Do AI assistant brand mentions drive real user behavior?

The paper measures what happens on the open web after a conversational assistant — ChatGPT, Claude, or Gemini — names a brand to a user who has no recent observed engagement with it. On an opt-in panel that joins pageview clickstream to the same users' chat conversations, "that user's same-name Google search rises +4.3 percentage points (pp) [3.1, 5.5], visits to the brand's own site +2.4 pp [1.4, 3.5], and brand-specific retailer-page visits +1.0 pp [0.3, 1.7] over matched backward placebos." The placebo is within-user: each outcome is differenced against the same user's rate over equal-width windows 14, 21, and 28 days earlier, so the comparison is a person against their own recent past rather than against other users. The panel has run in production at Scrunch AI since early 2026, and the paper reports a disclosure-safe aggregate of that deployed pipeline, excluding search-embedded surfaces like Google's AI Overviews because a "prompt" there is a search the user already typed.

The mechanism the paper gives for why this matters: last-click attribution and standard web analytics have no row for the assistant mentioned this brand, so none of this lift is visible to an advertiser's existing tooling. The naming event is upstream of the logged touchpoints (a Google search, a site visit, a retailer page) and gets credited to "organic" or "direct." The paper's own term for the mechanism is "acquisition-like reach": the effect concentrates among users with no recent observed brand engagement, while among already-engaged users a mention "coincides with an active episode without clearly accelerating it" — so the assistant's measurable role is bringing new people into the funnel, not speeding up people already in it. Three layered defenses rule out the obvious alternative explanation that the user was already going to search anyway: a pre-trend event study, a stance classifier (recommend / neutral / caution) that separates genuine recommendations from incidental references, and a same-category control showing the named brand moves far more than unnamed brands mentioned in the same response.

This sits alongside two neighboring notes on the same underlying phenomenon — assistants surfacing brands — from different angles. How consistent are AI brand recommendation lists across repeated prompts? measures whether a brand gets named at all and how stable that naming is; this paper measures what happens after a brand is named, to a specific user. Do AI search engines favor well-known brands over others? is similarly about citation rate rather than downstream user behavior. Do shopping agents respond more strongly to price and rating cues? and Does limiting product choices increase trust in voice assistants? both measure AI influence on a user's choice inside an agent-mediated session; this paper measures influence that plays out afterward, off-platform, on the open web, with no agent present at the moment of action.

The design is observational: the panel does not observe transactions, so the retailer-page visit is "purchase-adjacent," not a logged sale, and a reverse-causation scenario — a brand-specific intent shock that both triggers a recommendation and independently drives the user's search and visit, within the same session — "remains unidentifiable on observational data; a randomized or encouragement design would be needed to close it." The paper also does not claim this generalizes beyond the brand lexicon and two English-speaking markets it tested. What it does establish, carefully bounded, is that a brand named conversationally to an unengaged user is followed by more of that user's search, site, and retail activity than the same user's own recent baseline — a touchpoint that existing attribution models do not instrument.