavior change; an old screenshot or tutorial is not an integration contract. Do not reconstruct missing assistant prompts from a landing-page title.
In Google Analytics, current documentation defines an AI Assistant default channel. It excludes Google AI Overviews and AI Mode, which are included in Organic Search. Inspect source-level rows as well as channel totals; a channel label answers a classification question, not which answer or prompt convinced the visitor.
Keep crawlers in an operational report
Server logs can help check whether an assistant’s infrastructure can retrieve your content, but logs require their own validation and privacy controls. OpenAI documents different purposes for OAI-SearchBot, GPTBot and ChatGPT-User. A request from one of those agents does not establish that someone visited your site in a browser or purchased anything.
User-agent strings can be copied. Where a provider offers verification guidance, use it rather than treating any matching string as authenticated. Store only what your operational purpose requires. Do not paste raw customer URLs, tokens or log exports into a marketing spreadsheet.
Separate reporting lanes
Answer observed in a test -> citation/visibility observation
Verified crawler request -> retrieval/operations observation
Recorded AI-source visit -> acquisition observation
Verified linked payment -> attributed financial observation
These lanes are related, but their counts are not interchangeable.Build a small, useful AI-source report
Include reporting period, timezone, latest completed ingestion time, recorded visitors or sessions, landing pages, activation outcomes, verified new-customer payments, refunds and missing attribution. Use the same unit and conversion window for other channels so the comparison remains meaningful. Keep each currency separate.
For a synthetic example, imagine 80 recorded AI-source sessions and four verified new-customer payments linked under a stated 14-day rule. That is an observed 5% session-to-payment rate under that rule. It is not proof that every AI exposure converted at 5%, nor that the channel caused those purchases. If two refunds later reduce receipts, show them rather than retaining the launch-day headline.
With four purchases, one extra payment changes the story substantially. Inspect the actual landing-page intent and product fit before reallocating a budget. Compare mature cohorts and annotate changes to source classification; otherwise a reporting-rule update can look like channel growth.
Understand what remains invisible
Someone may read an answer, remember your name and return through a bookmark or a branded search. The later recorded visit does not reveal the earlier influence. A missing referrer cannot safely be relabeled AI merely because the landing page is often cited. Keep unknown journeys unknown.
Voluntary “How did you hear about us?” feedback can add context, but self-report and technical attribution measure different things. Preserve both fields rather than overwriting one with the other. Likewise, a manually tested prompt is a single observation affected by time and context—not a census of what every user sees.
Google documents AI Overviews and AI Mode within overall Web Search performance in Search Console. Do not equate all Google organic traffic with AI traffic, or claim an exact AI-only revenue figure from those aggregate totals.
Improve the content, not just the tracking label
Make important explanations accessible as readable text, answer specific questions, link related material and keep examples and source claims accurate. Google says its normal SEO practices remain relevant for its AI search features and does not require a special AI markup format. Inclusion is not guaranteed.
The practical next step is to identify one landing page with relevant recorded AI-source visits, test its journey to a useful product outcome and fix a concrete problem. If source information is missing, use the Direct-versus-Unattributed guide before inventing a new label. Attribution can connect evidence you possess; it cannot recreate every conversation that happened elsewhere.