Which technical data sheets get cited in AI answers? We asked 30 spec questions to find out

Cited a manufacturer or brand owner: AI Mode19 of 30

Cited a manufacturer or brand owner: Perplexity14 of 30

Cited a distributor, retailer or rep: AI Mode19 of 30

Cited a distributor, retailer or rep: Perplexity4 of 30

Cited a PDF: AI Mode10 of 30

Cited a PDF: Perplexity9 of 30

Cited an instant-quote platform: AI Mode4 of 30

Cited an instant-quote platform: Perplexity0 of 30

What 30 spec questions cited, Oct. 26, 2026

Item	Value (answers)
Cited a manufacturer or brand owner: AI Mode	19 of 30
Cited a manufacturer or brand owner: Perplexity	14 of 30
Cited a distributor, retailer or rep: AI Mode	19 of 30
Cited a distributor, retailer or rep: Perplexity	4 of 30
Cited a PDF: AI Mode	10 of 30
Cited a PDF: Perplexity	9 of 30
Cited an instant-quote platform: AI Mode	4 of 30
Cited an instant-quote platform: Perplexity	0 of 30

Answers out of 30 in each engine. One run, two engines; ChatGPT and AI Overviews not tested. 79 links with cut-off addresses were removed, so PDF counts are a floor. Source: Our citation log, Google AI Mode and Perplexity, Oct. 26, 2026.

When an engineer asks an AI tool for the tensile strength of 30% glass-filled nylon 6, somebody’s technical data sheet supplies the number. On Oct. 26, 2026, Google AI Mode answered “65 to 195 MPa,” with an average of 140 MPa, and the source it linked first was a compiled data sheet from Professional Plastics, a distributor, summarizing 130 grades. Perplexity, asked the same question that morning, pointed to a BASF Ultramid document instead, so the same question produced two different suppliers of record.

In our test, AI answers cited data sheets, guides and product pages that state the number in plain, readable text, and they did not much care whose domain the document sat on. Several of the documents they used were manufacturers’ files hosted by a distributor or a document-sharing site. For most manufacturers the fix is formatting and housekeeping, which is good news for a marketing team of one.

How we ran the test, and what it cannot tell you

We wrote 30 spec-style questions, seven or eight for each of four industries: Plastics and Rubber, Industrial Machinery and Equipment, Automation and Industrial Tech, and Building Products. They are what an engineer or specifier types when sizing or qualifying something: “IP rating needed for washdown conveyor motor,” “sensing distance of an 18mm inductive proximity sensor on aluminum,” “free area of a 4 inch drainable louver,” “fire rating of 5/8 type X drywall.”

We ran each one once, on Oct. 26, 2026, in Google AI Mode and in Perplexity’s default search, in signed-in accounts in Chrome, with Google placing us in southwest Ohio. For AI Mode we logged the outside links on the answer page, inline and in the sources panel. For Perplexity we logged the first URL behind each inline citation; where a chip bundled extra sources (“+1”), the extras are not in our log.

We did not test ChatGPT, and we did not test the AI Overviews that appear on regular Google results pages. Nothing here describes how either of them cites sources.

AI answers also change from run to run. When we re-ran one question in AI Mode a few minutes later, half of the cited domains were different. Read every count below as one snapshot from one day and one location, good for spotting patterns in what got cited, not for ranking sources or predicting your next result.

One more limit is ours. Our capture tool cut off some long addresses. We recovered 35 of them by finding the full, live address on the site and checking it matched what we captured. We could not recover 79, including 18 YouTube video links and 22 PDF links, so we removed them from the log and from every count. Long document addresses were the ones most often cut, so the counts below undercount PDFs, especially for Perplexity.

The full log is public. Download the dataset (CSV) for every logged citation, 239 links across the 60 answers plus one row for an answer that showed no links, with the question, engine, domain and our classification on each row. The 79 removed rows, with the domain and file name we did capture, are in a separate file, the excluded-rows log (CSV). The log covers one run in two engines; ChatGPT and AI Overviews were not tested.

