Almost nothing AI engines cite carries a file suffix. In the Machine Relations Index answer ledger, 187,650 of 201,266 cited-URL events (93.2 percent) were addresses with no file suffix at all. PDFs were 2,762 events (1.4 percent) and Markdown files 352 (0.17 percent). Only three events were an llms.txt file.
The rare file types are not spread evenly. A PDF was 8.1 percent of the cited-URL events on deep-tech hardware questions and 0.23 percent of those on AI-visibility questions. Markdown was concentrated on Perplexity: 322 of the 352 Markdown events came from Perplexity, and two thirds of all of them were GitHub or Hugging Face repository pages, not a publisher's own page served as Markdown.
This page counts cited-URL events. It does not show that an engine reads a PDF or a Markdown file, or that publishing one earns a citation: the ledger does not record what an engine fetched, only the address it listed.
What was counted #
- Population. Successful runs in the Index's buying-question panel from 2026-07-06 to 2026-11-09 on its 1,008 buying questions: 168 in each of six question shapes (best X, top lists, how to choose, is it worth it, problem first, X vs Y). Runs flagged as naming a brand in the prompt are excluded. The ledger export is dated 2026-11-10 08:10 UTC.
- Engines. ChatGPT, Claude, Gemini, Google AI Mode and Perplexity. Google AI Overviews are left out because this analysis was not run on them.
- Unit. One cited URL in one answer run, counted once however many times the run repeated it. The same URL cited in two runs counts twice. The set holds 201,266 such events from 23,729 runs that cited at least one address, 79,206 distinct URLs and 29,587 root domains.
- A file type. The suffix on the last segment of the URL path (the part after the final dot), lowercased, with the query string and fragment removed. No page was fetched or opened, so a suffix is a fact about the address and not about what the engine received.
- The questions are the Index's own. They were written to cover buying decisions in 24 verticals. They are not a sample of what buyers type.
Ninety-three percent of cited-URL events carry no file suffix #
Cited-URL events by the suffix on the address (share of all 201,266; 95 percent interval from resampling the 1,008 questions where one was computed):
| What the address ends in | Events | Share | 95% interval |
|---|---|---|---|
| No suffix | 187,650 | 93.24% | not computed |
| .html or .htm | 7,841 | 3.90% | 3.63 to 4.17 |
| 2,762 | 1.37% | 1.17 to 1.58 | |
| .php, .aspx, .asp, .shtml, .cfm, .cgi, .jsp, .cms | 967 | 0.48% | not computed |
| .md, .mdx or .markdown | 352 | 0.17% | 0.13 to 0.24 |
| .json, .xml, .csv or an Office file | 17 | 0.01% | not computed |
| .txt (3 of them llms.txt) | 3 | 0.0015% | not computed |
| Any other suffix, including numbers and domain names inside a name | 1,674 | 0.83% | not computed |
The no-suffix row is the ordinary web page: a slug with nothing after it. A PDF is about 1.4 of every hundred cited-URL events, and 42 of every hundred cited PDFs sit in two of the 24 verticals (below). Counted by distinct address, the 2,762 PDF events are 1,559 different PDFs, and 84 of them were cited five or more times, so the share is not one document repeated.
Perplexity and Gemini cite PDFs; Claude and ChatGPT almost do not #
Share of each engine's cited URLs that were PDFs or Markdown files, and the share of runs that cited at least one PDF (events matching, out of events; runs out of runs that cited anything):
| Engine | Cited URLs | PDF events | PDF share | Markdown events | Runs with a PDF |
|---|---|---|---|---|---|
| Perplexity | 107,152 | 2,174 | 2.03% | 322 | 1,088 of 7,964 (13.7%) |
| Gemini | 17,580 | 313 | 1.78% | 11 | 164 of 1,409 (11.6%) |
| ChatGPT | 36,286 | 174 | 0.48% | 8 | 156 of 7,346 (2.1%) |
| Google AI Mode | 6,178 | 19 | 0.31% | 2 | 19 of 883 (2.2%) |
| Claude | 34,070 | 82 | 0.24% | 9 | 74 of 6,127 (1.2%) |
| All five | 201,266 | 2,762 | 1.37% | 352 | 1,501 of 23,729 (6.3%) |
Perplexity's PDF share is more than eight times Claude's. Gemini cited nothing in most runs (see how many sources each engine cites), so its per-run figure rests on 1,409 runs, and one run in nine of those held a PDF. The PDFs are not one publisher's habit: 917 root domains supplied them, and the largest, ti.com, supplied 49 of the 2,762 (1.8 percent).
