MEASUREMENT · PRACTICAL FIELD GUIDE
Measure AI-influenced buying without inventing attribution
Keep observed referrals, customer recollections, AI citations and completed outcomes separate so your next marketing decision rests on evidence.
By Ashley Kays ·
An attribution gap is a question to investigate
Impact research describes shoppers who find information through AI and later navigate manually. That can leave an incomplete referral trail.
It does not follow that unexplained direct traffic came from AI. Bookmarks, missing campaign tags, privacy controls and other journeys can produce similar records. Treat the gap as a reason to improve your evidence, not to reassign traffic until the dashboard looks more impressive.
Create four separate views
| View | Record | Limit |
|---|---|---|
| Observed visits | Available source, medium, landing page and date | Untracked influence is missing |
| Customer-reported discovery | Optional answer about how they found you | Memory and response bias |
| AI answer observations | Prompt, service, timestamp and cited URL | A sampled answer, not a visit |
| Business outcomes | Accepted leads, orders, refunds and fulfilment | Not automatically attributable to one channel |
Use the same reporting window and time zone, but do not add these columns together. They describe different things. One customer may appear in several views.
Define the conversion before looking for the source
For an enquiry site, count a new accepted enquiry after the application confirms success. A button click and a validation failure are different events. For a store, reconcile analytics with actual orders and cancellations before discussing revenue.
Document consent behavior, known collection gaps and reporting delays. Keep names, email addresses and free-text enquiry contents out of analytics parameters. When tracking is unavailable, the business workflow still needs to work.
Use a small prompt panel for visibility
Write a handful of real buying questions, including constraints such as location, budget or compatibility. Repeat the same prompts on selected services under recorded conditions. Save the first answer, citation and factual checks. Keep branded accuracy questions separate from unbranded discovery questions.
Report the denominator: “Our page appeared in three of ten recorded runs” describes the sample. It does not mean a 30% market share or a predictable conversion rate. Do not discard inconvenient answers, failures or missing citations.
Add customer context without forcing a story
Where appropriate, ask an optional question such as “Where did you first hear about us?” Offer search, social, a recommendation, an AI assistant, another source and not sure. If people mention more than one, preserve that nuance in your private business records.
Use aggregate answers alongside observed referrals. Do not overwrite a measured source with a guess, or attach a customer’s private text to public reporting. Small samples should be shown as counts as well as percentages.
Choose a decision rule before the review
Pick one question the data can help answer: which landing page needs a clearer answer, whether a partner’s visitors are relevant, or whether a new guide earns useful next steps. Record the proposed change, expected behavior, review window and quality guardrail.
For example, better-qualified enquiries may justify keeping a clearer page even if total visits do not rise. More citations with no meaningful visits may justify continued observation rather than a larger budget. If you need causal evidence, design an appropriate comparison or experiment; attribution reports alone cannot provide it.
Put it to work
Use the lead-event verification guide to check one real outcome. Then build the four-view record above for a single landing page and a complete reporting period.
Sources and editorial context
Reviewed November 3, 2026. This is independent editorial analysis. The exercises, sample briefs and decision rules are our recommendations; they are not results reported by the sources or measured client outcomes.