What AI Tells Engineers: A Pilot Study of ChatGPT and Claude on Materials and Manufacturing Questions
Abstract. Engineers increasingly start material and process decisions by asking an AI assistant, yet no published study describes what those assistants cite when the question is technical. We asked ChatGPT and Claude 52 engineering questions spanning eight segments, from running-shoe foams to flip-chip underfills, covering material selection, comparison, process selection, troubleshooting, compliance and supplier discovery. We traced the 1,075 sources cited across 156 answers. Asked plainly, Claude searched the web for only 18 of 52 questions (35%) and never for troubleshooting or process questions; everything else came from training data. When answers were sourced, pages from established material suppliers made up 10% of Claude's citations and 31% of ChatGPT's. Patents (13% of Claude's citations), academic papers, trade press, manufacturers' guides and smaller suppliers' content supplied most of the rest. Named suppliers were often unsourced: only 39% of Claude's mentions of established suppliers were backed by the supplier's own pages. For supplier-discovery questions, Claude's lists leaned heavily on market-research report teasers and press-release wires. The two engines rarely agreed: for the same question, they shared on average 12% of cited domains. We describe what this means for how suppliers and manufacturers publish technical information, and the limits of a single-run pilot.
1. Why we ran this
A widening share of material and process decisions now begins in a chat window. In the 2026 State of Marketing to Engineers survey by TREW Marketing and GlobalSpec, 69% of technical buyers said they use generative AI during the purchasing process, up from 58% a year earlier, and 62% of the buying journey happens before an engineer contacts a vendor. The same engineers rate the trustworthiness of AI answers at 4.7 out of 10, and only 6% say an AI summary is usually enough. They use AI to frame the question, then verify.
That makes the sources behind an AI answer commercially important twice over: they shape the first framing, and they are where the engineer lands to check it. Yet the published work on "AI visibility" is built almost entirely on consumer and software questions. We found no study of what AI assistants cite for engineering questions, and no published figure for how often they cite a supplier's own pages, its datasheets, or anything else.
This pilot is a first measurement. It is deliberately small, and Section 6 states what it can't show.
2. Method
Questions. We wrote 52 questions in the way engineers actually ask them, some with a role preface ("I'm a footwear materials engineer…"), most without. Eight segments carry six questions each: footwear materials, semiconductor packaging, plastics extrusion, manufacturing process selection, engineering polymers, adhesives and coatings, sustainability and packaging, and elastomers and composites. Four more are explicit supplier-discovery questions ("Which companies supply…"). By intent: 24 material-selection questions, 10 comparisons, 7 process-selection questions, 5 troubleshooting questions, 5 supplier-discovery questions and 1 compliance question. The full list is in the appendix.
Engines and arms. Every question went to three arms on October 14, 2026, one run each:
- ChatGPT: the consumer product, logged out, with search on, from a US location.
- Claude, as asked: Claude Opus 5.5 through the Anthropic API with its web-search tool available. The model decides whether to search, as it does in the consumer product.
- Claude, sources requested: the same, with "Please cite your sources." appended to the question. Engineers often ask for references, and in this arm the model searched every time.
Sources. For each answer we recorded the URLs the engine cited and the larger set it retrieved but didn't cite. We counted each URL once per answer. Across the three arms, that came to 1,075 cited URLs.
Classification. We assigned each of the 1,097 domains we saw to a source type with model assistance, and reviewed the roughly 110 most-cited domains by hand. For charts, the types fold into seven families:
- supplier pages (material and chemical producers' own sites)
- manufacturers and service providers
- market research and press wires
- patents
- academic papers
- trade press, databases and reference
- everything else
Companies. We extracted the companies and trade names each answer named. Mentions of about 70 established material, chemical and component companies were matched to those companies' own domains to see whether a mention was backed by the company's own page.
3. Results
3.1 Claude answers most engineering questions from memory
Asked plainly, Claude ran a web search on 18 of the 52 questions. The other 34 answers came entirely from what the model learned in training. Search depended on the kind of question. Claude searched on all five supplier-discovery questions and the one compliance question, on 10 of 24 selection questions, on 2 of 10 comparisons, and on none of the troubleshooting or process questions (Fig. 1).
How often Claude searched the web, by question intent
Questions where Claude, asked plainly, ran at least one web search
Table view
| Count | Of | Share | |
|---|---|---|---|
| Supplier discovery | 5 | 5 | 100% |
| Compliance | 1 | 1 | 100% |
| Material selection | 10 | 24 | 42% |
| Comparison | 2 | 10 | 20% |
| Process selection | 0 | 7 | 0% |
| Troubleshooting | 0 | 5 | 0% |
With "Please cite your sources." appended, Claude searched on every question, running 3.0 searches per answer on average and at most 5. The difference between the two Claude arms is the difference between how engineers usually ask and how they ask when they want references.
