We asked ChatGPT and Claude 52 engineering questions. Here's who they cited.

Engineers have started asking AI the questions they used to bring to a datasheet, a distributor or a supplier's application engineer. Which TPU stays flexible below −10 °C? Why is my blown film showing gels? Injection molding or thermoforming for 5,000 enclosures? In the 2026 State of Marketing to Engineers survey, 69% of technical buyers said they use generative AI during purchasing.

So we asked the obvious next question: when an engineer asks, what does the AI actually cite?

What we did

We wrote 52 questions the way engineers ask them, across eight segments: footwear materials, semiconductor packaging, plastics extrusion, process selection, engineering polymers, adhesives and coatings, sustainability and packaging, and elastomers and composites. Four more asked outright which companies supply something.

On October 14 we put every question to ChatGPT and to Claude, and traced every source behind the 156 answers. Claude ran twice: once as asked, and once with "Please cite your sources." added, as engineers often do.

Five things we found

1. Claude mostly answers from memory. Asked plainly, Claude searched the web for only 18 of the 52 questions. For troubleshooting and process questions it never searched at all. Those answers came entirely from what the model learned in training, so the only content that could have shaped them is content that was public and readable long before anyone asked.

2. Supplier pages are a minority of what gets cited. Pages on material and chemical producers' own sites made up 44% of ChatGPT's citations and 20% of Claude's. Counting only established suppliers, the shares fall to 31% and 10%. Half of Claude's sourced answers cited no established supplier at all.

3. Patents and papers do a lot of the talking. Patents were 13% of everything Claude cited when asked for sources, and uspto.gov was its single most-cited domain. Academic papers were 18% of Claude's citations and 21% of ChatGPT's. When an engine describes your chemistry, it may be quoting your patent's claims or a decade-old paper rather than the grade you sell today.

4. Named isn't sourced. When Claude named an established supplier in a sourced answer, the supplier's own page backed the mention only 39% of the time. ChatGPT did better, at 73%. The rest of the time, someone else's page supplied the description.

5. The two engines read different internets. For the same question, ChatGPT and Claude shared on average 12% of their cited domains, and on 11 questions they shared none. A page that wins on one engine may not exist for the other.

What it means

For a materials supplier or a manufacturer, the uncomfortable conclusion is that most of what an engineer reads about your products in an AI answer isn't written by you. It comes from patents, papers, distributors' copies of your datasheets, trade press, and competitors' and smaller suppliers' blogs.

The practical conclusions:

• Two levers. For how-to questions, what the model learned in training matters most. For "who makes it" questions, retrieval matters most. Both depend on public, readable, durable pages.
• Segments differ. Process questions were won by manufacturers' own cost guides; packaging questions by papers and patents; polymer questions by suppliers' grade pages. One playbook won't fit every segment.
• Measure each engine, in rates. Twelve percent overlap and run-to-run variation make a single screenshot meaningless.

This was a pilot: one run per question per engine, one day, US location. The paper sets out the method, the limitations and the full question list.