AI can find the part. Can it show the part fits?

“2021 F-150, 3.5 EcoBoost, front pads and rotors, ceramic if you have it, and I need them today.” Every parts counter hears that sentence. AI assistants are starting to hear it too.

By Founder

Published

DIRECTIONA future direction. No present-tense capability claim.

Explore with AI

The industry already solved half of this

The aftermarket has encoded fitment for years. ACES ties a brand's part number and part type to a vehicle configuration, with qualifiers, a note, a position and a quantity. PIES describes the item.

Both are maintained by the Auto Care Association. ACES 5.0 and PIES 8.0 took effect on March 26, 2026. Marketplaces ingest them directly. Walmart's intake requires the brand ID, the manufacturer part number and a part-type ID.

Our read: nobody should rebuild this. The question is what survives the trip from the file to an AI answer.

What gets lost

The note. Free-text notes (“exc. HD brakes”, “w/ 2-piece rotor”) are a known problem. Auto Care's own conference material is about moving them into coded qualifiers to cut “the ‘did you mean’, interpretation of what was written”. A language model paraphrasing a note is exactly that interpretation.

The page. schema.org's isAccessoryOrSparePartFor points from one product to another. It cannot carry a vehicle configuration with qualifiers, position and quantity. If the application list lives only behind a vehicle selector, a crawler sees nothing.

The source. A marketplace fit badge matches against “part compatibility information within the listing”. Which file, which version and which date aren't shown to the buyer or an agent.

Today. Fitment lives in the catalog. Same-day stock lives in hubs and stores. Nothing standard joins the two for an assistant.

Fitment answers need different wording

“This will fit” is a promise only the seller can make. What an assistant can say is “the manufacturer lists this part for the 2021 F-150 3.5L, front, one set, per its ACES file dated September 30, with this condition: ‘Exc. Heavy Duty Payload Package’.” When the applications split on drive or brake package and the buyer hasn't said which, it asks. When nothing is listed, it says so, and doesn't call that “doesn't fit”.

An honest fitment answer names its source, its date and its condition.

Who's building

CarParts.com's assistant is described in a MACH Alliance case study as reachable through an MCP server, with fitment-check and VIN-decode tools. Algolia launched ACES/PIES-aware search in March. Partly is reported to have raised $50 million to build parts-data infrastructure.

The large sellers are covered. The brands and independent sellers whose part numbers the assistants talk about mostly aren't.

What we're building

Lehvel reads a parts business's own ACES and PIES data as sources and returns fitment with its source, its conditions word for word, and its date. It joins that with store availability when there's a current reading, and offers one request a person answers. It also checks what ChatGPT, Claude, Gemini and Perplexity say about the business's part numbers, including on vehicles the parts aren't listed for.

01 · Question
What does an assistant need to answer a parts request honestly: the right part, a fit it can stand behind, and a store that has it?
02 · Why it matters
The aftermarket already encodes fitment in ACES and PIES, and marketplaces ingest those files directly. We traced what survives the trip from the file to an AI answer, and what gets lost, before building on it.
03 · State
DIRECTION A future direction. No present-tense capability claim.
04 · What we built or tested
  • A tested fitment model in Lehvel's repo that reads ACES and PIES data as sources and keeps each condition word for word. No customer's data is loaded.
  • Not live: any agent-facing fitment tool or parts request for a parts business.
05 · Evidence
ACES 5.0 and PIES 8.0 (Auto Care Association)Took effect on March 26, 2026.
Walmart automotive fitment intakeRequires the brand ID, the manufacturer part number and a part-type ID.
Free-text notes (Auto Care conference material)Moving notes into coded qualifiers cuts “the ‘did you mean’, interpretation of what was written”.
schema.org isAccessoryOrSparePartForPoints from one product to another; can't carry a vehicle configuration with qualifiers, position and quantity.
eBay fit badgeMatches against “part compatibility information within the listing”; the file, version and date aren't shown.
CarParts.com (MACH Alliance case study, reported)Assistant described as reachable through an MCP server, with fitment-check and VIN-decode tools.
Partly (Unite.AI, reported)Raised $50 million to build parts-data infrastructure.
06 · What we learned
  • Nobody should rebuild ACES and PIES. The question is what survives the trip from the file to an AI answer.
  • Four things get lost: the free-text note, the application list hidden behind a vehicle selector, the source file and its date, and today's stock.
  • “This will fit” is a promise only the seller can make. An assistant can say what the manufacturer lists, per which file, dated when, with the condition quoted.
  • When applications split on something the buyer hasn't said, the assistant asks. When nothing is listed, it says so, and doesn't call that “doesn't fit”.
  • The large sellers are covered. The brands and independent sellers whose part numbers the assistants talk about mostly aren't.
07 · Limitations
  • The CarParts.com and Partly items are reported, not read at a primary source.
  • No customer's data is loaded, so the model hasn't read a real ACES file.
  • Nothing standard joins catalog fitment to same-day store stock, and we haven't built that join against a real store system.
08 · What changes next
Build it with one parts brand first: fitment with its source, conditions and date, store availability when there's a current reading, one request a person answers, and a check of what ChatGPT, Claude, Gemini and Perplexity say about the brand's part numbers.
Negative results are published as results. A state on this page is a claim about what is true today, not about what is planned; a limitation is a limitation.
← All of LabsNext: Technical product data was never designed for AI agents
These tools belong to the people who made them. Lehvel reviewed and, where noted, tested them. Nothing here is a Lehvel product, and no performance figure attributed to a vendor has been reproduced by us.

Explore with AI