3.2 Most of what gets cited isn't the supplier's own page
Fig. 2 shows where cited URLs came from in each arm.
Who gets cited when engineers ask AI
Share of cited URLs by source type
Table view
| Supplier pages | Manufacturers & service providers | Market research & press wires | Patents | Academic papers | Trade press, databases & reference | Other | Total | |
|---|---|---|---|---|---|---|---|---|
| ChatGPT | 44% (145) | 20% (67) | 0% (0) | 0% (1) | 21% (70) | 3% (10) | 11% (38) | 331 |
| Claude, sources requested | 20% (121) | 15% (91) | 3% (17) | 13% (77) | 18% (111) | 12% (73) | 20% (120) | 610 |
| Claude, as asked | 20% (27) | 16% (21) | 8% (11) | 5% (7) | 16% (22) | 11% (15) | 23% (31) | 134 |
ChatGPT leaned on supplier pages more than Claude did: 44% of its citations were material or chemical producers' own pages, against 20% for Claude in both arms. Counting only established suppliers (our list of about 70 large material and chemical companies), the shares fall to 31% for ChatGPT and 10% for Claude with sources requested. Half of Claude's sourced answers (26 of 52) cited no established material supplier at all.
The rest is a long tail. Claude's 610 cited URLs came from 342 distinct domains. Three source types stand out:
- Patents. Patents were 13% of Claude's citations, appeared in 62% of its sourced answers, and made uspto.gov its most-cited domain. ChatGPT cited one patent in 52 answers.
- Academic papers. Papers were 18% of Claude's citations and 21% of ChatGPT's, led by ScienceDirect, PubMed Central, ResearchGate and Wiley.
- Smaller suppliers' content. In Claude's sourced answers, 58 of 126 citations of supplier pages went to companies outside our established list, often content written for search ("TPE vs. TPU", "Top 10 TPU manufacturers"). On the midsole-foam question, foam and footwear manufacturers' blogs were retrieved and cited, while only one major polymer supplier had a page cited, by one engine.
Retrieval is not citation. Claude cited 43% of the pages it retrieved with sources requested; ChatGPT cited 29% of the pages it consulted.
3.3 Every segment has its own source mix
Pooling ChatGPT and Claude with sources requested, the mix of sources varies sharply by segment (Fig. 3).
Each segment has its own source mix
Share of cited URLs by source type, ChatGPT and Claude (sources requested) pooled
| Question segment | Supplier | Mfr. & service | Market research | Patents | Academic | Trade & ref. | Other |
|---|---|---|---|---|---|---|---|
| Process selection (n=106) | 16% | 52% | 0% | 7% | 3% | 4% | 19% |
| Engineering polymers (n=92) | 49% | 12% | 0% | 4% | 3% | 8% | 24% |
| Supplier discovery (n=99) | 37% | 30% | 15% | 0% | 0% | 11% | 6% |
| Adhesives & coatings (n=83) | 33% | 7% | 0% | 8% | 17% | 5% | 30% |
| Footwear materials (n=112) | 31% | 3% | 2% | 17% | 28% | 8% | 12% |
| Plastics extrusion (n=111) | 26% | 11% | 0% | 12% | 28% | 16% | 7% |
| Elastomers & composites (n=91) | 23% | 13% | 0% | 10% | 16% | 13% | 24% |
| Sustainability & packaging (n=124) | 19% | 17% | 0% | 3% | 25% | 9% | 27% |
| Semiconductor packaging (n=123) | 25% | 7% | 0% | 12% | 43% | 6% | 7% |
- Process selection: more than half of the citations (52%) went to manufacturers' and service providers' own guides, mostly smaller molders' cost breakdowns and calculators. On "When should I switch from 3D printing to injection molding?", neither engine cited the large online quoting platforms.
- Engineering polymers: supplier pages carried 49%. On the flame-retardant polycarbonate question, both engines cited suppliers' grade-level product pages.
- Semiconductor packaging: academic sources carried 43% and patents another 12%.
- Sustainability: academic sources and standards bodies, regulators and NGOs (grouped under "Other" in Fig. 3) shaped the answers.
A content strategy that works for one segment won't transfer to the next.
