How to Merchandise Inventory for AI

A shopper tells an assistant they have AED 95,000 all in, they want something reliable and low-mileage, and they would rather buy locally. Honda Civic, Subaru Crosstrek, Toyota Corolla Cross. Thirty seconds later they have three cars and, increasingly, three dealerships to call.

That shopper may never open forty vehicle pages. They may never compare dubizzle to your website themselves. The assistant does the first pass: it asks clarifying questions, searches, cross-checks a VIN, and decides which stores are even in the conversation. The dealer who still merchandises only for a human scrolling a results page is optimizing the part of the journey that is shrinking.

Every dealer wants to know how that shortlist got made. The honest answer is that it was not one process. There are four different ways a vehicle can enter an AI answer, they have almost nothing in common, and the one most dealers are optimizing for is the weakest of the four. This guide covers all four, what each one actually requires, why the marketplaces keep winning them, and the merchandising spec a listing has to pass before an assistant will confidently name you.

The question dealers are now asking

For years the work was merchandising inventory for consumers and optimizing websites for Google. That work is not obsolete. It is incomplete. The new question is narrower and harder:

Is this inventory clear, accurate, transparent, and structured enough that an AI agent can recommend it without guessing?

An assistant does not need a “Manager’s Special” or a 500-word templated description of how beautiful the car is. It needs facts it can understand and verify: VIN, mileage, trim, equipment, the real price, fees, history, certification, warranty, and whether the car is actually on the ground. When those are not clear, it does not call you to ask. It looks elsewhere. Marketplaces, history sources, other dealer listings. Whatever will let it answer without inventing a number.

What the assistant actually does with a AED 95,000 ask

Watch a real shopping conversation and the sequence is fairly stable. The shopper states a budget, a need, and a radius. The assistant then has to turn that into a shortlist. It cannot do that from a photo and a slogan. It has to resolve a set of filters, and each filter is a fact that either exists in text or does not.

Then come the questions dealers have been answering for humans for twenty years, and mostly answering badly for machines. How much can you rely on the advertised price versus the transaction price? Is a CPO claim real, or a badge on a photo? Are the rebates available to this buyer, or only if they qualify for a program the page never names? Which add-ons are required? The answers to those questions used to live in a desk conversation. They now have to live in the listing, because the assistant is doing the first pass without you in the room.

The four doors into an AI answer

When an assistant produces a car, the data behind it arrived through one of these paths. They are listed strongest to weakest, by how much control the source has over what gets said.

1. A partner app or built-in integration. Some marketplaces are not sources the assistant found; they are software running inside it. When the assistant needs listings, it calls that partner directly and renders what comes back. This is the highest-privilege door and it is closed to an individual dealer acting alone.
2. A merchant catalog. The assistant answers from a product index it built ahead of time by ingesting structured feeds from approved sellers. No crawling happens at question time. The data is as fresh as your last feed push.
3. Live retrieval. The assistant runs its own searches while the shopper waits, fetches whatever pages come back, and quotes what it can parse from the raw HTML. This is the door open to every dealer, and it is the one this guide’s siblings mostly cover.
4. Model memory. Whatever the model absorbed in training. Useful for “is the Crosstrek reliable,” useless for “what is on the lot today,” and impossible to update or correct.

Being good at door three does not get you into doors one or two. They are separate systems with separate applications, separate formats, and separate approval. That is the single most expensive misunderstanding in dealer AI strategy right now. A beautiful VDP that ranks on Google still loses to a partner app that never had to be discovered.

Why the marketplace keeps being the answer

If you have watched an assistant shop for a car, you have probably noticed it leaning on a large marketplace as its primary source. It is tempting to read that as an SEO outcome, as though the marketplace outranked you. It usually is not.

In January 2026, AutoTrader in Canada launched an app inside ChatGPT, the first automotive marketplace anywhere to do it. Listings render in the chat; one click takes the shopper to the marketplace, not to the dealer site. CarGurus publishes an MCP connector in ChatGPT and Claude that searches its listings. Separately, third-party analysis has found Cars.com to be the most-cited public automotive marketplace across ChatGPT, Google AI Overviews, and Google AI Mode. Expect dubizzle, DubiCars and YallaMotor to follow the same playbook in the Gulf; the mechanism is identical and the data they hold is already structured.

