How to publish part numbers and cross-references so AI assistants can read them
One page per part, the data printed as text, a cross-reference table that says how close each match is, and a way to check what the assistants see.
- 11 min read
Short answer
AI assistants can only repeat part data they can read as text. Give every part its own page at a stable address, linked with ordinary links, with the part number, maker, specifications, and cross-references printed in the page's HTML. Data that sits only inside a PDF or behind a search box may never be read. Add Product structured data from the same database, then check the result in Google, Bing, and the assistants themselves.
Why an assistant cannot find your part numbers
A buyer who used to type a part number into Google now also asks ChatGPT, Copilot, Perplexity, or Google's AI Mode: who stocks this, what replaces it, what are its dimensions. The assistant answers from pages it can read. If your catalogue is not among them, it answers from someone else's pages, or from memory, which is where wrong part data comes from.
Most industrial catalogues are hard to read for reasons that have nothing to do with AI. They were built for a person who already knows where to click:
- The catalogue is a PDF. Thousands of parts share one address, and the cross-reference list sits on page 212.
- Parts appear only after a search. There is a search box and a results list, but no page a link can point to.
- The page is assembled in the browser. The address exists, but the part data arrives by script after the page loads.
- Prices and data sit behind a login. A crawler cannot sign in, so everything behind the login does not exist for it.
- Part numbers are written one way only. The page says 6205-2RS and the buyer types 62052RS.
None of this needs an AI project to fix. It is publishing work on the website and the product data you already have.
How assistants read a catalogue
The assistants do not share one index. Each reads the web its own way, and what their makers have documented is short. The table separates what is documented from what has only been observed.
| Assistant | Reads from | What that means for a catalogue |
|---|---|---|
| Google AI Overviews and AI Mode | Google's search index | Google states there are no extra technical requirements: a page has to be indexed and allowed to show a snippet. A part page missing from Google is missing here too. |
| Microsoft Copilot | Bing's index | Check your part pages in Bing Webmaster Tools as well as in Google Search Console. |
| ChatGPT, Claude, Perplexity | Their own crawlers, plus search results | None of the three documents whether its crawler runs JavaScript. A study by Vercel in December 2024 observed that they did not. Treat text in the page's HTML as the only safe place for part data. |
The practical rule that follows covers all of them: one address per part, reachable by an ordinary link, with the data present in the HTML the server sends.
Give every part its own page
- One stable address per part number, such as /parts/6205-2rs. A search result, or an address that changes with the session, does not count.
- Reachable by ordinary links. Google states that it can generally crawl a link only when it is a standard link element with an address in it. A part that can be reached only by typing into a search box has no link pointing at it. Category pages that list their parts as links solve this.
- No data after a # in the address. Google generally does not use that part of an address for crawling or indexing, so a part shown at /catalogue#6205 is not a page.
- Listed in the sitemap, generated from the same database, so new parts are found without waiting for a crawl of the whole site.
- Open to crawlers. Check robots.txt and your CDN or firewall settings: some block AI crawlers by default.
- Variants on their own pages when buyers ask for them by number. If 2RS and ZZ are ordered as different parts, they are different pages that link to each other.
What to print on each part page
Write the page so that a person who lands on it from nowhere, with no menu and no context, can tell what the part is and whether it is the one they need. A crawler arrives the same way.
- The part number as you write it, and as buyers type it. Print the official form in the title and heading, and list the common variants in the text: without dashes, without spaces, with the old prefix.
- The maker and the brand, in words. A logo image does not say who made the part.
- A one-sentence description in the buyer's vocabulary: what the part is, its size, and what it is used for.
- Specifications in a table, with units. Bore, outside diameter, width, load rating, material, temperature range: one row per value, the unit in the cell.
- Cross-references, in the table described in the next section.
- What it fits or replaces: the machine, the assembly, the superseded number.
- How to get it: a link to request a quote, and stock or lead time only if you can keep it true.
- The datasheet as a download, alongside the page's own text.
Publish cross-references as a table that says how close the match is
A cross-reference is a claim: this number can stand in for that one. Buyers ask assistants for exactly that claim, and an assistant that finds two part numbers near each other on a page may report them as interchangeable whether you meant it or not. So say how close each match is, in the same row as the number.
| Their part number | Maker | Match | Note |
|---|---|---|---|
| A-6205-2S | Maker A | Direct replacement | Same dimensions, seal type, and clearance |
| B 6205 DD | Maker B | Same dimensions | Different seal material: check the operating temperature |
| C-6205-ZZ | Maker C | Alternative | Metal shields instead of contact seals: not for wet or dusty use |
| 6205-RS (old number) | Own range | Superseded by this part | Replaced in 2019; same fit |
- Use a real table, one row per number. The number, the maker, and the note stay together when the row is read or copied as text. A sentence listing ten numbers loses which note belongs to which.
- Name the kind of match with a small fixed vocabulary, such as direct replacement, same dimensions, alternative, superseded. Use the same words on every page.
- Publish only matches someone has checked. In many product families the base number is common across makers while each maker's suffixes are its own, and two suffixes that look equivalent can stand for different seals or clearances.
- Link in both directions where you stock both parts, so each page confirms the other.