3.4 How an AI builds a supplier list
Supplier-discovery questions show the starkest difference between the engines. Asked "Which companies supply underfill and mold compound materials for advanced semiconductor packaging?", Claude drew 9 of its 12 sources from market-research report teasers and press-release wires. ChatGPT cited 17 pages, every one of them a supplier's own (Fig. 4).
“Which companies supply underfill and mold compound materials for advanced semiconductor packaging?”
Sources behind each engine's answer
Table view
| Supplier pages | Market research & press wires | Trade press, databases & reference | Total | |
|---|---|---|---|---|
| Claude | 8% (1) | 75% (9) | 17% (2) | 12 |
| ChatGPT | 100% (17) | 0% (0) | 0% (0) | 17 |
Claude's answer was candid about it: "Most of the sources below are market research reports, which vary in quality. Treat their rankings and market-share numbers as rough estimates." The same pattern, less extreme, appeared for footwear TPU suppliers, where 5 of Claude's 20 sources were report teasers or press wires.
The model's own search queries show the other half of the mechanism. With sources requested, 49 of Claude's 158 searches (31%) contained a company or trade name, in 23 of 52 answers. Asked who supplies TPU for footwear, it searched for "TPU footwear suppliers BASF Elastollan Infinergy Covestro Lubrizol Huntsman Wanhua". The shortlist came from memory, and the search confirmed it.
3.5 Named, but not sourced
Being named is not the same as being the source. When an engine named an established supplier, we checked whether that supplier's own site was among the cited pages (Fig. 5).
When an engine names a supplier, is the supplier's own page behind it?
Share of established-supplier mentions backed by a citation to that supplier's own site
Table view
| Count | Of | Share | |
|---|---|---|---|
| ChatGPT | 43 | 59 | 73% |
| Claude, sources requested | 40 | 103 | 39% |
| Claude, as asked | 11 | 106 | 10% |
ChatGPT backed 73% of such mentions with the supplier's own page. Claude backed 39% when asked for sources and 10% when asked plainly, mostly because it hadn't searched. The rest of the time, someone else's page — a paper, a patent, a distributor, a teaser — described the product.
The spread between companies is wide:
- Best-sourced: Covestro was named in 17 answers, and its own pages were cited in 13 of them.
- Never retrieved: one of the world's largest chemical companies was named in 11 answers, yet its website was never retrieved in any of the 156. Its site is open to AI crawlers, so access doesn't explain the gap.
- Brands without their owner: trade names were often used as shorthand for a whole material class ("TPV, e.g. Santoprene"; "PPSU, e.g. Radel R"). One company that owns several such brands was never named at all.
3.6 The two engines read different internets
For the same question, the sets of domains ChatGPT and Claude cited overlapped by a mean Jaccard index of 0.12. On 11 of 52 questions they shared no cited domain at all. Being cited by one engine says little about the other.
3.7 PDFs and old pages
PDFs and downloads were 14% of ChatGPT's citations and 20% of Claude's, but they were different documents. Half of ChatGPT's PDFs were supplier documents, such as product brochures and datasheets. More than half of Claude's were patents. Engines read PDFs when they can reach them.
Cited pages are often old. Among the cited pages whose age the search tool reported (497 of Claude's citations), the median was 1.1 years, 38% were older than two years, and 20% were older than five. In one 2026 answer about footwear TPU suppliers, ChatGPT cited a supplier's news page from a 2016 trade show.
4. What this means for suppliers and manufacturers
There are two levers, not one. For how-to questions (troubleshooting, processing, comparisons), Claude usually answered from training data. Content shapes those answers only if it was public, crawlable and durable long before the question was asked. For supplier and compliance questions, retrieval decides, and what matters is whether the engine can find, read and quote a specific page.
Own the class explanation. Many selection answers compare material classes before naming anyone. Asked about midsole foams, Claude compared PEBA, eTPU and supercritical EVA in detail and named no brand at all. Suppliers that publish the clearest, most balanced explanation of a class decision are in position when the grade decision comes next.
Publish what papers and patents are currently saying for you. Where engines rely on papers and patents, as in semiconductor packaging and extrusion, the commercial grade is described in research and claims language. Plain-language grade pages with test methods and conditions give engines something better to quote.
Be reachable. In a separate audit of 77 supplier, service-provider, distributor and database websites, almost none blocked AI agents in robots.txt. Delivery was the real problem. Only 3 of 14 flagship grades had a full HTML datasheet on the supplier's own domain; others were delivered through JavaScript, third-party catalogs that bar crawlers, or firewalls.
