What 68 Auto Parts Sites Show AI Shopping Agents

I ran 68 auto parts sites through our agent-readiness checker between 7 and 10 November 2026, then opened their product pages and read the source. The finding that matters most to a parts store is not in the score. The best fitment data I found sat on a store that scores 0 of 9, and the agent files the checker counts said almost nothing about whether an AI agent can tell if a part fits a car.

It is the same checker and the same nine checks as our footwear study: MCP, WebMCP, A2A, ANP, UCP, AP2, OpenAPI, llms.txt and a Wikidata entry tied to the site. For parts I added the question the checker cannot answer: does a machine reading the product page learn which cars the part fits?

Auto parts AI agent readiness in numbers: 68 sites checked

Most of the 68 are parts stores, and eBay is in as the one marketplace. Some are not, and I kept them in because they came up along the way: a tyre shop, accessory shops, a garage equipment store, a tractor parts supplier, a motorcycle parts store, a repair shop, an industrial supplier, a parts maker, a distributor, a brand site and three corporate sites. The table at the end says which is which.

Which auto parts sites the agent readiness checker could read

Before comparing scores I split the sites by what the checker could actually read. A 0 means "publishes nothing for agents" only when the checker got answers from the store.

Of the 37 read in full, 17 publish nothing for agents, 15 have a single point and 5 score 3 or 4. Twelve of the fifteen single points are a Wikidata entry. That point comes from an entry about the company on Wikidata, not from anything published on the store's own site. The other three are an llms.txt the store wrote itself. One of those, AUTODOC's, has no sections, which the checker flagged as broken when it read Sonic Automotive's file two days later, so that point may not survive a rerun.

Shopify gives auto parts stores agent readiness points for free

Every site above 1 of 9 runs on Shopify, and most of what it scores comes from the platform. Shopify serves a UCP profile at /.well-known/ucp on every store and declares an MCP endpoint in it. Its storefronts also load a WebMCP adapter: on two of these stores, tradevehicleparts.co.uk and anjautoparts.com, a WebMCP-enabled Chrome listed 12 tools, from search_catalog and get_product to create_cart and proceed_to_checkout. Our checker counted that adapter from 9 November, so the three Shopify stores checked before then show 3 of 9 and would show 4 on a rerun (inferred from the two stores where we saw it).

So on Shopify a store starts at 3 or 4 without doing anything. On Magento (6 sites), WooCommerce (3) and Salesforce (1) it starts at 0, and every one of those I read in full sat at 0 or 1.

The free points do not reach the product data. On tradevehicleparts.co.uk the store sells a bumper as "Fits Fiat Scudo 2022-2024", while the Shopify catalog entry our checker matched to the same part number tells shoppers "Fitment: Fiat Scudo 2007-2016". An agent working from the catalog gets the other years.

The biggest auto parts retailers score 1 of 9 at most

I ran the biggest names I could list as their own group: AutoZone, O'Reilly, Advance Auto Parts, Pep Boys, NAPA, RockAuto, CarParts.com, Euro Car Parts, Halfords, GSF Car Parts, AUTODOC, Oscaro, kfzteile24, A.T.U, Norauto, Feu Vert, Mister-Auto, Repco, Supercheap Auto, Autobarn, Canadian Tire, Mekonomen and Inter Cars. That is 23 retailers across the US, UK, Europe, Canada and Australia. The list is mine, not a published ranking.

• None scores above 1 of 9. 9 were read in full, 6 in part, 7 refused our checker and CarParts.com never opened from Europe.
• Every point but one is a Wikidata entry, which lives on Wikidata, not on the store's site. The exception is AUTODOC's UK site, whose point is an llms.txt with no sections.
• None runs an MCP server, UCP, WebMCP, A2A or OpenAPI that our checker could find. Three publish an llms.txt (Halfords, AUTODOC UK, Canadian Tire), and two of those have no sections.
• The same company can look different by domain. autodoc.co.uk was read in full and serves an llms.txt. autodoc.de refuses our checker on the homepage and on llms.txt, and its Wikidata entry counts only there, because the entry names the .de domain.

eBay, checked as the one marketplace, scores 1 of 8 readable, also from Wikidata, with no agent files at the root of ebay.com.

Some of them also turn people away by location. NAPA refused me from my own connection and opened through a UK VPN, while our checker, running from a cloud network, read it in full. AUTODOC's German site and A.T.U I also opened through a VPN.

Fitment data and AI agents: where parts stores keep "does it fit"

This is the part a parts buyer cares about, and the nine checks do not touch it. Reading one product page per store, I found three places the fit answer lives.

Fitment in the product markup: three stores

• 1A Auto lists each car as its own isAccessoryOrSparePartFor entry with make, model and year, plus interchange part numbers. The most complete markup in the pilot, from a store that scores 0 of 9.
• AUTODOC marks up 14 vehicles down to engine and power, but they are the cars customers bought the part for, not the fit list. The page itself says the part fits cars the markup leaves out.
• ECS Tuning puts 71 fitment rows in its markup as text values ("Audi B5 A4 Quattro 1.8T"), readable but not a structured vehicle.

