TL;DR
This paper introduces "product follow-through," a metric that links AI assistant conversations to subsequent web browsing behavior. The authors reveal a massive attribution gap: while AI recommendations significantly drive shoppers to specific products, 89% of these visits leave no referral trace, as users prefer searching for product names rather than clicking assistant-provided links.
What Problem Does This Paper Address?
In the traditional e-commerce funnel, retailers track "where customers come from" using referral links and UTM parameters. However, the rise of AI assistants (like ChatGPT, Claude, and Gemini) has created a blind spot. When a shopper asks an AI for a recommendation, the assistant might name a specific product, but the shopper often navigates to the store by typing that product name into a search engine or address bar later.
Because the store’s analytics only see the final search click, they credit the traffic to SEO or "organic search," completely missing the AI's role in forming the initial purchase intent. This paper quantifies this "invisible" influence and proposes a way to evaluate AI assistants based on actual user behavior rather than just the quality of the text they generate.
Methods
The researchers utilized an opt-in consumer panel to link the same individuals' AI chat logs with their subsequent browsing data across 20 major online retailers. To isolate the effect of the AI's advice, they employed several rigorous comparison techniques:
- Earlier-Week Baseline: Comparing a user’s behavior after an AI answer to their own behavior at the same store 14, 21, and 28 days prior.
- Matched Moments: Comparing visits after an AI answer to "control" moments where the user was active online but hadn't received a recommendation.
- Swapped Answers: Testing if users visited products recommended to other people (to control for general product popularity).
- LLM Judges: The authors used language models to classify whether an AI answer named a specific product and whether a visited URL actually showed that specific product.
Datasets
The study draws from a large-scale, opt-in consumer panel in two English-speaking markets.
- AI Assistants: Conversations from ChatGPT, Gemini, Claude, Perplexity, and Grok.
- Retailers: Browsing data from 20 of the largest online retailers.
- Timeframes: A main study period (July–October 2026) and a five-month "held-out" replication period (February–June 2026) to ensure the results were robust.
Key Insights and Ideas
- The Search Box is the New Address Bar: Shoppers treat the search engine as a bridge. After an AI names a product, searches for that specific product name jump 24x compared to the baseline.
- Referral Links are Leaky: Even when assistants provide links, shoppers rarely use them. 89% of visits following a recommendation showed no sign of an AI referral. The link is most likely to be used by "new" or infrequent visitors; regular customers simply go to the store their own way.
- AI as a Browsing Catalyst: AI recommendations don't just lead to a "one-and-done" visit. Shoppers who follow an AI's lead are 5.2 percentage points more likely to view multiple products than they are during their usual visits. The AI gets them in the door, and then they keep browsing.
- Attribution Failure: Retailers are currently "flying blind." By the time a shopper reaches a product page, the "decision" was often made minutes earlier in a private chat that the retailer cannot see.
Key Results
- Product Follow-Through: Visits to a specific product within 24 hours of an AI naming it rose from a baseline of 0.02% to 2.07% (a hundred-fold increase in relative terms).
- High Intent: On days when a user visited a product page at a recommended store, 44% of those users viewed the exact product the AI had named.
- Assistant Variance: Follow-through rates varied by tool; ChatGPT saw a 2.2 point rise in follow-through, while Gemini saw a 1.0 point rise.
- Price Matters: Answers that quoted a specific price near the store's name saw a significantly higher follow-through (+0.93 points) than those that didn't.
Limitations and Future Work
The authors acknowledge that this is an observational study, meaning they cannot strictly prove the AI caused the visit (though their use of baselines and matched moments strongly suggests it).
Future directions include: * Conversion Data: The current study tracks page views, not confirmed purchases. Linking this to "order" data would be the "holy grail" for retailers. * Causal Testing: Randomizing which products an AI recommends to different users to measure the delta in traffic more precisely. * Agentic Shopping: As AI moves from "answering" to "acting" (e.g., AI agents that buy products on a user's behalf), the metric will need to shift from page views to successful transactions.
Summary updated