AI Recommended the Brand by Name, Then Linked Somewhere Else 69% of the Time. Your Inventory Already Lives on Those Sites.

There is now a number for where an AI answer sends the click after it names a business as its recommendation.

Shero Commerce, a Shopify agency, ran a citation test against Google AI Mode, ChatGPT and Perplexity across 60 buying categories, alongside a content audit of 1,000 Shopify stores. 883 of those stores returned live product data, and the study analyzed 8,573 product descriptions against every source the three tools cited.

The tools cited 1,851 sources. 2.8% of them were pages owned by a brand. 59% were publishers.

Then the sharper cut. The tools did name real brands as recommendations. Of 159 times a brand got picked by name, its own page was the cited source 31% of the time. The other 69% of the time an AI said the brand out loud and handed the reader a link to somebody else. In a third of the 60 categories, no brand got named at all.

Search Engine Journal covered it on August 26 under a headline that says the whole thing: AI tools recommend brands but cite other sites. One piece of housekeeping before anyone quotes these numbers at you: Shero published the research in July, so this is a measurement taken a few weeks ago rather than something that happened last week.

It is also the second dataset this month pointing the same direction. In August we covered RESONEO's capture of ChatGPT's retrieval pipeline, which found business and place questions coming back from Yelp, TripAdvisor and Google Maps feeds instead of from the business's own site. That study measured which pipe the answer traveled through. This one measures which domain got the credit at the end of it. Two different questions, one uncomfortable answer.

There were no car dealers in either dataset. Read on anyway, because the mechanism this study isolates is one the car business industrialized twenty years ago.

Where the copies actually were

The duplication numbers are the part that should stop a dealer cold.

20% of the 8,573 product descriptions appeared word for word, or near enough, on a domain the brand does not own. Another 22% of products shared near-identical copy with a different product inside the same store. Shero detected this with shingle fingerprinting on 8-word sequences and then verified the hits with quoted web searches, so these are literal string matches rather than a similarity score somebody eyeballed.

Here is the finding underneath it. The copies were not sitting on competitors' stores. They piled up on the retailers and marketplaces that resell the brand. Nordstrom. Amazon. Sally Beauty. One MCoBeauty product sentence turned up near verbatim on seven other domains, grocery chains included.

Substitute your own business into that sentence. A brand writes its copy once. The channel partners that resell its product republish that same copy at scale, from domains with more authority and more inbound links than the brand will ever have. When an AI assembles an answer, it holds a dozen copies of the same paragraph and picks one.

Microsoft has described how that pick gets made. From the Bing Webmaster blog in December 2025: "LLMs group near-duplicate URLs into a single cluster and then choose one page to represent the set." The same post, on syndication: "When your articles are republished on other sites, identical copies can exist across domains, making it harder for search engines and AI systems to identify the original source."

Bing stops there, and so will I. That post does not claim the unpicked copies never get cited. Several vendors are already quoting it as though it does.

Your store has been doing this on purpose since 2005

A franchised dealership is a reseller that publishes supplier copy and then pays to syndicate it further. That is the content architecture of a VDP, described honestly.

The vehicle description on your new car pages started at the manufacturer. Every other franchise dealer in your brand publishes a version of it. Then your inventory feed pushes the same vehicle, the same photos and frequently the same description out to Cars.com, CarGurus, Autotrader, Edmunds, KBB, Carfax listings, your OEM's national inventory locator and whatever else appears on the vendor invoice. You pay for that distribution. It has been a defensible decision for lead volume for two decades and it still moves cars.

It also constructs, deliberately, the exact structure this study measured. Original copy on a low-authority domain. Near-identical copies on high-authority marketplace domains that already outrank you for the model name and have for years.

The DTC brands in the study arrived at that position by accident, through resellers they do not control. Dealers arrived there on purpose, through vendors they write checks to every month.

What this means for your dealership

The recommendation and the click are now two separate events. For fifteen years they were one thing. A shopper searched, saw your store, clicked your store. This study pulls them apart. An assistant can name your dealership as the right place to buy the car and still route the reader to a listing site. When a GEO vendor shows you a report full of "mentions," ask whether it also shows citations, and ask which domain received the link.

