Define the outcome before counting visits
An AI referral becomes meaningful when it connects to a question about the website. Did visitors find the setup instructions, complete a qualified inquiry, or finish an intended purchase? Choose an outcome that reflects the page’s purpose. A higher visit count alone cannot tell you whether the experience helped anyone.
Document the event definition and its exclusions. A form submission might include spam, internal testing, and repeated attempts by the same person. If those cases remain in the total, call it submissions rather than qualified demand. Precise labels protect the interpretation before any attribution model enters the discussion.
Identify what the arrival actually shows
Inspect the recorded referrer, landing page, timestamp, and available campaign parameters. A recognizable assistant domain can support classifying a visit as an observed referral. Preserve the original value as well as the classification so that a later rule change does not erase the evidence behind the label.
A missing referrer is not proof that AI played no role. Applications, privacy settings, copied links, and intervening navigation can remove that signal. Conversely, a referral does not reveal the original prompt or prove that the visitor accepted the answer. Keep observed traffic separate from inferred influence.
Map a short, inspectable journey
Start with a landing page and follow the next meaningful steps. For a documentation page, that might mean opening the relevant setup section and reaching a completion page. Use consented, appropriately scoped analytics. Avoid collecting chat transcripts or sensitive query details simply because they might make a journey diagram more complete.
Keep visitor counts, sessions, and events distinct. One person can generate several visits and many events. Cross-device journeys may appear disconnected. State the unit used for every ratio, and check whether the numerator and denominator cover the same eligible population and observation window.
Run an illustrative landing-page review
Imagine an assistant links to a setup article, but visitors cannot find the prerequisite checklist. This is an illustrative workflow, not a customer result. Investigate the page before concluding that the referral source brings poor-quality traffic. The problem may be an incomplete handoff from answer to page.
- Confirm that the landing URL resolves to the intended article and preserves useful context.
- Check whether the first screen explains prerequisites and the next practical step.
- Verify the completion event with a documented test and exclude that test from reporting.
- Compare equivalent observation windows after an edit and inspect absolute counts alongside rates.
Use Kept AI evidence in context
Kept AI’s native Agent can analyze a 30-day conversion summary alongside crawler activity, crawlability, robots rules, and paths. Those views can guide investigation, but crawler access is not a human visit and a conversion summary is not proof of causation. Check each metric’s scope before combining observations.
Report with attribution limits attached
Describe results as outcomes associated with observed referrals unless the research design supports a stronger statement. Last-touch attribution can overlook earlier discovery; first-touch attribution can overlook later persuasion. A small number of outcomes can also make rates unstable. Show counts and explain the attribution rule in ordinary language.
Choose the next action from the evidence
Prioritize a specific fix: clarify prerequisites, repair a broken destination, or improve event definitions. Preserve the before-and-after record and note simultaneous campaigns or site changes. Referral analysis is strongest when it improves a verifiable visitor experience while keeping the limits of the measurement visible.