TRADER Corporation, January 2026 · CarGurus help docs, August 2026 · Semrush, September 2025, cited by Cars.com, November 2025

So the marketplaces are winning on all four doors at once, and the compounding is the point:

• They are inside the assistant. A partner app does not have to be discovered. AutoTrader did not win that slot by writing better meta titles, and dubizzle will not either.
• Their data was born structured. A marketplace holds normalized inventory for thousands of rooftops in one schema. Producing a compliant feed for a new program is a weekend of work for them and a procurement project for a dealer group.
• Their pages are built to be read. Server-rendered, consistently marked up, one URL per VIN, with a crawl history going back years.
• They aggregate the corroboration. When an assistant wants to check a fact about a specific VIN, the marketplace is usually where it checks. Your page is one vote. Their index is the ledger.

None of that is a reason to give up on your own site. It is a reason to be precise about which door you are actually working on, because the fixes are different. Door three is still yours: a server-rendered, fact-complete VDP that agrees with every other place that VIN appears. Door two is a feed problem. Door one is a partnership problem, and most rooftops will only get it through someone who already has the relationship.

The catalog programs, side by side

Every major assistant now runs some version of a merchant feed program. These evolve quickly and open on their own timelines, so treat specifics as directional and verify before you build. The shape, though, is consistent: each wants a structured inventory file, pushed on a schedule, in its own format.

OpenAI commerce documentation, 2026 · Microsoft Advertising · Google Merchant Center vehicle ads documentation · Perplexity Merchant Program, 2026

The gap worth noticing

Read that last column again. Google built a vehicle-shaped feed. The assistant-native programs built a general merchandise feed and have not yet extended it to cars. There is novin attribute in the OpenAI spec. There is no mileage field, no trim, no certification flag, no history report link. Title is capped around 150 characters. Description is plain text with a 5,000 character ceiling.

That gap is why a good listing today is not simply a matter of filling in a form. Until those specs grow vehicle attributes, the facts that decide a car deal have to survive inside the general-purpose fields, which means the ones you control on the page carry more weight, not less. Here is what a used car actually looks like in the OpenAI shape:

{
  "item_id": "1HGCV1F30LA512345",
  "title": "2020 Honda Accord Sport 1.5T",
  "brand": "Honda",
  "condition": "used",
  "price": "91500.00 AED",
  "availability": "in_stock",
  "url": "https://yourdealer.example/inventory/1HGCV1F30LA512345",
  "image_url": "https://yourdealer.example/img/512345-1.jpg",
  "seller_name": "Maple Ridge Auto",
  "description": "61,800 km. One owner. GCC specs, full service history. 1.5L turbo, CVT, front-wheel drive. Heated seats, adaptive cruise, CarPlay. Price includes 5% VAT; RTA registration and transfer fees extra."
}

Note what had to happen. The VIN became the item_id, because a stable unique identifier is required and the VIN is the only one a car has. Mileage, drivetrain, trim equipment, title status, and the fee disclosure all got pushed intodescription, which is a factual field OpenAI tells merchants to keep plain. That field is the closest thing to a vehicle spec sheet the format has. Filling it with “Manager’s Special, won’t last at this price!” wastes the only place the facts fit.

Advertised price, transaction price, and why the assistant cares

This is the merchandising problem dealers already know, now running inside a system that cannot be talked around on the lot. A shopper with AED 95,000 all in is not asking for the internet price. They are asking for the number they will write the check for. An assistant that cannot tell those numbers apart will either exclude the car or recommend it on a price that will not survive the desk, which is how you get a wasted lead and a model that learns your listings are unreliable.

