What Shoppers Take From AI Chat to Search
A shopper can reuse a product name from an AI recommendation to look up the item outside the chat, even when the store’s traffic report shows only a search visit.
Cornelia Reitinger had tried several websites looking for a yellow umbrella for her terrace. A chatbot found the one she wound up buying, she wrote in September. Before she bought it, she checked the seller and the umbrella’s full dimensions herself. She doesn’t say whether she clicked a link from the chat or searched for the umbrella after getting the answer.
That missing step matters less to a shopper than it does to a store trying to tell where its visitors came from. A person who gets a candidate from an AI chat can look it up by name, open a listing and check the details there. If they open the listing from a search result, a store’s traffic report can correctly say they came from search without showing how they knew what to look for.
A November preprint from two researchers at Scrunch AI matches assistant conversations and browsing records from the same volunteers in two English-speaking markets. Across 20 large online retailers, the researchers could look for the product from an answer in a later search query or on a page the shopper opened. Those records can show whether the same item turned up again, not whether the answer gave someone the idea or settled what to buy.
What a search visit leaves out
After answers that recommended a store, with or without naming an item, the researchers classified 54% of store visits as following a web search, almost the same as on the shoppers’ usual visits. The search category included a recent search even if there was no recorded click from it. Only 11% had an identifiable sign of an assistant link, such as a marker in the web address. That doesn’t rule out a link click for every unmarked visit. Nor does it give the assistant credit for all the traffic that followed its answer.
The search terms tell you something the entry route can’t. When a web search occurred in the 30 minutes before a shopper’s first visit to the named store after an answer, it named a product from the answer 28% of the time. In the same shoppers’ earlier comparison windows, the figure was 1%. From the route alone, both look like search. Seeing the query alongside the answer shows how specific the connection can be.
For the product-page comparison, the researchers focused on answers that named a specific item at a store the assistant had introduced and the shopper hadn’t visited in the previous week. Among days when those shoppers opened a product page at that store, 44% included an item from their own answer. When the researchers scored those same days against a different shopper’s answer matched on factors such as store, request type and product category, the share was 1.5%. The gap appeared in an earlier five-month period too.
That is a striking match, but it describes days with a product-page visit. Count all qualifying answers, including those followed by no visit, and about 2 in 100 were followed by a view of the named product at that store within a day. The equivalent rate when the researchers looked at those shoppers’ browsing 14, 21 and 28 days earlier was 0.02 in 100.
Swapping answers makes it harder to explain the result by saying everyone already goes to the popular item. It doesn’t tell us whether the shopper already had a particular item in mind before opening the chat. A shopper who had just heard about an umbrella might have asked whether it would suit the terrace and then opened a listing they were already inclined to look for. The researchers chose their main group after seeing initial results and don’t disclose how many volunteers were in the panel. The numbers describe these shoppers and stores, not online shopping as a whole.
Finding the model, checking the size
The researchers could see that browsing sometimes continued beyond a single item. With arrivals through detected assistant links left out, shoppers viewed at least two products on about 31% of visits after a recommendation, versus 26% on their other visits. There’s no way to tell from those page views whether someone was choosing between options, checking a recommendation or browsing for something else.
The researchers let a different size or color of the same item count as a product match. That’s reasonable when you’re studying whether the item shows up again. It’s not enough if you’re testing whether the advice helped someone find an umbrella that fits. Reitinger had a candidate, and she still needed to check who sold it and read the dimensions. A visit to the right listing might help someone reject a recommendation, too.
If you’re evaluating a shopping assistant, it’s worth seeing what happens when the shopper tries to pick up the task somewhere else. A direct link is useful, but a listing can be out of stock or available only in the wrong size. The exact model name gives the shopper something to search for at another seller; the size gives them something to check when they get there.
Sources
- Michael Iannelli and Alan Ai, Straight to the Product: What Online Retailers Miss When Shoppers Ask AI First, November 16, 2026 preprint, version 1.
- Cornelia Reitinger, Will agents replace websites?, September 9, 2026. First-person account, separate from the browsing study.