We Checked Whether Manufacturer Sites Are Actually AI-Ready
We recently published AI-Ready Manufacturer Guidelines — practical, standards-based recommendations for making a catalog readable by AI agents instead of just humans. Guidelines are easy to write in the abstract. So instead of leaving it there, we pulled five real product pages from real companies and checked them against our own list. Here's exactly what we found.
The method
We fetched the raw HTML of five product pages — no rendering, no login, just what a plain request actually returns, which is close to what a lot of AI retrieval looks like in practice — and checked each for schema.org Product structured data, whether that data included real engineering specs (not just name/price/SKU), and whether the page was even reachable at all.
| Site | Product JSON-LD? | Real structured specs? | Notes |
|---|---|---|---|
| SparkFun (RedBoard Qwiic) | Yes | No | Name, price, SKU, image only — zero specs in the markup. |
| Pololu (Micro Metal Gearmotor) | Yes | Sort of, but broken | Real numbers present, dumped as garbled text inside description. |
| Prusa (MK4S printer) | No | No | Zero structured data on a flagship product from a major manufacturer. |
| McMaster-Carr (bearings) | No | No | Couldn't even reach a stable individual product URL — real pages sit behind JS interaction. |
| Digi-Key (resistor) | — | — | Blocked the request outright (HTTP 403) before any check could run. |
The most interesting result wasn't the sites with nothing — it was Pololu
SparkFun and Prusa having incomplete or missing structured data is unsurprising. The more useful finding was Pololu, because they clearly tried. Their product markup does include a Product type, and the real spec numbers — voltage, RPM, stall torque — are genuinely present in the data. But they're dumped into the free-text description field as a garbled fragment of what used to be an HTML table: "12 V \n\t\t330 RPM, 80 mA \n\t\t1.3 kg⋅cm...". A human skimming the rendered page reads this fine. An agent parsing the structured data has to regex-guess its way through mangled whitespace to recover a single number — which is precisely the gap described in why AI agents will do the shopping: the information exists, but only in a form built for a person.
What this says about the actual size of the problem
None of the five sites we checked — including ones that clearly invested some effort into structured markup — expose real engineering specs in a form an agent can reliably use without extra parsing work. That's a stronger, more concrete version of the claim we made in compliance has to become machine-readable: the gap isn't hypothetical, and it isn't limited to small or under-resourced companies. It shows up even where a team has clearly already invested in basic SEO structured data.
What we'd actually recommend, concretely
The fix for Pololu's specific case is small and doesn't require rebuilding anything: keep the Product type, but move each spec out of the free-text description and into its own additionalProperty entry — {"@type": "PropertyValue", "name": "voltage", "value": "12V"} and so on. That's the exact mechanism schema.org already provides for this, described in more detail in our AI-Ready Manufacturer Guidelines. It's a markup change, not a redesign.
For McMaster-Carr and Digi-Key, the constraint is different — it's reachability, not markup quality — but the two aren't actually in the same position. Digi-Key's developer API is genuinely open and self-serve: anyone can register at developer.digikey.com and get OAuth credentials, no gatekeeping beyond signing up. That's the route Xibrary is already taking for electronics data through Nexar. McMaster-Carr does have a documented API too, but it's restricted to "approved customers" through a registration and vetting process, not something a developer can just sign up for. Combined with a site that's hard to reach through plain fetching in the first place, McMaster-Carr's catalog is meaningfully harder to get structured access to than Digi-Key's — we're not going to recommend fighting bot detection to get around that, but it's also not a simple "just use the API instead" the way it is for Digi-Key.
We'll keep spot-checking sites like this as we go — real evidence is a lot more useful than assuming the gap exists.