Four price shapes show up constantly, and only one of them is usable:

• A price that includes every mandatory dealer-imposed charge. Doc fee, freight you require, dealer-installed accessories that are not optional. Tax, title, and registration can sit outside it if the page says so in text. This is the only number that can pass an all-in budget filter.
• A price that excludes a fee the buyer cannot decline. The assistant cannot add a fee or the VAT it cannot see. The car looks AED 1,500 to AED 4,500 cheaper than it is, then fails when another source discloses the fee. That is a contradiction, not a bargain.
• A price that only exists with a rebate or a lender. College grad, military, conquest, loyalty, or “price with approved financing.” If the condition is not stated next to the number, the assistant will treat it as the price. If the condition is stated, the car may drop out of an unqualified AED 95,000 search, which is the correct outcome.
• No price. “Call for price” does not create a phone call from an assistant. It creates an excluded vehicle. Every budget-filtered query drops the car silently.

The same logic applies to add-ons. If nitrogen, etch, or a protection package is required to buy the car, it is part of the price. If it is optional, say so. An assistant matching “out the door” will not invent that distinction in your favor.

CPO, history, and claims that cannot be checked

Certification is the other place listings die quietly. “Certified” on a banner is a claim. “Honda Certified, 7 years or 160,000 km from original in-service date, 182-point inspection” is a fact. The first one cannot be corroborated. The second one can be checked against the OEM program the assistant already knows exists.

History works the same way. If you have a report, link it and say what it shows in text: one owner, no accidents reported, GCC specs, full agency service history. If you do not have it, do not imply it. An assistant that can pull a third-party history on the VIN will, and a listing that is silent while the history report is not silent is another contradiction.

Warranty remaining belongs in months and kilometres, not “still under warranty.” Remaining factory coverage, a certified wrap, and a third-party contract are three different products. Name the one you are selling.

The verification loop nobody plans for

Here is the mechanic that changes how you should think about merchandising. An assistant does not simply read your listing and repeat it. When a claim matters and the source is thin, it goes looking for corroboration, and it has an unusually good key to search on: the VIN is a globally unique identifier for that exact car, printed on your page, on the marketplace listing, on the history report, and in every syndicated feed you send.

So the assistant cross-references. Your VDP against the marketplace listing. The marketplace listing against the history report. All of them against whatever other dealer pages carry the same VIN. And the answer it gives reflects the reconciliation, not your page.

This produces three failure modes that have nothing to do with crawlability:

• Contradiction. Your site says AED 91,500, the marketplace says AED 86,000, the feed says something else again. A model resolving conflicting sources will hedge, cite the source it trusts more, or quietly drop the car rather than state a number it cannot support.
• Silence. The facts a shopper is filtering on, certification, warranty remaining, accident history, the real out-the-door number, are absent from your listing. The assistant either fills them from a third party or excludes the car from a filtered answer. “Under AED 95,000 all in” is a filter your listing has to be able to pass.
• Unverifiable claims. “Certified,” “fully loaded,” “priced to sell,” and “like new” are not checkable against anything. A system built to avoid asserting things it cannot support will route around them.

The old failure was being invisible. The new failure is being visible and inconsistent, which is worse, because the assistant resolves the conflict using someone else’s data.

The practical consequence: price and availability parity across every place your VIN appears is now a discovery input. Not a nice-to-have, not a compliance chore. If your site, your marketplace listings, and your feeds disagree, you have handed the assistant a reason to prefer a source you do not control.

Where this collides with advertising rules

The same discipline is arriving from the regulatory side, which is unusual and worth using to your advantage internally, because it means one project satisfies two departments.

UAE consumer-protection law, administered federally by the Ministry of Economy, rests on the same principle: the advertised price must be the price the customer actually pays, stated clearly, with VAT and any mandatory charges made plain. The patterns regulators everywhere move against are the same ones an assistant trips over: prices that omit required fees, prices that quietly exclude VAT, discounts not available to every buyer, required down payments, prices contingent on a particular finance arrangement, required add-ons, and vehicles that are not actually for sale.

Federal Law No. 15 of 2020 on Consumer Protection. Not legal advice; confirm specifics with your compliance counsel. Emirate-level and free-zone rules can add to the federal position.

