A customer may receive an AI product recommendation and later search for your brand on Google. Would your report show where that journey began? Research joining conversations to browsing records finds more brand searches, own-site visits and retailer-page visits after recommendations to users with no recent observed brand engagement. DijitalPi commentary: Measuring only traffic arriving through assistant links can miss part of the journey. However, this observational preprint contains no transaction or sales data and cannot establish that the recommendation alone caused the subsequent behaviour.
ACADEMIC WORK REVIEWED
From Prompt to Purchase: How AI Brand Recommendations Move Consumers on the Open Web
ResearchersMichael Iannelli, Alan Ai.
arXiv · v1: 9 June 2026 · Reviewed v2: 31 August 2026 · Preprint · Source language: English
English review: 28 September 2026 · The source date appears in the citation above.
Editorial review draft
The source has been checked; the review still carries its original editorial status. DijitalPi did not conduct a customer experiment for this paper.
LET'S READ THE RESEARCH TOGETHER
We first explain the researchers' question, method and findings. We then discuss how to interpret the results, clearly separating DijitalPi's commentary from the source.
01 / WHAT DID THE RESEARCHERS WANT TO UNDERSTAND?
What happens on the web after a brand recommendation?
The study examines the relationship between assistant brand recommendations and later brand searches and website visits. It tries to distinguish an earlier exposure invisible to last-click reporting from brand interest already underway.
Explaining the context · DijitalPi commentary
DijitalPi commentary: A customer may learn a brand name in a conversation and later arrive through search. Crediting the final visit to search does not reveal that earlier exposure. Conversely, a mention during a conversation with an existing user does not turn their subsequent visit into new interest created by AI.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
Scrunch AI researchers joined an opt-in panel’s clickstream, the sequence of visited web pages, to the same people’s ChatGPT, Claude and Gemini conversations. The observational analysis covers 2 English-speaking markets and uses a user-response-brand record as its unit. Google AI Overviews and AI Mode are excluded. Panel and subgroup participant counts are withheld as commercially sensitive.
Understanding the method · DijitalPi commentary
The headline comparison concerns genuine recommendations to people with no observed brand search, own-site visit or retailer engagement in the preceding 7 days. This does not prove they have never been customers. Behaviour in the following 7 days is compared with equal-length historical windows for the same user, starting 14, 21 and 28 days before the response. The empty immediately preceding week is not the baseline used to calculate the difference.
For each eligible user-response-brand record, the outcome is whether at least one qualifying brand search, own-site visit or retailer-page visit identifying the brand in its URL occurs within the window. Recommendations are separated from neutral mentions, with controls using unnamed brands in the same category. These checks address pre-existing interest but do not create randomized exposure.
03 / RESEARCH FINDINGS
Search and visits occur more often after recommendations.
In Table 3, for recommendations to users without recent observed engagement, the rate of Google searches containing the brand is 2.9% in the historical comparison and 7.2% in the following 7 days, a difference of 4.3 percentage points. Own-site visit rates are 3.1% and 5.5%, a 2.4-point difference. Brand-specific retailer-page visit rates are 0.8% and 1.8%, a 1.0-point difference. Denominators are eligible user-response-brand records, not unique customers, orders or all assistant users. These differences are observational contrasts, not relative percentage increases.
DijitalPi commentary: Measure direct assistant referrals, but do not assume those links capture the whole journey. Branded search and customer accounts of discovery can add context. Guessing an unseen origin and assigning all search visits to AI would introduce a different measurement error.
The paper’s non-customer definition depends on what is observable. Offline purchases and untracked app activity can remain invisible. Brand search, own-site visits and retailer visits are not compulsory steps traversed in the same order by everyone. Do not multiply these rates as successive sales conversions.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
The reviewed v2 is a preprint. Despite the purchase wording in its title, the study observes no transactions, revenue or completed sales. Brand interest arising within the same session could influence both recommendation and action, leaving causality unresolved. Both authors are affiliated with Scrunch AI and the data come from its commercial measurement pipeline; undisclosed sample sizes limit independent assessment.
Our open question at DijitalPi: where does the recorded visit source differ from the customer’s account of discovery? Define your measurement coverage before applying consumer-brand results from English-speaking markets to Turkish local services or B2B sales forecasts.
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
Ecommerce: searching after a recommendation
An illustrative sports brand dismisses AI as a channel because direct assistant referrals are scarce.
Could a Google visitor have seen an earlier recommendation?
Open the recommendation for example 01+
Separate discovery from the final traffic source.
In consent-based customer research, ask an open question about where the person first encountered the brand. Review answers alongside branded search and known referrals. Recollection can also be incomplete; do not reclassify all organic traffic as AI contribution.
EXAMPLE 02
Local services: a mention is not an endorsement
An illustrative renovation business reports every occurrence of its name as a positive recommendation.
Is the assistant recommending the business or merely naming it?
Open the recommendation for example 02+
Record the meaning of the surrounding sentence.
Classify sample answers as recommendations, neutral mentions or cautions. Have a person check service coverage and fit. Measure subsequent enquiries separately. Research on consumer brands does not provide a ready-made effect estimate for local services.
EXAMPLE 03
B2B: an existing user counted as a new prospect
An illustrative software brand counts an existing user’s support conversation as new customer acquisition.
Did the mention introduce interest or accompany existing use?
Open the recommendation for example 03+
Distinguish established relationships from new enquiries.
Separate support, renewal and new-product evaluation in your own permissioned records. A lack of recent recorded activity does not prove someone has never been a customer. Keep meetings, quotations and signed contracts as distinct outcomes.
TRY IT WITH YOUR TEAM
Make gaps in the discovery journey visible.
Record known traffic source, self-reported first exposure and subsequent enquiry in separate fields. Leave unobserved steps unknown. This illustrative proposal is neither evidence of sales uplift nor a completed DijitalPi experiment.
The academic work belongs to the researchers named above. This page contains DijitalPi's explanatory review and original business examples; it is not a full translation of the paper.
Cookie preferences