Our new usability research of 37 contributors throughout seven particular search duties clearly exhibits that individuals:
- Learn AI Mode
- Not often click on out, and
- Solely go away when they’re able to transact.
From what we all know, there isn’t one other impartial usability research that has explored AI Mode to this depth.
In Might, I printed an in depth two-part research of AI Overviews (AIOs) with Amanda, Eric Van Buskirk, and his group. Eric and I additionally collaborated on Propellic’s journey trade AI mode research.
We labored collectively once more to convey you this week’s Development Memo: a research that gives essential insights and validation into the behaviors of individuals as they work together with Google’s AI Mode.
Since neither Google nor OpenAI (or anybody else) offers consumer knowledge for his or her AI (Search) merchandise, we’re filling an important hole.
We captured display recordings and think-aloud classes by way of distant research. The 250 distinctive duties collected present a sturdy knowledge set for our evaluation. (The whole methodology is offered on the finish of this memo, together with particulars in regards to the seven search duties.)
And also you may be shocked by among the findings. We had been.
It is a longer publish, so seize a drink and settle in.
Picture Credit score: Kevin Indig
Government Abstract
Our new usability research of Google’s AI Mode reveals how profoundly this characteristic adjustments consumer conduct.
- AI Mode holds consideration and retains customers inside. In roughly three‑quarters of the whole consumer classes, customers by no means left the AI Mode pane – and 88 % of customers’ first interactions had been with the AI‑generated textual content. Engagement was excessive: The median time by job kind was roughly 52-77 seconds.
- Clicks are uncommon and largely transactional. The median variety of exterior clicks per job was zero. Yep. You learn that proper. Ze-ro. And 77.6% of classes had zero exterior visits.
- Individuals skim however nonetheless make choices in AI Mode. Over half of the duties had been labeled as “skimmed rapidly,” the place customers look on the AI‑generated abstract, kind an opinion, and transfer on.
- AI Mode delivers “web site sorts” that match intent. It’s not nearly assembly search question or immediate intents; AI Mode is citing sources that match particular web site classes (like marketplaces vs evaluation websites vs manufacturers).
- Visibility, not site visitors, is the rising foreign money. Contributors made their model judgments straight from AI Mode outputs.
TL;DR? These are the core findings from this research:
- AI Mode is sticky.
- Clicks are reserved for transactions.
- AI Mode matches web site kind with intent.
- Product previews act like mini product element pages (aka PDPs).
However earlier than we dig in, a fast shout-out right here to the group behind this research.
Along with Eric Van Buskirk’s group at Clickstream Options, I carried out the primary broad usability research of Google’s AI Mode that uncovers not solely essential insights into how folks work together with the hybrid search/AI chat engine, but additionally what sorts of branded websites AI Mode surfaces and when.
I need to spotlight that Eric Van Buskirk was the analysis director. Whereas we collaborated intently on shaping the analysis questions, areas of focus, and methodology, Eric managed the group, oversaw the research execution, and delivered the findings. Afterward, we labored facet by facet to interpret the info.
Click on knowledge is a superb first go for evaluation on what’s occurring in AI Mode, however with this usability research particularly, we primarily regarded “over the shoulder” of real-life customers as they accomplished duties, which resulted in a sturdy assortment of information to tug insights from.
Our testing platform was UXtweak.
Boost your skills with Growth Memo’s weekly expert insights. Subscribe for free!
Google’s own Sundar Pichai has been crystal clear: AI Mode isn’t a toy; it’s a proving ground for what the core search experience will look like in the future.
On the Lex Fridman podcast, Pichai said (bolding mine):
“Our current plan is AI Mode is going to be there as a separate tab for people who really want to experience that… But as features work, we’ll keep migrating it to the main page…” [1]
Google has argued these new AI-focused options are designed to level customers to the online, however in observe, our knowledge exhibits that customers stick round and make choices with out clicking out. In principle, this might not solely influence click-outs to natural outcomes and citations, but additionally scale back exterior clicks to advertisements.