Read that next to the verification loop. A price that excludes mandatory fees is both an advertising exposure and a citation liability, because it will not survive reconciliation against a source that discloses them, and it cannot pass an all-in budget filter. A listing that states the real number, and says plainly what is and is not included, is simultaneously the compliant version and the machine-readable version. That alignment will not always hold, but right now it does, and it is the strongest internal argument for doing this work.

What AI ignores, specifically

Merchandising copy written for a human browsing a search results page does close to nothing in an AI answer, and some of it actively costs you. The recurring offenders:

• Urgency and superlatives. “Won’t last,” “Manager’s Special,” “Best deal in town.” Unverifiable, so unusable.
• Templated descriptions at scale. The same 500 words with the year and model swapped are not a fact source, and near-duplicate copy across an inventory is a known low-value content pattern. Per-vehicle facts beat per-vehicle prose that says nothing.
• Facts that exist only in images. The window sticker photo, the fee schedule rendered as a graphic, the equipment list burned into a banner. If it is not text, it is not a fact the assistant can quote.
• “Call for price.” Covered above. It is an exclusion, not a lead strategy.
• Facts behind an interaction. Specs in an accordion that loads on click, payments behind a form, equipment in a tab rendered after the fact. Most AI crawlers do not execute JavaScript at all. In a joint Vercel and MERJ analysis of more than 500 million GPTBot fetches, there was no evidence of JavaScript execution; the ChatGPT and Claude crawlers fetch script files without running them. Gemini and AppleBot are the exceptions.Vercel and MERJ, 2025

What this looks like on a real listing

Same car. Same photos. The difference is whether a machine can use the copy.

Typical lot copy

MANAGER'S SPECIAL!!! This BEAUTIFUL Accord WON'T LAST at this price. Fully loaded, like new, must see. Call for details. We beat any advertised price. Finance available for everyone!

The same car, merchandised for AI

2020 Honda Accord Sport 1.5T. VIN 1HGCV1F30LA512345. 61,800 km. One owner, GCC specs, no accidents on the service history (report linked). Front-wheel drive, 1.5L turbo, CVT. Equipment: heated front seats, adaptive cruise, Honda Sensing, Apple CarPlay, dual-zone climate, 19-inch alloys. Advertised price AED 91,500 includes 5% VAT. RTA registration and transfer fees are extra. Not contingent on showroom finance. In stock at Sheikh Zayed Road, Al Quoz 3, Dubai.

The second version is not prettier. It is checkable. An assistant can extract a VIN, an odometer, a trim, a fee-inclusive price, a location, and a history claim it can test. The first version gives it nothing it is willing to repeat. If those two listings compete for the same AED 95,000 query, the first one is not in the race.

The fields that actually decide a car

Strip the marketing away and a vehicle listing has to answer a fixed set of questions, in text, in the initial HTML, and identically everywhere the VIN appears. This is the merchandising spec, and it is short.

• Identity. VIN, year, make, model, and the actual trim. “Accord” is not a trim. “Sport 1.5T” is.
• Condition and use. Exact odometer with its unit, new, used, or certified, title status, and number of owners if you know it.
• Equipment as discrete items. A list, not a paragraph. Heated seats, adaptive cruise, sunroof, tow package. This is what “low-mileage Crosstrek with CarPlay” matches against.
• Price, with its boundaries stated. The number, the currency, what is included, what is not, and any condition attached to it. A price that depends on financing with the captive lender is a different number, and saying so is what makes it usable.
• Fees, itemized. Doc fee, freight, dealer-installed accessories. See above on why this is now doing double duty.
• Provenance and coverage. Certification program by name, warranty remaining in months and kilometres, history report availability. “Certified” alone is a claim; a named program with a term is a fact.
• Availability and location. Whether the car is on the ground today, and the dealership’s real address. ChatGPT now supports location sharing for local queries, so “within 50 km” resolves against a real coordinate. A vague service area does not survive that.
• A stable canonical URL. One page per VIN that does not move. Every corroboration path and every feed points at it.

The structured data trap

One correction worth making, because plenty of dealers are being sold the wrong version of it.

