Google Just Published What Americans Actually Ask AI. Two of the Ten Biggest Gaps Are Comparison Shopping and Car Repair.

On July 23, Google published a research report almost nobody in the car business is going to read. It contains the most useful data anyone has put out this year about where AI actually intersects with a dealership, and it does not mention automotive once.

The report is called the AI & Economy ATLAS (Activity, Task, Landscape and Adoption Study), written by Google's Zanna Iscenko and Scott Strand. It's the company's first serious attempt to answer a question everyone in marketing has been guessing at for two years. Not how many people use AI. What they use it for.

The full report analyzes 14,653,926 de-identified conversations across the Gemini app, AI Mode in Search, and the Gemini API, collected between April 6 and April 19 of this year. For scale: as of May, AI Mode serves over a billion monthly users and the Gemini app another 900 million.

Then Google did the genuinely clever part. Rather than invent its own categories, it mapped every one of those conversations onto the Bureau of Labor Statistics' American Time Use Survey, the long-running government study where a representative random sample of Americans keeps a chronological diary of what they actually did for a full 24 hours. That let Google line up two numbers side by side for every activity in American life: the share of AI conversations about it, and the share of waking hours actually spent on it.

The gap between those two numbers is the whole story. And in the table of the ten daily activities most over-represented in AI conversations relative to time spent, two entries are the front and the back of your building.

"Comparison shopping." And "Vehicle repair and maintenance (by self)."

What the report actually found

Start with the headline number, because it reframes everything: 86.5% of the AI conversations in this dataset happen outside of work. Formal work tasks account for just 13.5%. Whatever you assumed AI was mostly for (coding, spreadsheets, corporate email), the data says it's overwhelmingly a household technology. Google notes this isn't a Gemini quirk either; it points to OpenAI's own published data showing roughly 69% of ChatGPT interactions occur in non-work, consumer contexts.

Globally, conversations cluster in leisure (26.4%) and education (20.7%), followed by household activities (11%), personal care (9%), consumer purchases (5.2%) and professional and personal care services (4.8%).

But the raw shares aren't the interesting part. The interesting part is what happens when you compare US conversation share against US time share. There, some categories tower over the time people actually spend on them:

• Government services and civic obligations: roughly 20x over-indexed. The single biggest outlier in the entire report.
• Professional and personal care services: more than 7x.
• Education: about 5.8x.
• Consumer purchases: nearly 3x.

And the mirror image, the things Americans spend enormous time on and almost never ask AI about:

• Eating and drinking: about 18x under-indexed. The starkest negative gap in the data.
• Travel, sports and recreation, and caring for household members all skew negative too.

Google's own summary of the pattern is that AI assistance "over-indexes in cognitive and bureaucratic tasks, and under-indexes in physical and location-bound activities."

That's accurate but too polite. Here's the blunter version.

AI usage isn't a picture of life. It's a map of friction.

Nobody asks an AI how to eat dinner. You do it every day, you're good at it, the stakes are zero, and if you get it wrong you try again tomorrow.

People ask AI about the DMV. Twenty times more than the time they spend there. They ask about financial services, health decisions, legal questions, education. What every one of those has in common is that it's infrequent, expensive, hard to reverse, procedurally confusing, and, critically, the other party knows more than you do.

That is not a description of daily life. That is a description of the handful of moments each year where a normal person feels out of their depth and doesn't want to look stupid asking.

Now read that list again and notice that you sell two of them.

Buying a vehicle is a large purchase a household makes rarely, with a price structure almost nobody fully understands, pitting a buyer who does this once in a great while against a business that does it all day. Authorizing a repair is the same asymmetry in miniature and far more often: a number you can't evaluate, for a problem you can't see, quoted by the only person in the room who can diagnose it.

You are not accidentally caught in the blast radius of AI adoption. By Google's own data, you are standing near the center of it.

We've seen this conversation before. We just never got to watch it.

Here's the part worth sitting with: those questions were never being asked on your website anyway.

"Is this a fair price?" "Do I actually need this or are they upselling me?" "Should I fix it or trade it?" "Is the extended warranty worth it?" For a century those went to the brother-in-law who knew cars, the guy at work with the tools, a dog-eared Consumer Reports, and later a forum thread. Dealers never saw that conversation, never influenced it, and mostly never thought about it. It was the invisible layer where the shortlist got made and the trust got decided, and by the time a customer reached you, it was already over.

What changed isn't that the conversation moved online. It moved online a long time ago. What changed is that the brother-in-law now has a billion monthly users, reads whatever is published on the open web, and answers by synthesizing whichever sources it found, which means for the first time in the history of the business, that private conversation has an input you can actually supply.

And Google just published a map of what he's being asked.

What this means for your dealership

The service line is the one to sit with. "Vehicle repair and maintenance (by self)" being a top-ten over-indexed activity is not primarily a story about losing work to DIY. Watch the actual shape of that conversation: it starts at "can I do this myself," and it ends in one of two places: a YouTube video, or "okay, what should this cost at a shop." That second branch is a service customer with their wallet out, forming a price expectation from whatever the AI could find. On most dealership websites, what it can find about your service pricing is nothing at all. So the number gets built from a national average, a parts-store blog, and a forum post from 2019, and then your advisor has to argue with it.

