ChatGPT Answers 'Dealer Near Me' From Yelp and Google Maps. New Data Shows Your Website Barely Enters the Room.

For most of this summer, ChatGPT was telling anyone who cared to look exactly where it got its information.

Every web result in ChatGPT's server stream carried a field called result_source, and that field named the pipeline that fetched it. A French search consultancy called RESONEO noticed, built a capture tool, and ran it until the field disappeared. Their writeup of what they found went up in July and got its first real trade coverage on August 11, when Search Engine Journal picked it up.

The corpus is not small. RESONEO logged 1,249 answers across 682 free-tier and 567 paid conversations, working out to roughly 88,000 search results drawn from 26,900 distinct pages across 6,400 domains, using free, paid and logged-out accounts in several countries. Then, in their words, "On 21 July, OpenAI turned off the light." The field vanished from the stream. Nobody outside OpenAI can run this measurement again.

So this is a photograph, taken through a window that is now closed. Read it that way. It is still the best look anyone has gotten at the machine that increasingly decides whether a shopper hears your dealership's name.

The four pipes

The headline finding is that ChatGPT does not have "a" search engine behind it. It has several, and it picks between them.

The result_source field had four values: labrador, bright, oxylabs and serp. RESONEO reads labrador as OpenAI's own in-house index. The others are commercial scraping and SERP vendors, with bright being Bright Data, effectively Google results collected on OpenAI's behalf.

Which pipe you get depends on who you are and how you asked. In RESONEO's numbers, a free account in instant mode came back 63.6% labrador, 34.7% bright, 1.6% serp. A paid account in thinking mode flipped almost exactly the other way: 74.6% bright, 23.6% labrador, 1.9% serp.

Sit with that for a second, because it breaks something dealers get sold every week. Two shoppers ask ChatGPT the same question about your store. One is on the free tier, one is paying twenty dollars a month. They are querying different systems. Any vendor handing you a single number for "your ChatGPT visibility" is averaging across a split they almost certainly did not control for. We made this argument in July off a statistical paper; this is the plumbing underneath it.

The line that should stop every dealer cold

Now the part that matters at a car store.

RESONEO broke their free-account results down by question type. For the row covering a business or a place, the split was 100% labrador, 0% Google, and 25% of results arriving with no snippet at all.

Then they explain where the words come from. Verbatim: "Here the copy does not come from your site at all, but from feeds: Yelp, TripAdvisor, Google Maps. Each provider is recognisable by its id format. The most telling fact: nearly one in five Yelp-labelled records carries a link copied from the business's Google Business profile, while Yelp itself shows the bare address. OpenAI does the stitching."

When a shopper on a free ChatGPT account asks about a local business, the description they read was not written by that business. It was assembled by OpenAI out of third-party listing feeds. And OpenAI is splicing records together across providers, taking a link from one source and attaching it to a listing from another.

Your website, the thing you pay for, the thing your vendor charges you to optimize, was not in that transaction.

We have done this before, and dealers lost the first round

If this feels familiar, it should.

In the late 2000s, dealers spent real money on websites while Google quietly built the Local Pack out of citation data: directory listings, aggregator feeds, business profiles. Stores that had obsessed over their homepage discovered that the three-result box above it ran on data they had never touched, some of it wrong, some of it entered by a franchise group's marketing coordinator in 2004. The winners were not the dealers with the prettiest site. They were the dealers who did the unglamorous work of cleaning up their name, address, phone and hours everywhere those records lived.

Then Google Business Profile matured, dealers got a handle on it, and the industry settled into the assumption that local data was a solved problem.

This data suggests a new layer just got built on top of that stack, and it draws from the same kind of feeds, with a new aggregator doing the stitching and no interface for you to correct it.

The honest caveat, because this deserves one

Two things about this research need saying plainly.

First, independent researchers looking at the same stream do not fully agree on what they saw. Suganthan Mohanadasan ran his own network-traffic analysis of ChatGPT's source selection, published July 14 and updated July 22. He identifies five pipes rather than four, adding a bing value he describes as newly rolling out and cohort-gated. More importantly, he reads labrador as licensed content and does not conclude that OpenAI runs an independent index at all. He describes the system as routing across commercial fetchers.