In August, I explored the truth behind Google’s personal product cannibalization with AI Mode and AIOs:
Proper now, in response to Similarweb knowledge, utilization of the AI Mode tab on Google.com within the US has barely dipped and now sits at simply over 1%.
Google AIOs at the moment are seen by greater than 1.5 billion searchers each month, and so they sit entrance and middle. However engagement is falling. Customers are spending much less time on Google and clicking much less pages.
However as Google rolls AI Mode out extra broadly, it brings the most important shift to Search (the most important buyer acquisition channel there may be) ever.
Conventional search engine optimization is extremely efficient within the new AI world, but when AI Mode actually turns into the default, there’s a probability we have to rethink our arsenal of ways.
Making ready for the way forward for search means treating AI Mode because the vacation spot (not the doorway), and determining easy methods to present up there in ways in which really matter to actual consumer conduct.
With this research, I sought out to find and validate precise consumer behaviors inside the AI Mode expertise when enterprise quite a lot of duties with differing search intents.
1. AI Mode Is Sticky
Picture Credit score: Kevin Indig
Key Stats
Individuals learn first and often keep contained in the AI Mode expertise. Right here’s what we discovered:
- The vast majority of classes had zero exterior visits: that means, they didn’t go away AI Mode (in any respect).
- ~88% of customers’ first interplay* inside the characteristic was with the AI Mode textual content.
- Typical consumer engagement inside AI Mode is roughly 50 to 80 seconds per job.
These three stats outline the AI Mode search floor: It holds consideration and resolves many duties with out sending site visitors.
*Right here’s what I imply by “interplay:”
- An “interplay” inside the consumer duties = the participant meaningfully engaged with AI Mode after it loaded.
- What counts as an interplay: Studying or scrolling the AI Mode physique for greater than a fast look, together with scanning a end result block just like the Buying Pack or Proper Pane, opening a service provider card, clicking an inline hyperlink, hyperlink icon, or picture pack.
- What doesn’t depend as an interplay: Temporary eye flicks, cursor passes, or hesitation earlier than partaking.
Customers are in AI Mode to learn – not essentially to browse or search – with ~88% of classes interacting with the output’s textual content first and spending one minute or extra inside the AI Mode expertise.
Plus, it’s fascinating to see that customers spend greater than double the time in AI Mode in comparison with AIOs.
The general engagement is far stronger.
Picture Credit score: Kevin Indig
Why It Issues
Deal with the AI Mode panel like the first studying floor, not a teaser for blue hyperlinks.
AI Mode is a contained expertise the place sending clicks to web sites is a low precedence and giving customers one of the best reply is the best one.
In consequence, it utterly adjustments the worth chain for content material creators, firms, and publishers.
Perception
Why do different sources and/or AI Mode analysis analyses say that customers don’t return to the AI Mode characteristic fairly often?
My principle right here is that, as a result of AI mode is a separate search expertise (a minimum of, for now), it’s not as seen as AIOs.
As AI Mode adoption will increase with Google bringing Gemini (and AI Mode) into the browser, I count on our research findings to scale.
2. Clicks Are Reserved For Transactions
Whereas clicks are scarce, buy intent isn’t.
Contributors within the research solely clicked out when the duty demanded it (e.g., “put an merchandise in your buying cart”) or in the event that they browsed round a bit.
Nonetheless, the searching clicks had been so few that we will safely assume AI Mode solely results in click-outs when customers need to buy.
Even prompts with a comparability and informational intent are likely to maintain customers contained in the characteristic.
- Buying prompts like [canvas bag] and [tidy desk cables] drive the best AI Mode exit share.
- Comparability prompts like [Oura vs Apple Watch] present the bottom exit share of the duties.
When contributors had been inspired to take motion (“put an merchandise in your buying cart” or “discover a product”), nearly all of clicks went to buying options like Buying Packs or Service provider Playing cards.