- Write the competitor's number as text. Buyers search by the number they have, which is often not yours.
Keep the PDF, but stop relying on it
Google does index PDF files, and has said it may use text recognition on scanned ones. That is the most any provider has documented: none of the AI crawler documentation says what happens to a PDF. Even where a PDF is read, a 300-page catalogue is one address for thousands of parts. An assistant cannot send a buyer to the row they asked about, and you cannot see in your analytics which part they wanted.
The catalogue PDF and the datasheets stay as downloads, because engineers want them. The data inside them goes on the part pages as text. If the PDF is exported from the same database as the pages, the two never disagree.
Add Product structured data, and know what it does not do
Structured data is a block of code in the page that restates its facts in a fixed vocabulary, schema.org. For a part page the type is Product, with the fields mpn (manufacturer part number), sku (your own number), gtin (the barcode number, if the part has one), brand, and additionalProperty for each specification with its unit.
{
"@context": "https://schema.org",
"@type": "Product",
"name": "6205-2RS deep groove ball bearing, 25 x 52 x 15 mm, sealed",
"mpn": "6205-2RS",
"sku": "YOUR-SKU-10482",
"brand": { "@type": "Brand", "name": "Your Brand" },
"url": "https://www.example.com/parts/6205-2rs",
"additionalProperty": [
{ "@type": "PropertyValue", "name": "Bore diameter", "value": 25, "unitCode": "MMT" },
{ "@type": "PropertyValue", "name": "Outside diameter", "value": 52, "unitCode": "MMT" },
{ "@type": "PropertyValue", "name": "Width", "value": 15, "unitCode": "MMT" }
],
"isSimilarTo": [
{
"@type": "Product",
"mpn": "A-6205-2S",
"brand": { "@type": "Brand", "name": "Maker A" }
}
]
}What is known about its effect, stated plainly:
- Google says no special structured data is needed to appear in AI Overviews or AI Mode. Microsoft says schema helps search engines and AI systems understand content. OpenAI, Anthropic, and Perplexity have published nothing on whether they use it.
- It will not bring rich results to a quote-only catalogue. Google's shopping listings require a price above zero, and its product snippets require a review, a rating, or an offer.
- It must match the visible page. Google's policy is not to mark up content that readers cannot see. Structured data restates the page.
- schema.org has no field for an interchangeable part. The closest, isSimilarTo, means a functionally similar product. It cannot carry "direct replacement" against "alternative", which is why the visible table does that job.
It is still worth adding, for one reason: when part pages are generated from a database, the structured data is one more template filled from the same fields. It costs little, it is the format search engines document, and it forces the product data to be clean. Do it after the visible page is right.
Shopping feeds and agent protocols: not yet for quote-based sellers
OpenAI and Google have both published ways for a seller to hand product data to an assistant directly. As of November 2026, OpenAI's product feed requires a price for every item, and Google's Universal Commerce Protocol is built around checkout and needs Google's approval before going live. Neither describes a seller whose price depends on the customer and the quantity.
For a company that sells by quote, there is nothing to join yet. The preparation is the same work as the rest of this guide: one record per part with an identifier, a maker, specifications, and an address. Whatever format arrives for business buyers will ask for those fields.
How to check what assistants can read
- View the page source (the browser's inspector shows the page after its scripts have run) and search it for the part number and one specification. If they are missing from the source, they arrive by script, and some crawlers will not see them.
- Search Google and Bing for one of your part numbers in quotation marks, together with your company name. If your part page does not appear, it is not indexed.
- Check a sample of part pages in Google Search Console and Bing Webmaster Tools: indexed, not excluded, no crawl errors.
- Ask the assistants. Take twenty part numbers and ask ChatGPT, Copilot, Perplexity, and Google's AI Mode who supplies each one, what its dimensions are, and what replaces it. Record whether the answer is correct, and which sites it cites.
- Read your server logs for the assistants' crawlers, which identify themselves by name. Which part pages they request tells you what they have found.
- Repeat monthly with the same twenty numbers. The answers change as the indexes update.
The order of work
- Clean the product data first: one record per part, one spelling per maker, units in their own field. Every later step reads from it.
- Run the check above on twenty parts, so there is a baseline.
- Publish part pages from the database, starting with the product lines buyers ask about most.
- Add the cross-reference tables, checked matches only, with the kind of match stated.
- Add category pages, the sitemap, and structured data.
- Run the check again a month later, and compare.
Part pages are website work: a catalogue where every part has its own page, its data as text, and a request form beside it. Once buyers can find the part, the next gap is the request itself: see how to automate RFQ intake. To learn where your catalogue stands today, send us its address and a few part numbers. We will tell you what a crawler can read and what it misses.
Sources
- Google: AI features and your website
- Google: make your links crawlable
- Google: managing crawling of faceted navigation URLs
- Google: file types indexable by Google
- Google: PDFs in Google search results (2011)
- Google: product structured data
- Google: structured data policies
- schema.org: Product
- Microsoft: optimizing your content for inclusion in AI search answers (October 2025)
- Vercel: the rise of the AI crawler (December 2024)
- OpenAI: product feed specification
- Google: Universal Commerce Protocol overview