What the 30 answers cited

	Google AI Mode	Perplexity
Answers with at least one source in our log	29 of 30	28 of 30
Source links in our log	180	59
Answers citing a PDF document	10 of 30	9 of 30
Answers citing a manufacturer or brand owner’s domain	19 of 30	14 of 30
Answers citing a distributor, retailer or rep	19 of 30	4 of 30
Answers citing an instant-quote platform	4 of 30	0 of 30
Links removed because the address was cut off	52	27

The two engines behaved differently. AI Mode linked a wide spread of sources, about six per answer in our log, and leaned on explainer pages, distributor knowledge bases and small suppliers’ blogs. Perplexity cited fewer sources, and a larger share of the links it did give went to primary documents. Counted as a share of each engine’s logged links, rather than by answer as in the table, 34% of Perplexity’s links went to a manufacturer or brand owner (20 of 59) against 20% of AI Mode’s (36 of 180), and 19% were PDFs (11 of 59) against 7% (13 of 180). AI Mode still cited manufacturers and PDFs in more answers, because it cites so many more sources per answer. Perplexity’s primary documents included an ExxonMobil Chemical HDPE guide, a Freudenberg nitrile O-ring material sheet, KSB’s pump lexicon, Rmax on polyiso R-value, Mersen and Intertek on control panel ratings.

On 14 of the 30 questions, the two engines shared at least one cited domain. On the other 16, a reader would have been sent to entirely different sources depending on the tool.

One AI Mode answer had no sources at all. For “hydraulic cylinder bore size for 10 tons at 3000 psi,” it did the arithmetic (2.91 inches minimum, so a standard 3-inch bore) and linked nothing. If your content depends on sizing questions, publish the calculator and the selection logic, because the AI tool will do the math itself.

What makes a technical data sheet citable?

The documents and pages in our log that carried the number an answer used had a few things in common, all visible on the pages themselves. The same habits help a page rank in ordinary search, as our page-one study of manufacturing SEO found.

The number is text, on the page that gets fetched

Greenheck’s louver page, cited by AI Mode for the free-area question, lists each 4-inch drainable model with its free area as plain text, for example “56% free area (4 ft. x 4 ft.)” for the ESD-435, next to the water penetration velocity. Architectural Louvers’ E4DS page, also cited, states “56.0% free area for a 48” wide x 48” high louver” with its first point of water penetration and wind load. Conveyors and Drives’ roller page, cited by AI Mode for roller load capacity (Perplexity cited two of its sister roller pages), says “156 lbs. @ 21” between rails.” Google’s guidance for AI features asks site owners to make “sure that important content is available in textual form,” and the cited pages did that.

The document is machine-readable all the way down

The Professional Plastics sheet behind AI Mode’s nylon figures is plain text: property, range, average, grade count. AutomationDirect’s ProSense M18 selection chart, cited by both engines for the proximity sensor question, gave our text tool the nominal sensing distances (5 mm flush, 8 mm non-flush) but not the material correction table, which the answer needed for aluminum. If your correction factors, derating curves or performance tables are images or complex layouts inside a PDF, expect some systems to read the headline value and miss the part that makes it correct.

It answers one question cleanly

ATAS International’s Tech Tips sheet on wind uplift, cited by AI Mode, explains UL 580, UL 1897 and ASTM E1592 in two pages. Banner Engineering’s Handbook of Photoelectric Sensing, cited by Perplexity for a sensor response-time question, has a section on sensing response time. It is a second edition from 1993, still on Banner’s servers and still being cited in 2026.

It is reachable without a form

Every cited document we opened loaded without registration. We did not test gated documents directly, and we did not find a gated one among the citations we checked. A data sheet behind an email form gives these tools nothing to read.

Page	What the page states	Cited for
Greenheck, conventional stationary louvers (ESD-435)	"56% free area (4 ft. x 4 ft.)" as plain text, next to the water penetration velocity	AI Mode: free area of a 4 inch drainable louver
Architectural Louvers, E4DS	"56.0% free area for a 48" wide x 48" high louver," with first point of water penetration and wind load	AI Mode: the same louver question
Conveyors and Drives, 3610-B05 roller	"156 lbs. @ 21" between rails"	AI Mode: roller load capacity (Perplexity cited two sister roller pages)
AutomationDirect, ProSense M18 selection chart (PDF)	Nominal sensing distances as text (5 mm flush, 8 mm non-flush); the material correction table did not come through as text	Both engines: sensing distance on aluminum

The pages that got cited stated the number in plain text. Where a table did not come through as text, the answer got the headline value and missed the correction. Source: greenheck.com, archlouvers.com, condrives.com and cdn.automationdirect.com, opened Oct. 26, 2026; our citation log.

