
Two in 100 Product-Naming Answers Lead to the Named Product's Page. Referral Reports Miss Most Visits After an Answer
A Scrunch AI panel study finds that 89% of visits after a store-recommending AI answer carry no sign of the assistant's link, so referral reports miss most of them. Among shoppers with no visit in a week, visits to the named product rise by about two per 100 specific-product answers; the paper names no retailers, withholds its panel size and sees no sales.
Neritus Vale
When an AI assistant names a product, some shoppers go to look at it, and most of their visits follow a web search. That is the finding of a paper by Michael Iannelli and Alan Ai of Scrunch AI, which links a panel’s assistant conversations to the same people’s browsing. It explains why a store’s referral report misses so many of those visits. It also puts the effect at about two extra visits per 100 answers, and it never sees a sale.
The paper’s best evidence is a swap. Take shoppers who had not visited a store for a week and then got an answer naming a specific product there. On days when they opened a product page at that store, 44% showed a product from their own answer, against 1.5% for a stranger’s matched answer. The match follows the answer, not the store’s bestsellers.
That share covers only what the paper calls “the narrower set of days with a product-page visit at the store”. For the same shoppers, visits to the named product within a day reach 2.07 per 100 answers, up from almost none in their earlier weeks. The authors chose this group after seeing first results. Its rise in store visits, 5.1 points, is smaller than the 10.5 points across all answers naming the store.
The case against referral reports rests on the route. In the paper’s count, 89% of visits after an answer recommending a store carry no sign of the assistant’s link; the store logs them as ordinary search or direct traffic. A web search precedes 54% of all visits after such an answer, the same share as the shoppers’ usual visits. The link does better with lapsed shoppers: it carries 42.5% of their extra visits after specific-product answers, against 16.8% for recent visitors.
Referral data have so far made assistants look marginal. About You’s Tarek Müller told Zalando’s partners that chatbots send online fashion stores under 0.5% of traffic, as WWD reported and we noted. Referral reports, the Scrunch paper’s authors write, “see the assistant best when it brings someone back, and least for recent visitors”. Across all visits after a recommending answer, they estimate that link-only reporting catches “about one in five of the extra visits”.
Search is also where a named store can lose the shopper. Iannelli and Ai liken a search for the product’s name to an address bar, but a results page offers many destinations. By our arithmetic from the paper’s figures, 9.3% of answers are followed by such searches, against 2.07% for visits to the named product’s page: more than four times as many. The paper does not say whether the two rates share a base, or where the rest of those searches ended. If the pattern holds for a widely stocked trainer, the results page becomes a contest.
Scrunch sells tools pitched at the blind spot its researchers describe. The company, bought by Sitecore in June, tracks how brands appear in AI answers. Its shopping product promises to show “which retailers AI funnels buyers toward”, under the heading “Know who gets the sale”. The paper’s proposed metric, product follow-through, needs “the answer text and the same person’s later page views”.
“We see visits, not purchases,” Scrunch’s researchers write in the paper.
Nor can an outsider check the base. The paper withholds panel size, per-analysis user counts and cell sizes as “commercially sensitive” and does not name its twenty retailers. Its worked example, a running watch, is invented for illustration, and it has no appendix. The paper does not say whether any of the retailers sells clothing.
The authors’ earlier studies show who is in the panel. Their June paper calls Scrunch’s pipeline a “measurement layer” that “has run in production at Scrunch AI since early 2026”. A revision of their July paper describes the panel as “AI-forward”, adding that assistant use in native mobile apps is “not observed”. Scrunch’s blog calls it “more AI-engaged and privacy-tolerant than the average buyer”, so 2.07 per 100 is not a rate for the average shopper. The November paper does not say whether it draws on that panel.
Their athletic apparel result from that panel is a mixed picture. The June study, “From Prompt to Purchase”, presented at a KDD workshop on 9 August, re-ran its brand analysis on athletic apparel brands. Searches for the brands rose 2.54 points and visits to the brands’ own sites 1.67 points (interval from −0.1 to 3.3), while visits to their pages at retailers moved 0.06 points over a week. “Apparel’s downstream is weak”, the authors wrote, but they also note that retailer URLs often carry no brand string, which makes retail visits hard to count. Scrunch’s blog post said “the lift reproduces, though the destinations differ”, naming destinations for beauty and electronics but not apparel.
An unnamed language model decides what counts. It labels each answer as naming a specific product at the store, suggesting the store or only mentioning it, and judges product matches from page titles and web addresses alone. A blind human reader confirmed 93% of 90 product matches after an answer and 54% of 39 in the baseline. The reader agreed with the answer-type labels on 75% of 102 answers and demoted 29% of the model’s recommendations to mentions. If the reader was right, the authors say, the recommending effects are understated.
Follow-through is pitched as a per-assistant score, but the paper prints it for ChatGPT and Gemini only. It rises 2.2 points after ChatGPT answers against 1.0 after Gemini’s. In the June study, a larger ChatGPT lead was judged “mostly compositional, a difference in who uses each assistant”. Unless compared within the same users, follow-through may rank audiences as much as assistants.
The effect is hard to dismiss on design. Swapped answers, a replication whose analysis plan was committed before any data were pulled, and comparisons with equally busy moments all point the same way. If answers naming a store run into the millions and those visits buy like other product-page arrivals, the undercount is worth paying to fix.
That case needs figures the paper does not give: how many answers name a store, and how many visits end in an order. Kaiser and Schulze’s referral study, which Frankfurt School billed as showing “ChatGPT’s limited role in online shopping”, at least followed visits to the till. It counts only referral sessions, its data end in July 2025, and visits from referral links convert below every established channel except paid social. Without both figures, two visits per 100 answers is a rate without a base.
A retailer can start with the paper’s own advice: report assistant referrals “next to product-page arrivals from search and direct traffic” and track product-name queries in search-console data. Tracking purchases, the paper says, “requires order data from the named stores”. Skipping that step means paying for visibility in answers whose sales nobody has counted.