AI searches for the brands it already knows

When an AI assistant searches the web, it doesn't type the engineer's question into a search box. It writes its own queries. Reading them shows how an answer is put together, and why some suppliers keep showing up while others never get a look.

What the queries looked like

In our pilot we asked Claude 52 engineering questions with "Please cite your sources." added, and recorded every search it ran: 158 in all. Of those, 49 (31%) already contained a company or trade name, in 23 of the 52 answers. The model was naming suppliers before it had read anything.

A few examples, verbatim:

• Asked who supplies TPU for footwear, its first search was TPU footwear suppliers BASF Elastollan Infinergy Covestro Lubrizol Huntsman Wanhua, followed by brand-specific searches for each.
• Asked which US contract manufacturers specialize in medical tubing extrusion, it searched medical tubing extrusion companies Zeus Nordson Teleflex Spectrum Plastics.
• Asked what adhesive bonds polypropylene without surface treatment, its second search was 3M Scotch-Weld DP8010 polyolefin bonding no primer, a specific product.
• Asked which plastic survives repeated autoclaving, it searched Radel PPSU steam sterilization cycles hospital disinfectants chemical resistance.

Memory picks the candidates, and search confirms them

The pattern is consistent with our earlier controlled experiments on how answer engines arbitrate between what they memorized and what they retrieve: models search harder for entities they already know. In practice, the shortlist an engineer sees often exists before any page is read. Search then fetches evidence for the candidates the model already had in mind, and the answer cites whatever it found.

Two things follow.

Incumbency compounds. A brand the model learned well in training gets searched by name, found, cited and named again. A brand it doesn't know doesn't get searched for. A newer supplier with a better grade can be absent from the answer without ever having been considered.

What the model finds when it searches for you matters. If a search for your brand surfaces a distributor's stale PDF, a patent, or a teaser page with outdated numbers, that's what gets quoted. The model came looking for you, and it found someone else's description.

What to do about it

If you're the incumbent: make sure that when an engine searches your brand and grade names, it lands on your page, an HTML page with current properties, test conditions, certifications and the current owner's name. Check what it finds today; it may not be yours.

If you're the challenger: you need to be in the material the model learns from and the sources it trusts before you can be searched for by name. Publish the clearest explanation of your material class, and earn coverage in the trade press and technical papers engines cite. When the model does come looking, there should be a page worth finding.

Either way: track which brand names appear in the engines' own search queries for your category. It's an early read on whose names the model already knows.

This pilot used one run per question; query wording varies between runs. The paper sets out the method and limitations.