You are likely paying for the page that gets cited instead of yours. The third-party sites most likely to be carrying your inventory copy are the same ones on your monthly marketing invoice. If AI answers keep favoring marketplace listings, your cost per lead does not disappear. It shifts further toward the vendors who already own the handoff. That is a budget conversation before it is an SEO one.

A thin-content audit will probably lie to you. 30% of the description fields in the study came back under 50 words. Then the researchers measured the rendered HTML instead of the field, and only 15.6% of cleanly measured stores were genuinely thin, 27 out of 173. 71% of the stores with thin description fields had rich content on the actual page, because modern themes render copy from places the field never sees. Your dealer platform behaves the same way. When somebody arrives with a deck declaring your VDPs empty, ask which layer they measured before you buy the remedy.

Structured data is not coverage. 59% of the stores emitted product structured data while their descriptions ran under 50 words or were missing outright. Marked-up emptiness is still emptiness. Dealer platforms ship Vehicle and Offer schema by default, which is worth having and is no evidence that there is anything on the page worth quoting.

The counter-case, honestly. Two of them, and both matter.

First, this is a Shopify DTC dataset covering beauty, apparel and consumer goods. Not one car dealer was tested. These percentages are not your store's percentages, and anyone who repeats them to you as automotive numbers is inventing data.

Second, the part the vendors quoting this study will skip: content depth barely predicted AI readiness at all. The Spearman correlation between the two was 0.13. Page speed against AI readiness came in at 0.02. Only the deepest quartile of content scored meaningfully higher, 44.9 against roughly 41. So "write longer VDP descriptions and the AI will cite you" is not a finding of this research. What the research establishes is where citations land and why duplicated copy makes any single page a poor candidate to represent its cluster. Rewriting the copy is a sound response to that problem. Promising a citation lift from it goes past the evidence.

One real disanalogy, and it changes the priority order. A DTC product page can live for years and accumulate links the entire time. A used VDP goes up, the unit sells, and the URL is gone. Whatever original text you write there gets a short window to be crawled, clustered and selected. Your model research pages, service and fixed ops pages, financing pages and the dealership's own about page carry no such expiry. That is where original writing compounds, and it is where most stores have the least of it.

What to do about it

1. Run the study's method on your own inventory this week. Take the description from five of your highest-priced new VDPs, paste one distinctive 8 to 10 word sentence into Google inside quotation marks, and count the domains that come back. If your sentence sits on forty franchise dealer sites and three marketplaces, you have your answer in ten minutes and it cost nothing.
2. Rewrite the top fifty units, not the whole feed. Shero's own conclusion is that "the highest-leverage fix is also the least technical," meaning store-specific editorial copy in place of supplier text, concentrated on the highest-intent products rather than the entire catalog. For a dealership that means the units carrying your gross, not every stock number on the lot.
3. Make your vendor separate mentions from citations in writing. A mention count with no cited domain beside it tells you nothing about where the traffic went. Put the distinction in the reporting requirement.
4. Confirm your platform renders descriptions server-side. View source on a VDP and search for a phrase from the description. If the phrase only appears after JavaScript runs, that is worth raising with your website provider before you spend a dollar on new copy.
5. Move budget toward pages that survive. Model overview pages, service menus, trade and financing explainers and local landing pages hold their URLs for years. Original text there is an asset. Original text on a unit that sells in six weeks is a rental.
6. Read your syndication agreement. Ask your feed provider what it is permitted to pass through and whether you can send a different or shortened description to marketplaces than the one on your own site. You may not have that lever. You should know either way.

The part worth remembering

The old bargain was simple. Write it once, push it everywhere, take the leads. That bargain assumed a search engine would hand the shopper a list and let them choose the source. AI answers close that loop themselves, and when they do, they pick one page to speak for every identical copy of a paragraph. Across 1,851 citations in this study, they picked a brand's own page 2.8% of the time.

Your dealership is the most syndicated business in your zip code. That was an advantage back when distribution was the scarce thing. It is a liability now that being the original is the scarce thing.

The ten-minute check in step one costs nothing and nobody has to sell you anything for you to run it. If you want a second set of eyes on what your VDP copy actually looks like to an AI assistant, book a walkthrough and we will run it against your store together.