Measure each engine, in rates. With 12% overlap between engines and answers that vary between runs, single screenshots mislead. Measurement needs repeated runs per engine.
5. Related work
Our earlier controlled experiments, run with the same Assay engine and published on homiere.com, study how answer engines arbitrate between memory and retrieval: When does the model look it up? and Where does a ChatGPT answer come from?. The survey figures in Section 1 are from TREW Marketing and GlobalSpec's 2026 State of Marketing to Engineers report.
6. Limitations
- One run per question per arm. Engines vary from run to run; individual percentages here carry real noise. The launch study will repeat each question three to five times and report spread.
- Two engines. We did not measure Gemini, Google AI Mode, Perplexity or Copilot in this pilot.
- Product surfaces. Claude was measured through its API rather than the claude.ai product, and ChatGPT was logged out. Logged-in and API behavior can differ.
- Scope. One day, US location, English only.
- Classification. Source types were assigned with model assistance and hand-checked only for the most-cited domains. Company matching relied on names and a fixed list of about 70 established suppliers.
- No accuracy grading. We measured who was cited, not whether the answers were technically correct. Grading answers against current datasheets is the next study.
Appendix: the 52 questions
| # | Segment | Intent | Question |
|---|---|---|---|
| 1 | Footwear materials | Comparison | I'm a footwear materials engineer. What foam should I use for a high-rebound running shoe midsole: PEBA, expanded TPU, or supercritical EVA? Compare energy return, density, durability and cost. |
| 2 | Footwear materials | Selection | Which TPU grades are best for a clear, abrasion-resistant shoe outsole that stays flexible below -10°C? |
| 3 | Footwear materials | Process | What's the best way to bond a TPU film upper to an EVA midsole without solvent-based primers? |
| 4 | Footwear materials | Supplier discovery | Who makes bio-based or recycled-content TPU suitable for footwear, and how does its performance compare with virgin TPU? |
| 5 | Footwear materials | Selection | I'm developing a knit shoe upper. What yarn materials give the best combination of stretch recovery, abrasion resistance and recyclability? |
| 6 | Footwear materials | Troubleshooting | How do I reduce compression set in an injection-molded EVA midsole? |
| 7 | Semiconductor packaging | Selection | I'm a packaging engineer working on a large-die flip-chip BGA. Which underfill materials should I consider to minimize warpage and survive 1000 cycles of -55 to 125°C thermal cycling? |
| 8 | Semiconductor packaging | Selection | What epoxy mold compounds are recommended for fan-out wafer-level packaging, and which properties matter most (CTE, modulus, filler loading)? |
| 9 | Semiconductor packaging | Comparison | What are the best thermal interface materials for a high-power AI accelerator package: indium solder TIM, liquid metal, or graphite pads? |
| 10 | Semiconductor packaging | Selection | Which low-Dk/low-Df dielectric materials are used for package substrates in 224 Gbps SerDes and other high-frequency designs? |
| 11 | Semiconductor packaging | Comparison | What die attach materials work for silicon carbide power modules operating at 200°C: sintered silver or solder? |
| 12 | Semiconductor packaging | Process | How do I choose between capillary underfill and no-flow underfill for a fine-pitch copper pillar package? |
| 13 | Plastics extrusion | Selection | I'm a plastics extrusion engineer. What resin should I use for medical tubing that needs to be kink-resistant, clear, and DEHP-free: TPU, TPE, or a PVC alternative? |
| 14 | Plastics extrusion | Troubleshooting | Why is my blown film line showing gels, and how do I tell whether they come from the resin, degradation, or contamination? |
| 15 | Plastics extrusion | Troubleshooting | What causes melt fracture in extruded LLDPE and which processing aids fix it? |
| 16 | Plastics extrusion | Selection | What's the best material for an extruded profile that needs UV resistance and flexibility outdoors for 10+ years? |
| 17 | Plastics extrusion | Selection | Which slip and antiblock additive packages work best for polyolefin film running on high-speed packaging lines? |
| 18 | Plastics extrusion | Troubleshooting | How do I run post-consumer recycled HDPE in extrusion without losing ESCR or getting odor problems? |
| 19 | Process selection | Process | I'm a manufacturing engineer. Should I use injection molding, thermoforming, or CNC machining for a 5,000-unit run of a polycarbonate enclosure? |
| 20 | Process selection | Process | What process should I use to make a lightweight lattice midsole: 3D printing (DLS or MJF), injection molding, or compression molding? |
