Can Independent Car Dealers Get Cited in AI Search? [Study]

I ran 69 used car buying questions through ChatGPT and Perplexity, logged every source the two systems retrieved and cited, and classified 262 domains by hand. The run produced 2,584 retrieved sources and 967 citations. This article reports what I found, in full, including the parts that contradict common advice and the parts I cannot explain.

Nobody has published a source level map of UK used car buying in AI search. This is that map: who owns the answers, what changes the answer, and what the data does not support. Everything here is open for anyone to re-run or disagree with.

01The Six Question Types I Tested

Each level adds one piece of information to the question.

I built 69 middle of funnel questions a UK used car buyer asks in real life. Not “is a Qashqai reliable”, because the motoring press owns that stage. These are the questions about where to buy and from whom, asked after the buyer commits to buying and before anyone walks onto a forecourt.

Every question sits at one of six levels of specificity. That ladder is the experiment.

1. Channel. “Where to buy a used car online UK”
2. Model. “Where to buy a used Ford Fiesta UK”
3. Requirement. “Where to buy a used car under £10,000 UK”
4. Place. “Best used car dealers in Dudley”
5. Model and place. “Where to buy a used Golf in Coventry”
6. Requirement and place. “Cheap used car dealers in Kidderminster”

Each question went to both platforms through their APIs on the same day, single turn, fixed settings. A six question product consideration control group sits alongside, to prove the contrast rather than test the hypothesis.

02Who Owns UK Used Car Answers

The whole citation set, before any slicing.

Across all 967 citations, the picture divides into four blocks: the marketplaces, the dealers, the public bodies, and a long tail of everything else.

n = 967 cited sources. Percentages rounded.

Three things stand out before any further analysis.

AutoTrader is not one of several marketplaces. It is the market. At 15.7% of all citations and 152 individual appearances, it outweighs the next single domain by more than three to one. The second most cited domain in a study about buying cars is gov.uk.

The motoring press has almost no presence. WhatCar, Auto Express, Autocar, Parkers, Carbuyer and Honest John together account for 2.4% of citations. These publications dominate organic search for used car research, and they barely register on questions about where to buy. Their authority does not transfer across the funnel.

Automotive specific sources take 78.2% of citations, against 21.8% for general purpose sites. AI reaches for vertical specialists on these questions rather than big general authorities, which is the opposite of the pattern most brand visibility studies report.

03Specificity Decides Who Gets Cited

Visibility climbs from one question in ten to every question in the set.

This is a gradient, not a switch, and it holds across all six levels.

Adding a model name changes nothing. “Where to buy a used Ford Fiesta UK” performs identically to “where to buy a used car online UK”, at 10% in both cases.

Adding a place changes everything. Adding a requirement on top of a place lifts it to every question in the set.

The gate is not how much the question knows about the car. The gate is whether the question contains a location.

04Independents Outcite the Marketplaces at Town Level

Not “get a look in”. Outcite them, narrowly, on share of sources.

At town level, independent dealers took 26.0% of all citations. AutoTrader, Motors, CarGurus and Cazoo together took 25.7%.

Ask a general question and the marketplaces own the answer outright. Name a town and the small dealers become, collectively, the most cited thing in the answer.

05The Two Platforms Disagree

Different magnitudes, identical direction.

ChatGPT leans harder on franchise groups and government sources. Perplexity leans harder on independents and motoring bodies. Both move in the same direction as questions become local, and ChatGPT’s twenty fold shift from 1% to 22% is the steeper climb of the two.

A pattern on one platform is a quirk of that platform. A pattern on two systems built independently by different companies describes something about how AI search handles local intent.

---

06Conversational Phrasing Made No Difference

A null result, published because the opposite advice circulates widely.

Twelve of the 69 questions were written the way a person talks. “I’m in Birmingham looking for a used car around £10k, which local dealers are worth going to?” I included them to test conversational phrasing against keyword style phrasing.

The difference vanished. 54% of keyword style questions cited an independent against 50% of conversational ones, at p = 1.0.

Independent survey work points the same way. BrightLocal’s roundup reports that 75% of ChatGPT users type keywords rather than natural language when looking for local services, which is Sagapixel research arrived at by a completely different method.

What moved the result was the information inside the question, a town and a budget, not the grammar wrapped around it.

07Three Quarters of Retrieved Pages Never Reach the Answer

Perplexity shows its working, and the discard pile teaches more than the citations.

Perplexity reports every source it retrieves, not only the ones it cites. Across this study, 73.3% of retrieved sources never appeared in an answer.

The system pulled up AutoTrader 215 times and left it out. Trustpilot 89 times. The AA 54 times.

Absence from an answer has two separate causes. Either the system never found the page, or the system found it and judged it did not answer the question. Those demand opposite fixes. The first is a discoverability problem. The second is a content problem.

A page retrieved and discarded 215 times is not invisible. It is unconvincing.

08One Citation in Twenty Sent a British Buyer to America

The finding I did not go looking for.

US domains took 5.1% of citations. Some of that is harmless: cars.com and CarMax appearing in a general answer about buying online.

