Did AI Send That Customer?
The newest blind spot: clients who arrive from an assistant's answer leave almost no trace. Seeing AI-referred work — imperfectly but usefully — before your dashboards can.
A new client books. Your analytics files them under "direct traffic" — they typed your name straight into the browser. Your source column says "found you online." Case closed, unremarkable.
Except what actually happened was this: they asked ChatGPT who handles cross-border tax nearby, the assistant named you and two others, they glanced at your reviews, and typed your name. The most consequential recommendation in your month — the new front door working exactly as designed — registered in your systems as nothing at all.
The AI measurement problem: assistant recommendations mostly arrive as brand-name searches and "direct" visits — the dashboard categories that mean origin unknown. The channel is real, growing, and nearly fingerprint-free, which means practices are currently making channel decisions on data that structurally undercounts the newest channel. You can't fix the fingerprints. You can triangulate.
The three imperfect instruments
Ask better, the human way. The source question already in your intake gains one follow-up for the vague answers: "found you online — searching, or did something recommend us?" Clients happily say "ChatGPT suggested you" — but almost never unprompted, because it doesn't yet feel like a normal answer. The prompt makes it sayable, and one line in the record makes it countable. This alone puts you months ahead of the dashboards.
Watch the shadows in the analytics. Three indirect signatures, none conclusive, all suggestive: brand-name search growth (people who already know your name — someone or something told them); "direct" traffic rising without a campaign to explain it; and, where your analytics shows referrers, the trickle of visits arriving from assistant domains — undercounted (many assistant hand-offs carry no referrer) but a floor, and a trend.
Correlate with the answer log. The monthly ritual tells you what assistants say; this page's instruments tell you what arrives. When "started being named for estate work in March" lines up with "estate enquiries citing 'found you online' rose through spring," you have the channel's honest picture — assembled from three imperfect sources, the way all real attribution works.
Why bother measuring what you can't measure well?
Because the decisions are arriving anyway. Whether the AI-visibility work deserves next quarter's attention; whether the content assistants cite is earning its keep; whether the "direct" growth is brand strength or a counting artifact — each turns on this channel's rough size. A triangulated "somewhere around a fifth of new enquiries, and growing" is decision-grade; waiting for perfect tracking means deciding blind while competitors who asked the extra question decide informed. Imperfect instruments, honestly labelled, beat missing ones — the whole instrument-panel principle, applied to the fog.
Questions practices actually ask
Will this get easier? Somewhat — referrer signals and assistant-side citations are slowly improving. But recommendation-shaped journeys (told a name → typed the name) are inherently low-fingerprint, like word of mouth always was. The human question will stay the best instrument, exactly as it has for referrals forever.
Should AI-referred clients change how I sell to them? They arrive pre-shortlisted — often having seen exactly what the assistant said about you. Treat them like referred clients: verification-ready, fast to book, no hard sell needed.
My analytics shows zero visits from AI domains. Channel's not real for me? Zero referred visits is compatible with a healthy AI channel — the hand-off usually breaks the trail. Trust the triangulation: the prompted source question and the answer log outrank the referrer report here.
Where does this number live? A source-tag in the record ("AI-recommended"), rolled into the index card's source mix monthly. When it earns a visible share, the whole cluster upstream has proven itself in your own numbers — the dogfood closing its loop.
Part of Measure What Matters — the instrument panel.