That is a real disagreement about a load-bearing fact, and I am not going to pretend it is settled. What both analyses agree on is the part dealers should act on: multiple retrieval pipelines exist, the one you get varies by account and cohort, and the pipeline that serves local business questions on free accounts is feeding on listing data. Mohanadasan's own example of the variance is blunt: "Same page, same week, arrived at David's browser stamped bing and at mine stamped bright."

Second, both efforts are outsiders reverse-engineering a server stream, not OpenAI describing its own product. Mohanadasan flags his own limits, noting his counts are "1 account and directional." OpenAI's crawler documentation says OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features" and says nothing about an in-house index. And neither study tested car dealerships specifically. They tested local businesses as a category.

So treat the mechanism as well evidenced and the exact plumbing as contested. That is still enough to change what you do this month.

What this means for your dealership

Your Google Business Profile is doing double duty now. It was already the thing that decided your Local Pack presence. On this evidence it is also feeding the answer a ChatGPT shopper reads. Every stale hour, every wrong phone number, every service department that still lists Saturday hours it dropped in 2024 is now propagating into a channel you cannot see.

Yelp matters again, and most franchised dealers have ignored it for a decade. We flagged this in July when OpenAI licensed Yelp's review database into ChatGPT for local recommendations. This capture is the first outside look at how that licensing actually shows up in an answer. If OpenAI is pulling Yelp-labelled records and stitching them with links from elsewhere, an abandoned Yelp page with a nine-year-old photo and a wrong sales manager is live inventory in an AI answer. You do not have to love Yelp. You have to make sure the record is correct.

Consistency across feeds is now a technical requirement, not hygiene. The stitching behavior is the risk. When one provider has your old address and another has your new one, something downstream picks. You want every source saying the same thing so there is nothing to pick between.

Your site still decides the non-local questions, and the first 200 characters are the whole game. Plenty of the car-shopping journey is not "dealer near me." It is "is a certified pre-owned warranty worth it," "what should I put down on a lease," "best three-row SUV under 45k." For those pages, RESONEO found the stored snippet "is cut off just past two hundred characters, and it is taken from the start of the page's rendered body," and that when a page has an H1, the snippet contains it eight times out of ten. Your H1 and your opening sentence are the excerpt. Not the meta description you wrote, not the paragraph three screens down.

Getting retrieved is not the same as getting read. RESONEO found that out of roughly eighty pages pulled, only one is actually opened by the model, and that a page the model opens ends up cited 74% of the time against 7% for a page that merely sat in the grounding set. Being in the pile is close to worthless. Being the page worth opening is everything.

Nothing about this is fast. RESONEO measured 13% of snippets running more than a month behind the crawl. A price change, a new store name after an acquisition, a relocated service drive: assume the AI layer is weeks behind you, and stop putting perishable claims in the first two lines of a page.

What to do about it

1. Audit your Google Business Profile this week, every rooftop and every department. Hours, address, phone, categories, primary photo. Treat it as customer-facing copy, because it now is.
2. Claim and correct your Yelp listing. You are not doing this for Yelp traffic. You are doing it because a feed you do not control is being read aloud by something a shopper trusts.
3. Run a name, address and phone consistency check across the major aggregators and fix the mismatches. Old DBAs from a prior owner are the usual culprit at a franchised store.
4. Rewrite the H1 and the opening 200 characters on your highest-intent pages so they read as a clean, quotable answer with no hero-image alt text or boilerplate in front of them.
5. Confirm you are not blocking OAI-SearchBot in robots.txt or at the WAF. OpenAI states plainly that sites opted out of it "will not be shown in ChatGPT search answers." We have written about why blanket bot-blocking backfires.
6. Stop buying a single AI visibility number. Ask any vendor selling you one whether they test free and paid accounts separately. The pipeline split says the answer changes.

The window closed, the shift did not

The most useful thing about this research is also the most uncomfortable: it existed because of a labeling field that OpenAI has since removed. We got one clear look, and now the room is dark again.

That is going to be the pattern. The systems deciding which dealership a shopper hears about are getting less observable, not more, and the measurement industry growing up around them is selling confidence it cannot actually source. The dealers who do well here will be the ones who stop chasing scores and go fix the underlying records, because clean, consistent, well-structured data is the one input that works no matter which pipe fetches it.

That is the same answer it was in 2009. It is just worth more now.

If you want a read on what an AI assistant currently finds when it looks for your store, book a walkthrough and we will go through it with you.