Why AI gets your product specs wrong, and how to fix it

AI tools often quote the wrong fire, thermal or certification figures for building products. Here is how manufacturers can improve AI product data accuracy by keeping one master record, publishing specs as HTML and auditing third-party listings.
Arabella Cronin
November 16, 2026
Professional construction worker installing drywall with gloves and helmet inside a building.

A specifier asks an AI tool for a fire-rated board suitable for a corridor wall. The answer comes back confident and clear. The Euroclass is wrong, and it is attributed to your product. Nobody at your end sees it happen. This is an AI product data accuracy problem, and it is landing on manufacturers' doors now.

The risk is not just embarrassment. A wrong fire classification, U-value or certification can lead to a rejected submittal, a specification query or a lost enquiry. In the worst case, someone relies on it.

Here is why it happens, and what you can do about it.

Why AI gets product specs wrong

AI tools do not look up your product the way a specifier does. They draw on whatever text they have seen or can retrieve about it, then predict the most likely answer. If that text is inconsistent, old or hard to read, the answer will be too.

Technical products make this worse. Fire classifications, thermal performance values, acoustic ratings and certifications all depend on exact test conditions. A figure for one board thickness, one substrate or one system build-up gets quoted as if it applies to the whole range. Humans spot the caveat. A language model often drops it.

The tool also has no way of knowing which source is the official one. A current datasheet and a three-year-old distributor listing look much the same to it.

Where the conflicting figures come from

When we look at how manufacturers appear online, the same problems come up again and again.

  • Specs locked in PDFs. Datasheets, technical brochures and test reports often hold the only full set of figures. PDFs are harder for AI tools to read reliably. Tables get flattened, footnotes get separated from the values they qualify, and scanned documents may not be read at all.
  • Different figures in different places. The website says one thing. The latest datasheet says another. The distributor's product page says a third. Each one is "correct" at a different point in time.
  • Outdated datasheets still in circulation. Old versions stay indexed on your own site, on distributor sites and in downloads shared by the trade. Nothing tells a machine which version is current.
  • Third-party listings that drift. NBS entries, merchant catalogues, BIM libraries and trade directories are often filled in once and rarely revisited. Changes to a product or its certification do not always reach them.
  • Missing structured data. Without clear markup, a product page is just a block of text. Nothing tells a machine that "30" is a fire resistance period in minutes, or that "0.021" is a thermal conductivity in W/mK.

None of these are unusual. Most manufacturers have all five. The result is a mess of conflicting figures, and AI tools pick from it.

How to improve AI product data accuracy

You cannot edit what an AI tool says. You can make the correct answer the easiest one to find. That is the practical core of any attempt to increase visibility in AI answers, and of how to optimise content for tools like these. Accuracy comes first, visibility second.

1. Keep one set of figures per product

Decide where the master record lives. It might be a product information spreadsheet, a PIM or a controlled document. Every other place should draw from it.

For each product, agree and record:

  • fire classification, with the standard and test reference
  • thermal values, with the thickness or build-up they apply to
  • acoustic performance, if relevant
  • certifications (BBA, CE or UKCA marking, third-party approvals), with certificate numbers and expiry dates
  • the date the figures were last checked

Then name one person who owns it. Technical and marketing teams both need to sign off changes. Most inconsistencies start when they do not talk to each other.

2. Put specs on the page as HTML

Do not rely on a PDF to carry your technical data. Put the key specifications on the product page as real, readable text, ideally in a properly marked-up table.

Include the conditions with the figure. "Class A2-s1, d0 when tested to EN 13501-1 at 12.5 mm" is far more useful, to a person and a machine, than "A2 fire rating".

Keep the PDF for downloads and full test reports. But link to it from a page that already states the headline figures. This is the sort of structure we build into website design projects for manufacturers: technical data as part of the page, not something hidden behind a download.

Add structured data too. Product schema, with specification properties where they apply, helps search engines and AI tools interpret the page correctly. A Webflow build makes it straightforward to template this across a whole range, so it stays consistent as you add products.

3. Make third-party listings match

Audit where your products appear outside your own site:

  • NBS and other specification platforms
  • distributor and merchant websites
  • BIM object libraries
  • trade directories and association listings
  • retailer and marketplace pages, where relevant

Compare each against your master record. Correct what is wrong and request updates where you do not control the listing. Send distributors a clean, current data pack and ask them to replace the old one. Withdraw or clearly mark superseded datasheets.

This is dull work. It is also where a lot of wrong answers originate.

4. Check regularly

Set a routine, not a one-off project.

  • Every quarter, ask several AI tools about your key products, using the questions a specifier would ask. Record the answers.
  • Compare the output to your master record. Note any errors and trace them to a source.
  • Re-check after every certification renewal, product change or datasheet update.
  • Keep a simple log. Patterns show up quickly.

Often the wrong figure turns out to come from a single outdated listing. Fix the source and the answer tends to improve over time.

What you cannot control

It is worth being honest here. AI output cannot be fully controlled. The same question can give different answers on different days. Tools update, retrieval changes, and some will still get things wrong however tidy your data is.

So treat this as risk reduction, not a guarantee. Clean, consistent, well-structured data gives AI tools the best chance of getting it right. It also helps everyone else: specifiers, estimators and merchants who read your pages and trust what they find.

It is also sensible to make your own pages the clear source of truth. A short note on technical pages, such as "Always check the current datasheet and certificate before specifying", protects you if an AI tool, or a person, gets it wrong.

Start with the figures

Most of this work has little to do with AI itself. It is about having one reliable version of your technical information and making it easy to find, read and reuse. That is good practice for the people who specify your products, and it happens to be what AI tools need as well.

If you are not sure where your product data conflicts, or how your product pages hold up, get in touch. We can help you audit what is out there, restructure specifications on your website and set up a process that keeps the figures right.

Arabella Cronin
November 16, 2026