Connect GA4 — read-only, one property, no tag to install. Citino matches each AI-referred session back to the prompts whose stored answers cited the page it landed on, so you get sessions and conversions against individual questions instead of one undifferentiated "AI" row.
When more than one prompt qualifies and none is clearly the strongest, the session is not split and not assigned to the likeliest. It goes to unmatched.
Resolving ambiguity toward the residue is why the residue is 29% and not 4%. Every matched figure opens to the answers underneath it. → Answers
Both parts drawn at the same weight. The 360 sessions nobody can explain are shown as exactly that.
This is a heuristic, and the page says so where the number is. There is no click ID on an answer, no UTM the engine carries forward, no callback from the model. There is a referrer, a landing page, a timestamp and a corpus of answers we collected ourselves. The column header reads matched, not source.
Spread the unmatched 29% across the prompts that did match and the values sum neatly to 1,240. It also invents those sessions' causes and inflates every prompt by the same flattering factor. If a vendor's AI attribution has no unmatched bucket at all, the residue did not disappear; it was allocated.
One question out of 54 is carrying more than most of the rest put together. Until you can see that, effort gets spread evenly across a prompt set whose value is not spread evenly at all.
A prompt can send volume and convert nothing; another sends a tenth of the traffic and most of the demos. Neither ranking is the visibility ranking.
Conversions are your own GA4 key events, unchanged. → Act
Your GA4 will show sessions from copilot.microsoft.com and from claude.ai. Both appear by name and carry a not sampled label. There is no measured answer set behind them, so they land in the 29% by construction rather than by failure.
Scraping a consumer interface would produce rows that look identical to measured ones — a wrong number indistinguishable from a right one corrupts every number beside it. And sampling Claude through its API is not sampling the claude.ai website.
Counted in the total, excluded from prompt value, labelled in both places.
That session carries no AI referrer. It is not in the 71%, not in the 29%, and not in the 1,240 at all. For a brand at 10.3% visibility that unmeasured effect is plausibly the larger one.
Prompt value is therefore a floor, not a total. A number that is definitely too small, stated as such, is more useful than one that is probably too big and cannot be checked.
And a session is not a sale. The conversions here are the key events already defined in your property, with the same last-touch limitation they have always had. → Methodology
Read-only, one property, and the match runs against answers already collected — so the first report covers the thirty days behind you, not the thirty ahead.