Researchers Redirected Every Search to Google's AI Mode for a Week. Clicks to Websites Fell 18.8 Points and the Users Tried to Leave.

For one week in March, a group of American Google users had nearly every search they typed redirected into AI Mode. Not offered. Redirected. A browser extension caught the query and sent it to Google's conversational search experience whether the person wanted it or not.

Clicks out to websites fell by 18.8 percentage points.

That number comes from a preregistered field experiment run by researchers at the University of Pennsylvania and Northeastern University, posted to arXiv on August 18 and written up by Search Engine Journal yesterday. The study enrolled 1,100 US participants who used Chrome and Google as their defaults, ran them through three baseline days of ordinary search, then randomly assigned them to one of three worlds: AI Overviews and AI Mode hidden, Google exactly as it ships today, or every search forced into AI Mode. It is a causal read on AI Mode, a surface the authors note is still understudied, and the numbers in it go well past the one everybody has been arguing about.

The click-through drop was the headline. The rest of the study is what a dealer should actually read.

Google's own prediction is the one that failed

Here is the part that makes this more than another traffic-panic study.

The researchers wrote seven hypotheses before they collected a single data point. One of them, H3, predicted that forcing people into AI Mode would increase the number of times they searched per day. They did not invent that guess. They took it from Google, which has said publicly that its AI features drive more searching. In August 2025, Liz Reid, VP and Head of Google Search, wrote that "total organic click volume from Google Search to websites has been relatively stable year-over-year" and that "average click quality has increased and we're actually sending slightly more quality clicks to websites than a year ago."

The experiment went the other way. Assignment to AI Mode cut search sessions by 0.92 per day. Among heavy Google users, the people running more than four search sessions a day, the drop was 2.01 sessions per day.

Read that twice. The heaviest searchers, the ones who live in Google, searched the least once AI Mode was all they had.

Every other perception measure moved the same direction. Trust in information found on Google fell. Perceived usefulness fell. Satisfaction fell. Agency fell. Personalization and relevance fell, even though the researchers had predicted those would rise, again following Google's own claims about AI-driven personalization. And 11.2 percentage points more participants went and searched on Bing, DuckDuckGo or Yahoo instead.

The resistance is the real finding here, more than the loss.

The sentence every dealership should tape to a monitor

At the end of the week the researchers asked the AI Mode group an open question: what did you think? They coded 309 answers. Negative sentiment led at 33.6 percent, positive at 29.3 percent.

One complaint category should stop a dealer cold. In 15.3 percent of responses, people said they had trouble getting to a specific website they were trying to reach. One participant put it exactly:

"when I just wanted to land on the webpage I would type in, it would tell me about the website instead."

That is a navigational search. Somebody typed a business name into Google because they wanted that business's website, and the AI answered a question they had not asked.

Now think about what share of your dealership's organic traffic is navigational. Somebody types your store name plus "hours." Somebody types your store name plus "service." Somebody types your store name and a model because they saw the unit on a third-party site and now want to see it on yours. That traffic is your highest-intent, lowest-cost, closest-to-the-sale traffic, and it is the traffic a conversational answer layer is best positioned to absorb, because the answer is short and Google already has it from your Business Profile.

Another 13.4 percent of responses complained about limited links and source diversity. Fewer links on the page means fewer slots, and fewer slots means your store is competing with the OEM, Cars.com, CarGurus, Reddit and the local news site for a list that got shorter.

The number nobody is quoting

Buried in the secondary results is the figure with the biggest dollar sign attached for a franchise store.

The fraction of users clicking on ads dropped by 42.7 percentage points in the AI Mode condition. The reason is simple and the authors say so plainly: AI Mode did not surface ads at the time of the experiment. Force every search into a surface with no ad inventory and ad clicks go to nearly nothing.

Do not read that as a forecast for your paid budget. It is a measurement of what AI Mode was in March, before monetization. The authors flag this themselves, writing that their estimates reflect "an early snapshot" and that they expect user satisfaction to shift away from this pre-monetization baseline once Google starts selling the surface.

Google is already doing exactly that. At Google Marketing Live in May, the company said it is testing two new ad types built with Gemini, Conversational Discovery ads and Highlighted Answers, alongside AI-powered Shopping ads and a Business Agent for Leads, all labeled "Sponsored."

So the honest version for a dealer is this. The surface that measured the worst for trust, satisfaction and agency is the surface Google is now building your next generation of paid inventory inside. The 42.7 point figure is the size of the hole that ad inventory is being poured into, and the perception numbers are the mood of the audience it will be poured on.

What this means for your dealership

Your organic reporting will get worse before it gets clearer. Google finished rolling out its AI performance report to every Search Console property on August 31, and it counts impressions and withholds clicks. You cannot separate AI Mode from ordinary search inside it. That means the 18.8 point effect, if it reaches your store, will show up in your GA4 as a soft organic decline with no named cause.

The soft decline may be understated. First-party data published in August found 22.4 percent of AI Overview traffic landing in GA4 as Direct rather than Organic Search. Traffic that does arrive gets filed in the wrong bucket, so the channel looks worse than it performs and Direct looks better than it earned.

This is a second study contradicting the "low quality clicks" defense. In July we covered a randomized study showing the clicks AI Overviews absorb are as engaged as the ones that survive. This one adds the user side of it. The people losing those clicks were less satisfied, not more.

Brand and navigational queries are where the exposure sits. A conversational layer is very good at answering "what are the hours at Smith Ford." It is much weaker at handing over a specific stock number on a specific unit at a specific price. Your VDP survives on specificity. Your hours page does not.

Do not assume everyone stays on Google. An 11.2 point shift toward Bing, DuckDuckGo and Yahoo in a single week is small in absolute terms and large as a direction. Those engines feed Copilot and other AI surfaces you are probably not measuring at all.

What to do about it

1. Pull a branded query report in Search Console for the last 12 months and chart branded impressions against actual branded sessions in GA4. A widening gap between the two is interception, and it is the earliest signal you will get.
2. Put the navigational answers where a machine can lift them cleanly. Hours, address, phone, department hours, and current stock on every relevant page, in text, matching your Google Business Profile exactly. If the AI is going to answer instead of refer, make sure it answers correctly.
3. Separate branded from non-branded in every organic report you look at. They are about to behave differently, and a blended organic number will hide it.
4. Verify your site in Bing Webmaster Tools this week if it is not already. It costs nothing and it is the only view you get of the engines people defect to.
5. Ask your paid agency which surfaces your Search and Shopping ads are eligible to serve on right now, and how they will report AI Mode placements separately when those formats reach general availability. Get the answer in writing before the inventory exists, not after.
6. Stop treating your AI visibility and your website traffic as one number. They are separating. Track appearing in the answer and earning the click as two different outcomes with two different fixes.

The comforting read on this study is that users pushed back, and Google has a commercial reason to listen. The uncomfortable read is that Google is going to keep pushing, monetize the surface, and the pre-monetization version people disliked will be replaced by one with ads in it.

Either way, the dealership that knows its own branded-versus-unbranded split, its own Direct misattribution rate, and its own AI impression trend will spend the next year making decisions while everyone else argues about someone else's data.

Want to see how your site currently reads to the AI layer before any of this lands? Take a look at our AI compatibility work, or book a walkthrough and we will go through your numbers with you.