Why can't I just count referrals from ChatGPT and Gemini?

Because most AI-influenced visits never carry an AI referrer. Similarweb (June 2026) found 55.9% of AI-influenced visits arrived via a search engine and only ~19.9% arrived directly. People re-search the brand after an AI names it, so the visit lands as branded search or direct traffic. Counting only labelled AI referrals captures a fraction of the real total and undercounts the channel.

How do I tell delayed AI visits apart from ordinary direct or search traffic?

Use behavioural fingerprints, not a single referrer. Watch for a lift in branded search that tracks your AI citations, a rise in 'Direct' sessions landing on deep, content-rich pages rather than the homepage, and a higher new-user share on those sessions. None is proof on its own, but together, and correlated in time with your tracked AI visibility, they give a defensible estimate of the delayed AI-driven demand.

How to measure AI visits that arrive days after the prompt

Q: What attribution window should I use for AI-referred traffic?

At least 7 days, not a same-session or same-day window. Gener8's July 2026 data shows the prompt-to-visit rate rising 9x to 10x between 1 hour and 72 hours after a prompt, and Similarweb measured its 2.5x visit lift over a 7-day window. A last-click, same-session model attributes almost none of that to AI. Match your window to the behaviour: intent planted in an AI answer often converts days later.

To measure AI visits that arrive days after the prompt, widen your attribution window to at least seven days, treat branded search and direct as AI-influenced channels rather than waiting for a labelled referral, and read the deep-landing-page and branded-search fingerprints as your signal. AI-referred traffic is late and indirect by nature — the measurement has to match. This how-to turns the prompt-to-visit lag into a repeatable method.

Step 1: Widen the attribution window to at least 7 days

The default same-session or same-day window is the first thing that hides AI. Gener8's July 2026 data shows the prompt-to-visit rate climbing 9x to 10x between the 1-hour mark and 72 hours after a prompt; Similarweb measured its 2.5x visit lift over a 7-day window. Set a look-back of at least seven days — longer for research-led categories like B2B SaaS, where visits keep accumulating past 72 hours. A last-click model that credits only the final touch will assign almost none of this to AI.

Step 2: Stop waiting for a labelled AI referral

Most AI-influenced visits never carry an AI referrer. Similarweb found 55.9% of them arrived via a search engine and only about 19.9% arrived directly — the person re-searches the brand after an AI answer names it. So instrument the channels where the visit actually lands:

Branded search. Track branded-query impressions and clicks in Search Console as a first-class AI signal, not a vanity metric. A lift here that follows your rising AI citations is AI demand arriving through search.
Direct-to-deep-page. Capture the dark traffic that lands with no referrer, as covered in why your AI traffic shows up as "Direct".
Labelled AI referrals. Still worth a dedicated GA4 channel — just treat it as the visible minority, not the whole channel.

Step 3: Baseline the fingerprints before you change anything

You can only spot a lift against a baseline. Before a visibility push, record your normal levels of: branded-search volume, the share of "Direct" sessions landing on non-homepage URLs, and the new-user share on those deep-landing sessions. Someone typing your URL lands on the homepage; someone acting on an AI recommendation lands deep, on the specific page that answered them. A rise in deep, first-time "Direct" sessions is one of the cleanest fingerprints of AI-driven demand.

Don't hunt for a single referrer that proves AI sent the visit — it usually isn't there. Watch three signals move together: branded search up, deep-landing direct sessions up, new-user share up, all in step with your tracked AI citations. That correlation is the honest evidence.

Step 4: Segment by vertical and intent

A blended number lies. Gener8's prompt-to-visit rate ranges about 4x across categories (B2B SaaS 14.1% vs Apparel & Fashion 3.5% at 24 hours), and the shape of the lag differs too. Split your measurement by product line or intent type so a fast-converting category doesn't mask a slow one — and so your benchmark compares like with like.

Step 5: Tie delayed visits back to the prompt that caused them

A late, indirect visit is only half the story; the other half is which AI answer set it in motion. Pair your traffic view with upstream visibility tracking — which prompts surface your brand, on which engines, and whether your pages are the cited source. When a branded-search lift follows a spike in citations for a specific topic, you have connected cause to effect across the lag. This is the join that last-click analytics can't make, and it is the core of the reporting stack that ties AI visibility to converting traffic.

Step 6: Give it a quarter, and read it as a curve

Two lags stack here. The citation lag means a new page may not be cited for weeks; the prompt-to-visit lag means each answer's visits then arrive over days. Add the three-month citation cliff and the honest evaluation window for a new page is a full quarter. Read AI's impact as a curve that builds and decays, not a same-day spike — and refresh competitive pages before the cliff rather than judging them dead early.

The method in one line: widen the window, measure search and direct as AI channels, fingerprint the deep-landing sessions, and connect the delayed visit to the prompt behind it. Doing that continuously, across every engine and over time, is exactly what Buffy Intel is built to provide.