Google announced in June 2025 that it was retiring vehicle listing structured data along with six other types, and removed the supporting reports and documentation that September. The free vehicles-for-sale treatment on Search and Business Profiles went with it. If your vendor is still selling the Google vehicle rich result as the deliverable, that product no longer exists.

Google Search Central, June and September 2025

The wrong conclusion is to rip the markup out. Google retiring a visual treatment in its own search results says nothing about whether an assistant parsing your page finds machine-readable facts useful, and every other engine reads schema.org independently of what Google displays. Keep the markup. Just stop measuring it by a rich result that is gone, and note that Google now points sites at standard Product schema. Our vehicle schema markup reference covers what to emit.

A twenty-minute audit you can run today

Pick one vehicle you would like an assistant to recommend, then:

1. View source on the VDP and search the raw HTML for the price and the odometer reading. Not the rendered page; the source. If they are not there, no amount of feed work fixes door three.
2. Search the VIN. Note every place it appears. That is the corroboration set the assistant will use.
3. Compare the prices across that set. Any disagreement is a contradiction you are asking a model to resolve without you.
4. Check whether your advertised price includes mandatory fees, and whether the page says so in text. Then check whether that price depends on a rebate or a lender the shopper may not qualify for.
5. Read the CPO and history language the way a skeptic would. Named program and term, or a badge? Report linked, or implied?
6. Read your robots.txt. Confirm GPTBot, ClaudeBot, PerplexityBot, Googlebot, and Bravebot are not blocked. Claude retrieves through Brave’s index, so blocking Bravebot blocks Claude. See which AI crawlers to allow.
7. Ask an assistant your own shopper’s question. “Best low-mileage Crosstrek within 50 km of [your city] for AED 95,000 all in.” Note who it names and where the data came from. Then ask it how it treated advertised price versus fees.

What is coming: the agent on the other side

Feeds are the discovery layer. A second layer is being built on top for the transaction, and two competing standards are emerging. The Agentic Commerce Protocol comes from OpenAI and Stripe and was built around product discovery and checkout for general merchandise. The Universal Commerce Protocol was co-developed by Google and Shopify, is backed by a coalition including Walmart, Target, Etsy and Visa, and keeps checkout on the merchant’s own stack. Microsoft Merchant Center now supports UCP feeds in the U.S.

Neither is built for a vehicle transaction. Cars are not SKUs: mixed new and used, prices that move daily, regulated advertising, financing, trade-ins, and a lead that has to land in a CRM as ADF. That is why vehicle-specific agent protocols are starting to appear, and why it will be a long time before an agent completes a car deal end to end.

That is not the part to plan for. The part to plan for is cheaper and closer: the shopper’s agent contacting your store on their behalf. Not a person filling out your lead form, but software asking your store a question and expecting a structured answer. The dealers who will handle that well are the ones whose facts are already machine-readable, because the same discipline serves both.

The one honest caveat

Shopping catalogs are shared by every seller. This is the single surface where your vehicle can sit next to another dealer’s, because that is how a shared index works. Even there, a correctly built feed carries your own link and your seller attribution, so the inquiry routes to you. Your own listing pages and any agent endpoint stay free of competitors. The trade-off is specific to shared catalogs, and being in the index generally beats the alternative.

Why this is hard to keep up yourself

Look back at the table. Four surfaces, four formats, four delivery mechanisms, four refresh cadences, and specs that revise on someone else’s schedule. Google drops your vehicles after three days without a successful feed. ChatGPT accepts updates every fifteen minutes, which is only an advantage if something is actually pushing them. And underneath all of it, every price has to stay identical to your website and your marketplace listings, on a lot where prices move daily.

This is exactly the kind of plumbing that is correct the week it is set up and quietly wrong a month later.

VIN Index takes the one inventory feed you already produce and does the rest: normalizes and VIN-decodes every vehicle, publishes a server-rendered listing page built to be read without JavaScript, generates and maintains the outbound feeds in each surface’s own format with your attribution attached, refreshes them roughly every four hours, and keeps every number consistent across all of it. To get your inventory into that shape, join the waitlist. The free plan costs nothing and takes one feed.