Some of it is not. The question “where to buy a used car on finance with bad credit UK” returned American subprime motor lenders, the US Federal Trade Commission and the US Consumer Financial Protection Bureau. None of those lend in Britain, and none of their consumer protections apply here. A buyer in a vulnerable financial position asked a British question and received an American answer.

The questions that leaked worst had one thing in common: they described a situation that exists in both countries, in language that exists in both countries. Finance, credit, consumer rights. The more the question resembled a universal problem, the more likely the answer crossed the Atlantic.

I flag this as a finding rather than a conclusion. Measuring it properly needs a study designed around it, which is the next thing on my list.

---

09What This Means for an Independent Retailer

What the data supports, and what it does not.

Specificity has always been the small operator’s lever

This finding restates, inside AI search, something the SEO industry established about organic search. Search Engine Land’s guide puts it plainly: longer, more specific phrases carry lower volume but higher intent and less competition, which makes them easier to rank for, and the guide singles out smaller websites as the main beneficiaries.

Instead of battling giants for generic terms, you become the exact answer to a specific question.

Neil Patel, writing on long tail search strategy

BrightEdge extends the same logic to generative systems, arguing specific queries carry greater citation potential because they let AI systems identify focused answers.

My data puts a number on that argument for one industry. In UK used car buying, the specific thing that unlocks it is a place name, and a model name does nothing at all.

The local opportunity is opening, not closing

BrightLocal’s 2026 Local Consumer Review Survey records that use of ChatGPT and generative AI for local recommendations climbed from 6% to 45% in a single year. The audience arrived.

The SOCi Local Visibility Index 2026 records that fewer than half the businesses leading Google local search results also appear in AI local recommendations. Winning Google local carries no guarantee here. That gap is the opportunity, and it closes as competitors notice it.

BrightLocal’s Consumer Search Behavior research adds the stakes: 72% of consumers assess three or fewer businesses before deciding, and only 1% look at more than ten. An AI citation places you inside a shortlist that was already short.

The four moves, and one to skip

The final row is a null result. I include it because the opposite advice circulates widely.

On schema markup

Mark up your business properly. Use the AutoDealer type, fill in addressLocality with your town and areaServed with the surrounding towns buyers actually type, and keep opening hours and contact details current. That takes an afternoon.

Then stop thinking about it. This study measured nothing about schema, so I have no evidence it changes citation rates either way. Treat markup as hygiene rather than strategy. The variable that separated a 3% citation rate from a 26% one was whether the question named a place, and the work that follows from that is the content on your pages, not the JSON underneath them.

---

10Method

Every step, in enough detail to run it again and get the same answer.

11Why I Left Google AI Overviews Out

A deliberate exclusion, not an oversight.

Google’s AI Overviews reach more people than ChatGPT and Perplexity combined, so leaving them out needs an explanation.

The unit of measurement differs. ChatGPT and Perplexity accept a full question through an API and return an answer with its sources attached. AI Overviews is a feature inside a search results page, reached through a third party scraping service rather than an API. The retrieval path, the caching behaviour and the response format all differ. Mixing them into one dataset compares things that are not alike.

Including it forces the wrong question format. Keeping AI Overviews comparable means writing short keyword strings, because that is what a search box receives. This study needed conversational questions with budgets and locations in them. Designing the whole question set around the least flexible surface costs more than the extra platform gains.

It fires unpredictably on commercial questions. AI Overviews appears reliably on informational queries. On commercial ones like “best used car dealers in Dudley”, Google frequently serves the local pack and ads instead. A platform answering an unknown fraction of the question set produces a denominator nobody can report honestly.

Adding it is the obvious next step, run as a separate arm on the identical question set so the two remain comparable rather than blended.

12What This Study Does Not Establish

The limits, stated plainly, so you can weigh them yourself.

Citations are not customers. I hold no data on whether an AI citation sends anyone to a forecourt. Everything here measures visibility, which sits upstream of revenue and is not the same thing.

The geography is narrow. Town level questions covered the West Midlands and Shropshire. Whether the same gradient holds in Cornwall or Aberdeen remains untested.

Cells are small. Four regional questions. Eight to sixteen questions at each depth level. The direction is clear and the effect size is large, which is why it survives a sample this size. Individual percentages carry wide intervals, printed on every chart rather than hidden.

One run, one day. AI answers move. I have not yet measured how stable these citation sets are across weeks, and existing industry research on citation volatility suggests they move considerably. A second run is scheduled and I will publish the comparison whichever direction it goes.

Four questions are missing. Of 138 intended API calls, four ChatGPT calls failed on quota or timeout. Their absence is recorded rather than papered over.

Configuration shapes the platform gap. ChatGPT ran on a cost optimised model with a reduced web search context. Perplexity ran on its low preset. Figure 3 compares those two configurations, not the two companies in general.

Correlation, not causation. I measured which sources appear. I did not test what makes them appear. Section 09 describes how to state the information that separated a 3% citation rate from a 26% one. It does not prove that stating it produces the citation.

13Frequently Asked Questions