"Comparison shopping" over-indexing means the shortlist forms where your analytics can't reach. This is the quiet consequence: a shopper can do the entire consideration phase (three brands, trim levels, financing structures, which local stores seem trustworthy) without generating a single session in your GA4. By the time you get a measurable visit, you're not competing for consideration. You're confirming a decision that was already made somewhere you couldn't see. We've written before about what happens to the top of the funnel when the shortlist forms off-site; this is Google's own data describing the same thing from the other direction.

The 20x government-services number touches you more than you'd think. Title and registration, temp tags, out-of-state purchases, doc fees, what a rebate actually requires, how a tax credit works. Dealers handle this paperwork every day and treat it as back-office; consumers treat it as the most confusing part of the transaction and are clearly asking machines about it. Fair warning on this one, though: Google itself flags in a footnote that the sample was drawn in early April, so tax-season timing may be inflating the category.

What this report does not say

The corpus of vendor emails you'll get about this in the next three weeks will overstate it, so let's be precise about the limits, several of which Google states plainly:

• AI Overviews is not in this data. Neither is Google Workspace, Maps, Translate, or Gemini Notebook. ATLAS covers the Gemini app, AI Mode, and the API, full stop. The surface most dealers actually worry about, the summary sitting on top of the search results, is explicitly excluded, and Google elsewhere describes it as used by over 2.5 billion people. This is a partial view.
• It measures conversations, not outcomes. In Google's words, ATLAS "measures behavioral interactions, not definitive productivity outcomes," and "does not capture the ultimate productive output the user is working toward." There is no click data in it. It cannot tell you whether a single one of those comparison-shopping conversations sent anyone to a dealership website, as Search Engine Journal noted in its coverage.
• The granular findings are probabilistic. Google says so directly: classifications are inferred from conversational text, and highly granular occupational or household activity findings "carry more uncertainty than broader, major-group trends." Treat that top-ten table as directional, not as a measurement.
• It's a two-week sample. April 6-19, 2026. One snapshot, and one that lands in tax season.

Which means: nobody can sell you an "ATLAS-optimized" anything, and if a vendor tries, that tells you what you need to know about the vendor.

What to do about it

1. Publish your service pricing in plain language, per store. Ranges are fine: the job, a real price range, and how long it takes, for the vehicles you actually see. "Front brake pads and rotors on a 2021 Explorer at our shop" is a page almost no dealership has. This is the single highest-return content on a dealership site right now, it costs one service advisor an afternoon, and virtually no dealer has it, which is exactly why an AI has to build the answer out of somebody else's numbers.
2. Write the honest "can I do this myself" pages. Which jobs a reasonably handy owner really can do, which ones need a lift or a scan tool, and where the warranty line is. It feels like arguing against your own revenue. It isn't. These pages put you in the conversation that decides whether they call you at all, and they're the kind of content these systems reward and cite.
3. Answer the paperwork questions. Title, registration, temp tags, out-of-state buys, what documents to bring, how your doc fee works and what it covers. Describe your own process rather than giving generic tax or legal guidance. That's both more useful and more citable than the syndicated OEM blog post sitting on 400 other dealer sites.
4. Build real comparison content, including the vehicles you don't sell. "Comparison shopping" is over-indexed for a reason. A page that honestly compares your model against the two cross-shops and says who each one is right for will get used. A page that pretends the alternatives don't exist gets skipped by both shoppers and machines.
5. Stop grading these pages on sessions. If the consideration phase is happening inside a conversation you can't measure, pageviews will underrate this content badly. Watch branded search, direct traffic, and what customers say on the phone. "I read that your brake job runs about..." is the conversion event here, and it lands in a phone call, not an analytics dashboard. The service drive was already the most undermarketed asset in the building; this is the strongest argument yet for fixing that.
6. Check what an AI says about your store today. Ask it what a repair costs at your dealership, what your fees are, why someone should buy from you. Whatever it answers is what your next customer is reading, and if the answer is vague or wrong, that's not a marketing problem, it's an inventory-of-facts problem you can fix this week.

The bottom line

Google set out to measure AI's effect on the economy and accidentally published a customer-research document for the car business.

The finding underneath all the charts is simple: people bring AI the decisions that are rare, costly, and lopsided: the ones where they suspect the person across the desk knows something they don't. Buying a car is that. Approving a repair is that. Those conversations have always happened, always mattered, and were always invisible to you.

They're still invisible. But now they're being answered from published sources, which means for the first time you get a vote, and the dealers who take it will be the ones who wrote the boring, specific, honest pages nobody thought were marketing.

If you want to see what an AI currently says about your store's pricing, your fees, and your service department (and whether the systems now shaping those answers can even read your site), run your dealership through our AI compatibility test, and we'll walk through the real answers with you.