Google’s AI Skips Local Car Repair Searches. It Owns the Questions That Come First.

The AI answer layer has not arrived for local repair

Nine of the twelve metros we tested returned zero AI Overviews across all six service searches. Google answered with a map pack instead. Broken out by what the driver typed:

Only Tampa, Nashville and Salt Lake City produced an AI answer at all, and only on the vaguer, more diagnostic phrasings. “Brake repair” and “auto ac repair” — the two most transactional searches in the set — returned nothing in any of the twelve metros.

The questions that come first

Ask Google something a driver actually wonders about — whether the dealer is a rip-off, how to spot an honest shop, what a rebuild should cost — and the AI answers almost every time. Twenty-seven of our thirty advice questions produced an AI Overview. Only three did not: where should I get my transmission rebuilt, how often should I change my transmission fluid, and best place to get brakes done.

These answers do real work. They set the price the driver expects, decide whether a dealership is even in the running, and hand over a checklist of what a trustworthy shop looks like — all before a single shop name is searched for.

There is no gatekeeper here

Across 32 answers Google cited 156 distinct domains, and 125 of them appeared exactly once. Nothing dominates. Sorted by what kind of business owns the site:

Reddit appeared in 62.5% of answers, YouTube and Quora in 50% each. But the striking column is the right-hand one: seventy different independent repair shops were quoted as sources, against twenty dealership websites. A single-location shop — Rohnert Park Transmission in California — was cited in three separate answers, more often than J.D. Power.

Put beside our sports nutrition study from the same week, the contrast is close to total. In supplements, three publishers supplied the sources for 77% of answers and brand-owned websites accounted for 5.1% of citations — you cannot write your way in, because Forbes and Healthline firewall editorial from commerce. In auto repair, the answer is substantially assembled from ordinary businesses’ own content.

A finding we expected to go the other way

We went in testing whether AI answers steer drivers toward dealership service departments — which, had it been true, would have made AI discovery a new front in the Right to Repair argument. It is not true, but the picture is more interesting than a simple win for either side. Of the 27 advice answers, 13 directly compared dealerships with independents, and they split by dimension rather than picking a winner.

On cost, independents take it outright. All seven answers that addressed price came down the same way. Google states it without hedging — “Yes, dealership service is almost always more expensive than an independent shop” — and puts numbers on it: independent shops at “$110–$170 per hour” against dealership service centers at “$130–$250+ per hour, typically $20 to $40 higher than local independent shops due to brand overhead and specialized equipment.”

On competence, dealers take it. Nine of the thirteen credit dealerships for factory training, proprietary diagnostic software, warranty work or recalls. Asked directly whether dealership mechanics are better trained, the answer opens “Yes” — qualified immediately by the note that independents “often have a broader overall experience across many different makes and models,” but a yes nonetheless.

The single most consequential thing Google tells drivers is legal rather than commercial: under “the federal Magnuson-Moss Warranty Act, car makers cannot force you to use their dealerships for routine maintenance or repairs, nor can they void your entire warranty just because an independent shop touched your car.” That is the independent sector’s central argument, delivered unprompted to anyone who asks. We are reporting all of this because we set out to test it, not because it flatters anyone.

What this means for a repair shop

The practical read is not “ignore AI.” It is that the leverage sits somewhere other than where it is being sold. Nobody needs to win an AI Overview for “brake repair Denver” — Google is not writing one. What is winnable is the layer above it, where the questions are general, the answers are constant, and the sources are shops exactly like yours that happened to publish a clear answer.

That is an unusually open door, and it is open precisely because there is no gatekeeper to negotiate with. It is also unlikely to stay this way: an answer surface sourced from 156 domains, 125 of them cited once, is an immature one. Consolidation is the normal direction of travel.

One defect worth naming, because it shows how loose the sourcing currently is: five of the cited domains were not American — UK and Australian garages and motoring bodies, including rac.co.uk and johnhughes.com.au — supplying answers to US-located queries about US pricing.

Methodology

Data collected 5 August 2026, United States, English, desktop. 102 queries submitted to Google through the DataForSEO live SERP endpoint with AI Overview loading enabled: 72 local (12 metros — Phoenix, Charlotte, Denver, Tampa, Columbus, Nashville, Portland, San Antonio, Kansas City, Salt Lake City, Raleigh, Las Vegas — against six service templates) and 30 national advice questions. Every query that errored was retried until it returned; there are no unrecovered failures in the sample. Baltimore was excluded deliberately, because SEMPITE has clients there.

For each answer we recorded the AI Overview body text and every domain in its reference block. Each cited domain was then classified by the type of business that owns it, using each domain’s own homepage title and meta description rather than the title of the page Google cited — a dealership’s cited page is frequently a service page, which makes cited-page text a misleading signal. Classification is rule-based and deterministic, so it can be reproduced exactly: known platform, publisher and chain lists, then a test of whether a homepage advertises selling cars versus only servicing them, with a manufacturer name in the domain breaking ties. It scores 94% against a hand-labeled validation set, and the pipeline refuses to emit results below 85%. Domains the rules could not decide are reported as unclassified rather than guessed.

Limits, stated plainly. This is one snapshot; AI Overviews are non-deterministic and vary by user, location and time, so individual queries will not always reproduce. It measures Google’s AI Overview only — not Google AI Mode, ChatGPT, Perplexity or Gemini, which are different surfaces with different sourcing. Absence of an AI Overview is not absence of AI: it means this specific surface did not render for that query. The local sample is twelve metros, not a census. And the 7.8% of citations we could not classify would, in the worst case, shift the source mix by that much.

The weakness we would attack first, if this were someone else’s study. Local queries were submitted at national scope with the city named in the query text — “brake repair Phoenix AZ” — rather than from a request geolocated to Phoenix. Those are not identical, and Google’s local surfaces are geo-sensitive, so a genuinely geolocated request could plausibly behave differently. Results were also collected on desktop, while local service intent skews mobile. Both caveats bear on the 6.9% local figure specifically. They do not bear on the citation-structure finding, which is the study’s main result: 225 of the 256 citation slots come from the national advice questions, where geolocation is not a factor.

Full query list, per-answer citation records and the complete domain classification with its reasoning are available on request under CC-BY.