Fitment in the page, not the markup

FCP Euro has the fullest fit table I saw, down to engine code and the brake size the car left the factory with, all in the HTML a crawler receives. It scores 0 of 9 and none of that table is in its markup. Winparts.fr lists 53 vehicle rows with engine codes, Pelican Parts gives year ranges per body variant, and exist.ua names 15 makes and about 90 models in the description. A crawler reading the page can find the cars. An agent that trusts structured data only cannot.

Fitment behind a vehicle picker

RockAuto, Summit Racing, Euro Car Parts, Halfords, AutoZone, Advance Auto Parts and Oscaro answer "does it fit" only after the visitor picks a car, by script. The served page names no vehicle, or only a make. Mister-Auto goes furthest: its page text says the part "est compatible avec tous les véhicules de cette liste" (compatible with every vehicle in this list) while the page data holds an empty vehicle list, under a footer notice from TecDoc forbidding copying of its data.

Bot protection decides which AI agents reach auto parts stores

Fifteen sites refused our checker outright, and I wanted to know if that meant agents are refused too. On two Dutch stores that are very likely one operator (same catalogue, same prices, same protection), mijnautoonderdelen.nl and carpartsdirect.nl, I tested it:

1. Our checker, and PowerShell requests from a home connection under both a Chrome and a GPTBot name, all got 403.
2. A real browser opened both stores normally. With the user-agent set to Googlebot, both showed Cloudflare's "verifies you are not a bot" page.
3. ChatGPT in an ordinary chat with search listed live battery prices from the store, to the cent, including a product it was not first asked about.
4. Claude in the chat could not open the page. Claude driving a real browser on my own computer could.

So the protection there passes real browsers and at least one AI company's fetcher, and refuses requests it cannot verify, our checker included. Two other stores went further and blocked me as a person in a normal browser: motointegrator.pl and repco.com.au. Mister-Auto runs DataDome instead of Cloudflare and opened for me while refusing the checker.

For a store owner this means "could not check" in our report is a statement about our checker, not about agents. The record that shows which agents your protection lets in is your own bot management dashboard, read per agent. Turning protection off is the wrong fix, because it also lets in the scrapers and card testers it exists to stop.

Product data errors AI agents read on auto parts pages

Where there was markup, I often found it saying something the page does not:

• Halfords puts the EAN barcode in the mpn field, and Brembo's part number only in the product name.
• Mister-Auto writes its rating and review properties in lower case (aggregaterating, ratingvalue) and its breadcrumb as BreadCrumbList. Schema.org names are case-sensitive, so a strict reader drops them.
• Pelican Parts serves product microdata with an empty name and an empty image.
• anjautoparts.com names the store as the brand of the part, and Shopify localises the price and currency to the visitor.
• avtopro.ua shows three different price ranges for one part across its markup, its page and its meta description.
• Advance Auto Parts has no Offer in its markup at all, so no price or stock reaches an agent from it.

How an auto parts store can raise its AI agent readiness

1. Put the fit list in the markup. One isAccessoryOrSparePartFor entry per vehicle, with make, model, years and engine where it matters. 1A Auto shows it can be done at scale.
2. Serve the fit list in the HTML. A list that loads only after a vehicle pick is invisible to anything that does not pick one.
3. Fix identifiers. The maker's part number in mpn, the barcode in gtin, property names spelled the way schema.org spells them.
4. Check your bot protection per agent. Look up which AI fetchers it passes and which it challenges, and allow the verified ones you want. Do not switch it off.
5. Tie a Wikidata entry to your domain. Several stores have an entry that lists a different domain, so it does not count for the site they sell from.
6. If you publish an llms.txt, give it sections. Two files in this pilot had none and do not meet the format they are named after.

To see where your own store stands on the nine checks, run it through our agent readiness checker. It is free and needs no login.

All 68 auto parts sites and their agent readiness results

Auto parts agent readiness study method and limits

Dates, checker version and what was read

Checks ran from 7 to 10 November 2026. The checker changed twice during the pilot, both on 9 November: it started counting Shopify's WebMCP adapter, and it stopped reading a bot-check page as a store's homepage. Stores checked before those changes are marked in our working notes. Product pages were read by hand, one page per store where a page was read, so a fitment finding describes that page, not the whole catalogue.

What the checker cannot see

The nine checks read public files and a Wikidata entry. They do not measure fitment, price accuracy or which real agents a store's bot protection lets in. Those findings come from reading pages and from the agent tests above, on a handful of stores, so they show what happens, not how often.

What this means for auto parts stores and AI shopping agents

I started this pilot expecting the agent files to be the story. They are close to empty everywhere, from RockAuto to AutoZone, and the stores that score highest are the ones whose platform did the work for them.

What decides whether an agent can sell a part is the fit list, and most of these stores already have one. FCP Euro, winparts.fr and Pelican Parts show it to every visitor. Moving that list into the product markup, one vehicle per entry, is work a parts store can start this month, and no protocol helps an agent that cannot tell whether the part fits the car.

Want to know where your own store stands? Run it through our agent readiness checker, free and without a login, then open one of your product pages and look for the cars in its markup.