Picture Credit score: Kevin Indig
18% of exits had been attributable to customers exiting AI Mode and going straight to a different web site, making it a lot more durable to reverse engineer what drove these visits within the first place.
Research transcripts affirm that contributors typically share out loud that they’ll “go to the vendor’s web page,” or “discover the product on Amazon/ebay” for product searches.
Even when evaluating merchandise, whether or not software program or bodily items, customers barely click on out.
Picture Credit score: Kevin Indig
In plain phrases, AI mode eats up all TOFU and MOFU clicks. Customers uncover merchandise and kind opinions about them in AI Mode.
Key Stats
- Out of 250 legitimate duties, the median variety of exterior clicks was zero!
- The immediate job of [canvas bag] had 44 exterior clicks, and [tidy desk cables] had 31 clicks, accounting for two-thirds of all exterior clicks on this research.
- Comparability duties like [Oura Ring vs Apple Watch] or [Ramp vs Brex] had only a few clicks (≤6 complete throughout all duties).
Right here’s what’s fascinating…
Within the AIOs Overviews usability research, we discovered desktop customers click on out ~10.6% of the time in comparison with virtually 0% in AI Mode.
Nonetheless, AIOs have natural search outcomes and SERP Options under them. (Individuals click on out much less in AIOs, however they click on on natural outcomes and SERP options extra typically.)
Zero-Clicks
- AI Overviews: 93%*
- AI Mode: ~100%
*Remember that contributors of the AIO usability research clicked on common natural search outcomes. The 93% pertains to zero clicks inside the AI Overview.
On desktop, AI Mode produces roughly double the in-panel clickouts in comparison with the AIO panel. On AIO SERPs, complete clickouts can nonetheless occur by way of natural outcomes under the panel, so the page-level fee will sit between the AIO-panel determine and the basic baseline.
An vital be aware right here from Eric Van Kirk, the director of this research: When evaluating the AI Mode and AI Overview research, we’re not precisely evaluating apples to apples. On this research, contributors got duties that will immediate them to depart AI Mode in 2/7 questions, and that accounts for almost all of outbound clicks (which had been fewer than three exterior clicks). Alternatively, for the AIO research, essentially the most transactional query was “Discover a moveable charger for telephones below $15. Search as you sometimes would.” They weren’t instructed to “put it in a buying cart.” Nonetheless, the insights gathered relating to consumer conduct from this AI Mode research – and the sample that customers don’t really feel the necessity to click on out of AI Mode to make extra choices – nonetheless stands as a strong discovering.
The larger image right here is that AIOs are like a reality sheet that steers customers to websites ultimately, however AI Mode is a closed expertise that not often has customers clicking out.
What makes AI Mode (and ChatGPT, by the best way) difficult is when customers abandon the expertise and go on to web sites. It messes with attribution fashions and our capacity to know what influences conversions.
3. AI Mode Matches Web site Kind With Intent
Within the research, we assess what kinds of websites AI Mode exhibits for our seven duties.
The kinds are:
- Manufacturers: Sellers/distributors.
- Marketplaces: amazon.com, ebay.com, walmart.com, homedepot.com, bestbuy.com, goal.com, rei.com.
- Assessment websites: nerdwallet.com, pcmag.com, zdnet.com, nymag.com, usatoday.com, businessinsider.com.
- Publishers: nytimes.com, nbcnews.com, youtube.com, thespruce.com.
- Platform: Google.
Picture Credit score: Kevin Indig
Buying prompts path to product pages:
- Canvas Bag: 93% of exits go to Model + Market.
- Tidy desk cables: 68% go to Model + Market, with a visual Writer slice.
Comparisons path to critiques:
- Ramp vs Brex: 83% Assessment.
- Oura vs Apple Watch: cut up 50% Model and 50% Market.
When the consumer has to carry out a status verify, the result’s cut up model and publishers:
- Liquid Loss of life: 56% Model, 44 % Writer.