| 21 | Process selection | Process | What's the best process for overmolding TPE onto a polypropylene handle, and how do I get good adhesion? |
| 22 | Process selection | Process | When does it make sense to switch from 3D printing to injection molding for a plastic part? What volume breakpoints and tooling costs should I expect? |
| 23 | Process selection | Selection | Which 3D printing process and material is best for functional end-use parts that need to survive 120°C and chemical exposure? |
| 24 | Process selection | Process | What's the best way to manufacture a small thin-walled metal part with complex geometry in low volume: metal injection molding, investment casting, or DMLS? |
| 25 | Engineering polymers | Selection | What plastic should I use for a reusable medical device housing that must survive repeated autoclave sterilization and hospital disinfectants? |
| 26 | Engineering polymers | Comparison | PEEK vs PEI (Ultem) vs PPSU: which is best for a high-temperature, chemically resistant structural part? |
| 27 | Engineering polymers | Comparison | Which nylon is better for an under-hood automotive part exposed to coolant at 130°C: PA66 GF30, PPA, or PA46? |
| 28 | Engineering polymers | Selection | What's the best material for a soft-touch overmold on a consumer electronics device that resists sunscreen and skin oils? |
| 29 | Engineering polymers | Comparison | TPE vs silicone vs TPU for a wearable watch band: which has the best comfort, durability and skin compatibility? |
| 30 | Engineering polymers | Compliance | Which flame-retardant polycarbonate or PC/ABS grades meet UL 94 V-0 at thin walls for an electronics enclosure? |
| 31 | Adhesives & coatings | Selection | What adhesive should I use to bond polypropylene or other low-surface-energy plastics without surface treatment? |
| 32 | Adhesives & coatings | Selection | What are the best PFAS-free alternatives for durable water repellent (DWR) finishes on outdoor apparel? |
| 33 | Adhesives & coatings | Selection | Which structural adhesive is best for bonding carbon fiber composite to aluminum in an automotive application? |
| 34 | Adhesives & coatings | Comparison | What conformal coating should I use for a PCB exposed to condensation and salt spray: acrylic, silicone, urethane, or parylene? |
| 35 | Adhesives & coatings | Troubleshooting | How do I improve paint adhesion on TPO automotive parts? |
| 36 | Adhesives & coatings | Selection | What hot-melt adhesive works best for a high-speed packaging line running freezer-grade cartons? |
| 37 | Sustainability & packaging | Selection | I'm a packaging engineer. What mono-material alternatives exist to replace multilayer PET/PE laminates in recyclable flexible packaging? |
| 38 | Sustainability & packaging | Selection | What bio-based polymers can replace polycarbonate or ABS in consumer products without a big performance loss? |
| 39 | Sustainability & packaging | Comparison | How does chemically recycled PET compare with mechanically recycled PET for food-contact packaging? |
| 40 | Sustainability & packaging | Selection | What are the best alternatives to PVC for flexible applications as regulations tighten? |
| 41 | Sustainability & packaging | Selection | Which materials can replace PFAS-based fluoropolymers in seals and gaskets? |
| 42 | Sustainability & packaging | Selection | Which compostable or biodegradable polymers (PLA, PHA, PBAT) actually work for injection-molded parts? |
| 43 | Elastomers & composites | Comparison | What elastomer should I use for a seal exposed to biodiesel at high temperature: FKM, HNBR, or FFKM? |
| 44 | Elastomers & composites | Selection | What polyurethane system should I choose for a cast urethane wheel that needs high load capacity and low rolling resistance? |
| 45 | Elastomers & composites | Selection | Which carbon fiber prepreg systems are best for out-of-autoclave curing in sporting goods? |
| 46 | Elastomers & composites | Comparison | What's the best foam for protective padding that must absorb repeated impacts: EPP, EVA, polyurethane, or rate-dependent materials like D3O? |
| 47 | Elastomers & composites | Selection | Which thermoplastic elastomer types (TPV, SEBS, TPC-ET, TPU) are best for a soft gripping surface that must resist oils? |
| 48 | Elastomers & composites | Selection | What material should I use for a living hinge that needs to survive a million flex cycles? |
| 49 | Supplier discovery | Supplier discovery | Who are the leading suppliers of thermoplastic polyurethane (TPU) for footwear, and what are their flagship grades? |
| 50 | Supplier discovery | Supplier discovery | Which companies supply underfill and mold compound materials for advanced semiconductor packaging? |
| 51 | Supplier discovery | Supplier discovery | Which contract manufacturers in the US specialize in medical-grade tubing extrusion? |
| 52 | Supplier discovery | Supplier discovery | What are the best online services for getting custom injection-molded parts quoted quickly? |