A customer can read an answer on Google and leave without visiting your website. Does AI search actually reduce clicks to external sites? In a preregistered US field experiment, participants assigned to use Google AI Mode for their searches clicked out less and rated their search experience less favourably. DijitalPi commentary: Track appearing in an answer, receiving a visit and acquiring a customer as separate outcomes. The findings describe Chrome and Google use in March 2026; they do not measure Turkish-language search, the Turkish market or long-term voluntary adoption of AI Mode.
ACADEMIC WORK REVIEWED
AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence
English review: 29 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?
Do people leave for external sites when search supplies the answer?
The researchers tested how Google AI Overviews and conversational AI Mode change publisher clicks, search behaviour and user experience. Preregistration means the hypotheses and analysis plan were set before examining the results.
Explaining the context · DijitalPi commentary
DijitalPi commentary: Answering an informational question on the search page does not guarantee a visit to the business that produced the content. Fewer clicks also need not mean a better experience. Measuring these outcomes separately makes the paper useful for thinking about visibility and traffic reporting.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
A randomized field experiment recruited US adults whose primary browser was Chrome and primary search engine was Google in March 2026. The behavioural analysis included 1,100 participants. After 3 baseline days of normal search, participants spent 7 days with normal search, hidden AI features or all searches redirected to AI Mode.
Understanding the method · DijitalPi commentary
The normal-search comparison was not entirely AI-free: AI Overviews could appear and AI Mode was available by choice. A Google interface change disrupted the hiding intervention; only 51.1% of AI Overviews were successfully hidden. The authors therefore report the effect of assignment for AI Mode and a local effect accounting for actual implementation for the hiding condition. These are different estimators, not two simple differences between group averages.
Click-through rate is external clicks originating from Google divided by Google searches, calculated per user. Google-internal clicks are excluded from the numerator and vertical searches such as images, news and shopping from the denominator. Each AI Mode conversation turn counts as a separate search. This is not a percentage change in visitor numbers or sales.
03 / RESEARCH FINDINGS
Assignment to AI Mode reduced external click-through.
Relative to normal Google search, assignment to AI Mode reduced user-level external click-through rate by 18.8 percentage points. The local estimate for hiding AI features, accounting for actual implementation, was an increase of 8.8 percentage points against the same comparison. These are adjusted effect estimates, not relative percentages or raw before-and-after rates. Trust, satisfaction and perceived usefulness also declined under AI Mode; hiding AI features produced no significant differences in those experience measures.
DijitalPi commentary: Keep separate records for appearing as a source and receiving a website visit. Whether fewer visits mean fewer qualified enquiries requires additional measurement. The study did not measure sales or revenue losses, so its click-rate estimate cannot serve as a business revenue forecast.
The hiding estimate assumes assignment affected outcomes only through AI Overview exposure. If the interface change affected the experience in other ways, that interpretation weakens. Compulsory routing of all searches to AI Mode also differs from choosing the mode when it feels useful.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
The preprint provides experimental evidence of fewer external clicks under the studied short-term search arrangement. The younger and more educated sample, compulsory use and short follow-up constrain generalization. Turkish-language search, Türkiye and long-term adaptation need separate research. The authors declare no competing interests; the work received partial US National Science Foundation support.
Our open question at DijitalPi: can your website distinguish informational visitors from those progressing towards a product or service enquiry? The illustrative applications below address that distinction; they do not claim the same effect has been measured in Türkiye.
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
Ecommerce: an answer without a guide visit
An illustrative retailer’s buying guide appears as a source while traffic to the guide weakens.
Did the citation bring a customer to the store?
Open the recommendation for example 01+
Separate source visibility from website behaviour.
Record date, language and search mode for the same product questions. Track guide visits, product-page visits and orders separately. Note stock, campaign and seasonal changes. These are proposed DijitalPi measurement steps; they cannot alone establish why traffic changed.
EXAMPLE 02
Local services: an answer is not a booking
An illustrative repair business suspects that customers find maintenance answers on the search page.
Is seeking information the same as requesting a service?
Open the recommendation for example 02+
Separate informational questions from service requests.
Track maintenance questions separately from service-area and appointment questions. Make hours and coverage clear on the website. Record calls and form enquiries independently; do not assume every contact followed an AI answer.
EXAMPLE 03
B2B: fewer guide visits, a changing offer
An illustrative software company sees fewer guide visits while also changing its product offer.
Does less traffic necessarily mean fewer qualified enquiries?
Open the recommendation for example 03+
Check lead quality and concurrent changes.
Examine guide, pricing and enquiry pages separately. Record business needs and progression to a meeting. Include the offer change in the report. Do not translate the paper’s click-rate estimate into a predicted revenue loss for the company.
TRY IT WITH YOUR TEAM
Track the steps from visibility to enquiry separately.
For important customer questions, identify the data source for citations, website visits, qualified enquiries and sales. Keep visits with unknown origins separate. This is a proposed DijitalPi measurement exercise, not a completed client experiment.
Stephanie T. Wang, Jeffrey Gleason, Yakov Bart, Christo Wilson, Danaé Metaxa. Computer and Information Science, University of Pennsylvania; Khoury College of Computer Sciences, Northeastern University; D’Amore-McKim School of Business, Northeastern University.
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