Google itself exhibits up on buying duties:
- Retailer lookups to enterprise.google.com seem on Canvas Bag (7%) and Tidy desk cables (11%).
Try the top-clicked domains by job:
- Canvas Bag: llbean.com, ebay.com, rticoutdoors.com, enterprise.google.com.
- Tidy desk cables: walmart.com, amazon.com, homedepot.com.
- Subscription language apps vs free: pcmag.com, nytimes.com, usatoday.com.
- Bottled Water (Liquid Loss of life): reddit.com, liquiddeath.com, youtube.com.
- Ramp vs Brex: nerdwallet.com, kruzeconsulting.com, airwallex.com.
- Oura Ring 3 vs Apple Watch 9: ouraring.com, zdnet.com.
- VR arcade or good house: sandboxvr.com, enterprise.google.com, yodobashi.com.
Firms want to know the enjoying discipline. Whereas basic search engine optimization allowed principally any web site to be seen for any consumer intent, AI Mode has strict guidelines:
- Manufacturers beat marketplaces when customers know what product they need.
- Marketplaces are most well-liked when choices are broad or generic.
- Assessment websites seem for comparisons.
- Opinions spotlight Reddit and publishers.
- Google itself is most seen for native intent, and typically buying.
As SEOs, we have to think about how Google classifies our web site primarily based on its web page templates, status, and consumer engagement. However most significantly, we have to monitor prompts in AI Mode and have a look at the location combine to know the place we will play.
Websites can’t and received’t be seen for all sorts of queries in a subject anymore; you’ll have to filter your technique by the intent that aligns together with your web site kind as a result of AI Mode solely exhibits sure websites (like evaluation websites or manufacturers) for particular kinds of intent.
Product previews present up in about 25% of the AI Mode classes, get ~9 seconds of consideration, and folks often open just one.
Then? 45% cease there. Many opens are fast spec checks, not a clickout.
Picture Credit score: Kevin Indig
You’ll be able to simply see how some product suggestions by AI Mode and on-site experiences are fairly irritating to customers.
The post-click expertise is vital: basic greatest practices like critiques have a big effect on making essentially the most out of the few clicks we nonetheless get.
See this instance:
“It appears to be like prefer it has a whole lot of optimistic critiques. That’s one factor I might have a look at if I used to be going to purchase this bag. So this could be the one I might select.”
In buying duties, we discovered that model websites take nearly all of exits.
Compared duties, we found that evaluation websites dominate. For status checks (like a immediate for [Liquid Death]), exits to manufacturers and publishers had been cut up.
- For transactional intent prompts: Manufacturers take up most exits when the duty is to purchase one merchandise now. [Canvas Bag] exhibits a robust tilt to model PDPs.
- For status intent prompts: Model websites seem alongside publishers. A immediate for [Liquid Death] splits between liquiddeath.com and Reddit/YouTube/Eater.
- For comparability prompts: Manufacturers take a again seat. [Ramp vs Brex] exits go largely to evaluation websites like NerdWallet and Kruze.
Given customers can now straight checkout on ChatGPT and AI Mode, shopping-related duties may ship even fewer clicks out.[2, 3]
Due to this fact, AI Mode turns into a totally closed expertise the place even buying intent is fulfilled proper within the app.
Clicks are scarce. Affect is plentiful.
The info offers us a actuality verify: If customers proceed to undertake the brand new manner of Googling, AI Mode will reshape search conduct in methods SEOs can’t afford to disregard.
- Technique shifts from “get the press” to “earn the quotation.”
- Comparisons are for belief, not site visitors. They scale back exits as a result of customers really feel knowledgeable contained in the panel.
- Retailers ought to optimize for decisive exits. Give costs, availability, and proof above the fold to transform the few exits you do get.
You’ll have to earn citations that reply the duty, then win the few, high-intent exits that stay.
However our research doesn’t finish right here.
Right now’s outcomes reveal core insights into how folks work together with AI Mode. We’ll unpack extra to think about with Half 2 dropping subsequent week.
