The Dealer Was Named in 1 of 27 AI Answers. One Page Later, It Was 24.

The Results

Source: ChatGPT, three prompts across nine markets, 27 responses per reading. Baseline July 13, 2026 before the page went live. Re-measured July 21 and Aug. 13.

The Challenge

This dealership sits in Georgia, a town of about 16,000 people on the coast.

The customers are not in the dealer’s home city. They are in Jacksonville, 70 miles south. In Savannah, 78 miles north. Scattered across south Georgia and north Florida in cities the store will happily drive a truck to.

So the dealership does what a lot of stores do. It offers free home delivery, and it says so on the website.

Or rather, it said so in a picture.

The delivery offer lived in the homepage banner graphics. A designer set the words "free delivery" in a slider image, and that image sat on the site doing its job for human visitors, who could see it perfectly well.

But LLMs and search engines cannot read a picture of a sentence.

So when a shopper in Jacksonville opened ChatGPT and asked which GMC dealer would deliver a truck to their house, this store was not in the answer. We measured it before we changed anything. Across three delivery questions in nine markets, 27 responses in total, the dealership was named once.

What this means for you: your website probably makes promises that no AI assistant can see. Delivery terms, financing offers, service guarantees, price-match policies. If it is set in a banner graphic, it may as well not be written down.

The Hypothesis

Ashley, the analyst on this account, had a specific read on the problem.

The offer was real. The store genuinely delivers, genuinely covers 200 miles, and genuinely handles the purchase online. None of that was in dispute. The problem was that none of it existed as text a machine could quote.

So the hypothesis was narrow. Take the value proposition out of the image and write it as an actual page. Say how far. Say what it costs. Name the cities. Answer the questions a delivery shopper would ask, in the words they would ask them.

There was no delivery page on the site before this. Not a thin one, not an old one. It did not exist. We built it to find out whether the machines would pick it up.

What this means for you: this was not a rewrite or an optimization. It was writing down something the store had been saying out loud for years.

How We Tested

Three questions, the kind a delivery-minded shopper actually types:

1. "GMC dealer who will deliver to my house in [city]"

2. "2026 GMC Sierra 1500 that can be delivered to my house in [city]"

3. "New GMC dealership that does online purchases with home delivery in [city]"

Run across nine markets: Jacksonville, Orlando, Tallahassee, and Gainesville in Florida, and Savannah, Macon, Augusta, Albany, and Valdosta in Georgia.

Three questions times nine markets is 27 responses per reading. We took three readings. One on July 13 before the page went live, one on July 21, and one on Aug. 13.

Every prompt was run twice, two different ways. Once by hand in a fresh chat with no history, and once through the API. 

That second part matters more than it sounds. Most studies like this get run once, in somebody's own logged-in account, with months of chat history shaping the answer. That does not measure what a stranger sees. Running clean and cross-checking is the difference between a result and an anecdote.

The Outcome

The page was working inside 8 days

On July 13, before the page went live, the dealership appeared in 1 of 27 ChatGPT answers.

On July 21, eight days later, it appeared in 15.

That is the number we did not expect. The working assumption in this industry is that AI visibility takes a quarter to move, if it moves at all. This took a little over a week.

What this means for you: if you have been putting this off because you assumed the payoff was months away, that assumption is wrong. At least for a page that answers a question nobody else on your block has answered in writing.

A month later, 24 of 27

By Aug. 13, the dealership was named in 24 of 27 answers. Eight of the nine markets name the store on all three questions.

That is a gain of 85 percentage points.

Broken out by question, each scored across the nine markets:

The two questions that started at zero are the ones worth staring at. Before the page, there was no combination of words a shopper could type about ordering a Sierra online and having it delivered that would surface this store. Now there is.

One number went down, and that is normal

The core delivery question did not climb in a straight line. It went from 1 market to 9, then back to 8.

We are pointing that out because it is what actually happened, and because it is not alarming. Rankings move. A page that sits at position 3 in Google one week and position 6 the next is an ordinary Tuesday, and nobody writes a warning label about it. LLM answers behave the same way. You watch the trend across readings, not any single reading.

What this means for you: treat AI visibility the way you already treat rank tracking. One check tells you almost nothing. Four checks tell you where you stand.

Orlando never took, and we cannot tell you why

Orlando is the one market where the dealership still does not appear, on any of the three questions.

We tried three explanations and none of them held up.

We thought Orlando might be too far, sitting at 210 driving miles against a stated 200-mile delivery radius. But Tallahassee is 231 miles, and the store is named there on all three questions.

We thought the page might not mention Orlando. It does, by name, in a section listing eligible cities.

We thought Orlando might simply have more GMC dealers competing locally, giving ChatGPT no reason to reach 210 miles for another one. There is something there. Orlando has 11 GMC dealerships within 50 miles, more than any other market we tested. But Jacksonville has exactly as many GMC stores within 30 miles as Orlando does, and Jacksonville works fine. 

So we are leaving it open. Eight markets took, one did not, and the honest answer is that we do not know what makes Orlando different. Yet. We’ll continue to monitor visibility over time.

The Thing We Did Not Expect

In Savannah, before the page, the dealership was mentioned once, in the last line of the answer, as an afterthought.

After the page, it moved into the map panel at the top and became a numbered recommendation, with the free delivery terms quoted back and Savannah specifically identified as a covered market.

Here is the part worth sitting with. In that same answer, the assistant listed a local competitor first, carrying a 4.5-star rating. Our client appeared second, carrying 2.4 stars.

A shopper comparing those two ratings side by side would not pick our client. The assistant recommended the store anyway, because it was the one that answered the question being asked.

We are not going to tell you reviews do not matter. They matter, and a 2.4 is a problem worth fixing on its own terms. But the assumption that a strong star rating is what earns you a place in an AI answer did not hold here. What earned the placement was a page that said, in plain text, exactly what the shopper wanted to know.

What this means for you: reputation and answerability are two different jobs. You need both. But if you are waiting to fix your rating before you write the page, you have the order backwards.

What This Costs You

Somebody has to write it, and it has to be specific. A page that says "we offer delivery" does nothing. This one works because it states the radius, the cost, the exclusions, and the cities by name. Ambiguity is what got the store 1 of 27 in the first place.

The terms have to stay accurate, because the machine will quote them. This is the part most people miss. ChatGPT is now repeating this dealership's delivery terms to shoppers in nine markets. Whatever you publish, make sure the desk will stand behind it.

It needs maintenance nobody will remember to do. Delivery radius changes. Fees change. Coverage areas change. The page has to change with them, and there is no alert that tells you it has drifted.

What We Are Not Claiming

This is 27 responses, not a large study. Nine markets, three questions, one assistant. Two runs each that agreed with one another, which is a consistency check, not a large sample. We would not publish this as a general finding about AI search. It is one dealership, one page, one month.

We tested ChatGPT only. Not Gemini, not Perplexity, not Google's AI Overviews. Those may behave differently and we have not looked yet.

What Happens Next

This is reading one of an ongoing measurement, not a finished study.

We are re-running these prompts over the next three to six months, and we will publish what we find whether it confirms this or complicates it. The interesting question now is not whether the page works. It clearly does something. The question is where the effect stops.

Hats off to Ashley for running this one.