Your data sheet may be cited on someone else’s website

In our log, AI Mode cited Maxim Integrated’s IO-Link handbook from a copy served by Mouser, a distributor, for the cable-length question. Perplexity’s louver answer cited a 76-page louver selection document that a user had uploaded to Scribd. The Professional Plastics sheet that opened this piece compiles other companies’ grade data on a distributor’s server.

The removed rows show more of the same. They include a Pilz safety relay operating manual on a distributor’s server, a Danaher encoder handbook on another distributor’s site, an RTP Company flame-retardant polypropylene data sheet in another plastics company’s upload folder, a louver submittal drawing in MSC Industrial’s data sheet library and an Evonik VESTAKEEP PEEK document on a reseller’s site. We identified those from the domain and file name but could not capture full addresses, so they are not counted anywhere above.

When this happens, the engineer lands on a copy you do not control. It may be an old revision, and it will not have your current contact details, your distributor locator or your quote form. If your values changed, the AI answer may be quoting the old ones. The fix runs through your distributors’ product data as much as your own site; our guides to e-commerce for manufacturers who sell through distributors and to distributor PIM and product data cover how that content flows.

1. 1

   Your data sheet, current revision

   Published on your site

2. 2

   Copied elsewhere

   A distributor or a document-sharing site hosts its own copy

3. 3

   The AI answer cites the copy

   In our log: Maxim's IO-Link handbook on Mouser; a louver selection guide on Scribd

4. 4

   The engineer lands on the copy

   Possibly an old revision, with no current contact details, distributor locator or quote form

Two fixes: give every data sheet a permanent URL with a revision date in the footer, and ask distributors to link to it instead of rehosting. Source: Our citation log, Oct. 26, 2026.

What did not seem to matter

Brand size did not decide it. Small suppliers’ blog posts showed up next to Siemens and Rockwell Automation. Sites we could not classify, many of them small suppliers’ explainer posts, made up 77 of AI Mode’s 180 logged links.

Age did not decide it either. The 1993 Banner handbook was cited, and so was a 1977 Indian standard on flat belt conveyors, IS 8597, served from law.resource.org and cited by both engines.

Product pages that sell did not lose to neutral content. Conveyors and Drives’ roller page, Greenheck’s product listing and MBCI’s wind uplift article are commercial pages. They were cited because they held the specific value.

In your industry

Plastics and Rubber. Resin questions went to resin makers’ documents (ExxonMobil, RTP) and to distributors’ compiled sheets and product pages (Professional Plastics, Boedeker, Alro). If you are a molder, the material number is not yours to win; publish your processing guidance (shrink, wall thickness, cycle effects) as text. For ABS shrinkage, AI Mode cited a string of mold makers’ and molders’ blog posts.

Industrial Machinery and Equipment. Sizing questions were answered from explainers and from pump and conveyor makers’ reference pages, such as KSB’s NPSH entry. Put ingress ratings, load ratings and speed limits on product pages as text, not only in manuals.

Automation and Industrial Tech. Component makers’ knowledge pages did well (ifm and Keyence on IO-Link, Keyence on safety performance levels, Schneider Electric on sensing distance). So did distributors’ knowledge bases. Standards questions went to standards bodies, UL and test labs such as Intertek.

Building Products. Code questions went to code sites such as UpCodes and to manufacturers whose documents explain the test (MBCI, ATAS, Sika). Louver and panel questions went to product pages with performance values in HTML. Answer specifiers’ questions on the product page and keep the submittal PDF consistent with it.

What to do Monday

1. Pick your ten most-requested data sheets. Open each PDF and try to select the text in every table. Anything you cannot select is invisible to many of these systems; rebuild it as real text.
2. Put the five to eight values engineers screen on (ratings, ranges, capacities, test standards) on the HTML product page as text, with units, matching the PDF exactly.
3. Take forms off data sheets and submittals. Gate the configurator or the CAD library if you must, never the spec sheet. While you are there, check that your robots.txt lets AI search bots in.
4. Search your part numbers and document titles in quotes, plus “pdf.” List every copy hosted by someone else. Ask distributors to link to your live URL instead of rehosting, and put a revision date and your URL in the footer of every document so copies point home.
5. Give every data sheet a permanent URL that survives revisions, and redirect old file names to it.
6. Find your oldest PDFs still online. If the values are wrong, update or retire them with a redirect; a 30-year-old document can still be the cited source.
7. Once a month, ask your three most common spec questions in two AI tools and save the full cited URLs. Our published log is a reasonable template; save complete addresses, which we learned the hard way.