However for many who like to dig into particulars, the methodology of the research is included under.
Methodology
Research Design And Goal
We carried out a mixed-methods usability research to quantify how Google’s new AI Mode adjustments searcher conduct. Every participant accomplished seven stay Google search prompts by way of the AI Mode characteristic. This design permits us to look at each the mechanics of interplay (scrolls, clicks, dwell, belief) and the qualitative reasoning contributors voiced whereas finishing duties.
The duties:
- What do folks say about Liquid Loss of life, the beverage firm? Do their drinks enchantment to you?
- Think about you’re going to purchase a sleep tracker and the one two accessible are the Oura Ring 3 or the Apple Watch 9. Which might you select, and why?
- You’re getting insights in regards to the perks of a Ramp bank card vs. a Brex Card for small companies. Which one appears higher? What would make a enterprise change from one other card: charge element, eligibility wonderful print, or rewards?
- Within the “Ask something” field in AI Mode, enter “Assist me buy a water-resistant canvas bag.” Choose one that most closely fits your wants and you’ll purchase (for instance, a digicam bag, tote bag, duffel bag, and so on.).
- Proceed to the vendor’s web page. Click on so as to add to the buying cart and full this job with out going additional.
- Examine subscription language apps to free language apps. Would you pay, and in what scenario? Which product would you select?
- Suppose you might be visiting a pal in a big metropolis and need to go to both: 1. A digital actuality arcade OR 2. A wise house showroom. What’s the identify of town you’re visiting?
- 1. Suppose you’re employed at a small desk and your cables are a large number. 2. Within the “Ask something” field in AI Mode, enter: “The system cables are cluttering up my desk house. What can I purchase right now to assist?” 3. Then select the one product you assume could be one of the best answer. Put it within the buying cart on the exterior web site and finish this job.
Thirty-seven English-speaking U.S. adults had been recruited by way of Prolific between Aug. 20 and Sept. 1, 2025 (together with contributors in a small group who did pilot research).*
Eligibility required a ≥ 95% Prolific approval fee, a Chromium-based browser, and a functioning microphone. Contributors visited AI Mode and carried out duties remotely by way of their desktop pc; invalid classes had been excluded for technical failure or non-compliance. The ultimate dataset comprises over 250 legitimate job data throughout 37 contributors.
*Pilot research are carried out first in distant usability testing to determine and repair technical points – like screen-sharing, job setup, or recording issues – earlier than the principle research begins. They assist refine job wording, timing, and directions to make sure contributors interpret them appropriately. Most significantly, pilot classes affirm that the info collected will really reply the analysis questions and that the methodology works easily in a real-world distant setting.
Periods ran in UXtweak’s Distant unmoderated mode. Contributors learn a job immediate, clicked to Google.com/aimode, prompted AI Mode, and spoke their ideas aloud whereas interacting with AI Mode. They got the next instructions: “Suppose aloud and briefly clarify what attracts your consideration as you evaluation the knowledge. Converse aloud and hover your mouse to point the place you discover the knowledge you might be in search of.” Every participant accomplished seven job sorts designed to cowl numerous intent classes, together with comparability, transactional, and informational eventualities.
UXtweak recorded full-screen video, cursor paths, scroll occasions, and audio. Periods averaged 20-25 minutes. Incentives had been aggressive. Uncooked recordings, transcripts, and occasion logs had been exported for coding and evaluation.
Three educated coders reviewed every video in parallel. A row was logged for UI components that held consideration for ~5 seconds or longer. Variables captured included:
- Structural: Fields describing the setup, metadata, or construction of the research – not consumer conduct; embrace knowledge like participant-ID, task-ID, system, question, order of UI components clicked or visited throughout the job, kind of web site clicked (e.g., social, group, model, platform), area identify of the exterior web site visited, and extra.
- Characteristic: Fields describing UI components or interface elements that appeared or had been accessible to the participant. Examples embrace UI factor kind, together with buying carousels, service provider playing cards, proper panel, hyperlink icons, map embed, native pack, GMB card, service provider packs, and service provider playing cards.
- Engagement: Fields that seize lively consumer interplay, consideration, or time funding. Contains studying and a focus, chat and query conduct, together with click on and interplay conduct.
- Consequence: Fields representing consumer outcomes, annotator evaluations, or interpretation of conduct. Annotator feedback, effort score, the place data was discovered.
Coders additionally marked qualitative themes (e.g., “velocity,” “skepticism,” “belief in citations”) to assist RAG-based retrieval. The analysis director spot-checked ~10% of movies to validate consistency.
Annotations had been exported to Python/pandas 2.2. Placeholder codes (‘999=Not Relevant’, ‘998=Not Observable’) had been eliminated, and categorical variables (e.g., appearances, clicks, sentiment) had been normalized. Dwell occasions and different time metrics had been trimmed for excessive outliers. After cleansing, ~250 legitimate task-level rows remained.
Our retrieval-augmented era (RAG) pipeline enabled three levels of research:
- Information readiness (ingestion): We flattened each participant’s seven duties into particular person rows, cleaned coded values, and standardized time, click on, and different metrics. Transcripts had been retained in order that structured knowledge (resembling dwell time) could possibly be related to what customers really mentioned. Aim: create a clear, unified dataset that connects conduct with reasoning.
- Relevance filtering (retrieval): We used structured fields and annotations to isolate patterns, resembling customers who left AI Mode, clicked a service provider card, or confirmed hesitation. We then searched the transcripts for themes resembling belief, comfort, or frustration. Aim: mix conduct and sentiment to disclose actual consumer intent.
- Interpretation (quant + qual synthesis): For every group, we calculated descriptive stats (dwell, clicks, belief) and paired them with transcript proof. That’s how we surfaced insights like: “external-site duties confirmed greater satisfaction however extra CTA confusion.” Aim: hyperlink what folks did with what they felt inside AI Mode.
This pipeline allowed us to question the dataset hyperspecifically – e.g., “all contributors who scrolled >50% in AI Mode however expressed mistrust” – and hyperlink quantitative outcomes with qualitative reasoning.
In plain phrases: We will pull up simply the best group of contributors or moments, like “all of the individuals who didn’t belief AIO” or “everybody who scrolled greater than 50%.”
We summarized consumer conduct utilizing descriptive and inferential statistics throughout 250 legitimate job data. Every metric included the depend, imply, median, normal deviation, normal error, and 95% confidence interval. Categorical outcomes, resembling whether or not contributors left AI Mode or clicked a service provider card, had been reported as proportions.
Analyses lined greater than 50 structured and behavioral fields – from system kind and dwell time to UI interactions, sentiment. Confidence measures had been derived from a JSON evaluation of consumer sentiment by way of transcripts of all customers.
Every job was annotated by a educated coder and spot-checked for consistency throughout annotators. Coder-level distributions had been in comparison with affirm steady labeling patterns and inside consistency.
Thirty-seven contributors accomplished seven duties every, leading to roughly 250 legitimate duties. At that scale, proportions round 50% carry a margin of error of about six proportion factors, giving the dataset sufficient precision to detect significant directional variations.
Pattern measurement is smaller than our AI Overviews research (37 vs. 69 contributors) and is supposed to find out about U.S.-based customers (all contributors had been residing within the U.S.). All queries happened inside AI Mode, that means we didn’t straight examine AI vs non-AI situations. Suppose-aloud might inflate dwell occasions barely. RAG-driven coding is barely as sturdy as its annotation inputs, although heavy spot-checks confirmed reliability.
Contributors gave knowledgeable consent. Recordings had been encrypted and anonymized; no personally figuring out knowledge had been retained. The research conforms to Prolific’s ethics coverage and UXtweak TOS.
Featured Picture: Paulo Bobita/Search Engine Journal