PDF share differs thirty-five-fold across verticals #
Share of cited URLs that were PDFs, for selected verticals of the 24, including the highest and the lowest (events matching, out of events):
| Vertical | Cited URLs | PDF events | PDF share |
|---|---|---|---|
| Deep-tech hardware | 9,938 | 803 | 8.08% |
| Review-sparse industrial | 6,333 | 361 | 5.70% |
| Logistics and freight | 5,379 | 93 | 1.73% |
| Martech and advertising | 4,236 | 29 | 0.68% |
| Consumer finance | 11,850 | 50 | 0.42% |
| Sales, GTM and CRM | 3,459 | 12 | 0.35% |
| HR and talent | 8,626 | 29 | 0.34% |
| Emergent prosumer | 9,688 | 29 | 0.30% |
| AI visibility and GEO | 13,237 | 30 | 0.23% |
Deep-tech hardware and review-sparse industrial questions produced 8.1 percent of all cited URLs (16,271) and 42.1 percent of all cited PDFs (1,164). The most-cited PDF domains in the hardware vertical are manufacturers, agencies and trade bodies: ti.com (49 events), energy.gov (27), nvidia.com (19), cern.ch (18), semiconductors.org (17) and nxp.com (16). The spread between the top and bottom vertical is thirty-five to one.
By question shape the spread is narrower. A PDF is 1.9 percent of the addresses cited on problem-first questions (505 of 26,691) and 1.7 percent on how-to-choose (554 of 32,031), against 0.8 percent on best-X questions (307 of 39,663). Descriptions of a situation reach PDFs more than twice as often as best-X questions do.
Across all 2,762 PDF events, 290 (10.5 percent) sat on government domains (.gov and .mil), 212 (7.7 percent) on university domains, and 81 (2.9 percent) on arxiv.org (29), usenix.org (18), aclanthology.org (19), ieee.org (14) and nature.com (1). The other 2,179 (78.9 percent) were on 763 other domains, led by company sites such as ti.com, bigcommerce.com and nvidia.com. That split uses a domain-name rule, not a review of each document.
Markdown: concentrated on Perplexity, and mostly a repository #
Of 352 Markdown events from 209 distinct addresses on 29 root domains, the engines split as follows (events): Perplexity 322, Gemini 11, Claude 9, ChatGPT 8, Google AI Mode 2. Perplexity's Markdown rate is 0.30 percent of its cited URLs. The other four engines combined, 0.03 percent.
Two thirds of the Markdown events are not a publisher serving its own page as Markdown:
| What the Markdown address is | Events | Distinct addresses | Root domains |
|---|---|---|---|
| A file inside a GitHub repository | 190 | 117 | 1 |
| A file in a Hugging Face dataset or repository | 47 | 27 | 1 |
| A publisher's own page or documentation served as .md | 115 | 65 | 27 |
| All Markdown | 352 | 209 | 29 |
The 115 publisher events are the case that matters to anyone serving a Markdown twin of their pages. They are real: AWS documentation (27 events), milvus.io (21), nvidia.com (14, twelve of them one product page), agiletech.vn (10), adyen.com (5). They came from 105 runs, and 97 of the 115 events (84 percent) were Perplexity's. For 28 of the 65 publisher addresses, the same page without the .md was also cited in some run in the window, on any engine; for the other 37, only the Markdown form was cited (an exact-address match, which would miss a locale or www variant).
The Markdown events are concentrated in technical questions. They came from 91 of the 1,008 questions, and ten questions hold 148 of the 352 events (42 percent). Among the publisher events, AI infrastructure has 44 and AI security and privacy has 25 of the 115. In AI visibility and GEO the ledger holds 3 Markdown events in 13,237 (two of them publishers), and in professional services 1 in 2,699.
Markdown events by month, all five engines (the Perplexity API has returned 10 sources for nearly every answer since 2026-10-07; see what changed in Perplexity's collection):
| Month | Cited URLs | Markdown events | GitHub or Hugging Face | Publisher's own |
|---|---|---|---|---|
| July (from the 6th) | 34,087 | 28 (0.08%) | 13 | 15 |
| August | 24,008 | 8 (0.03%) | 8 | 0 |
| September | 36,591 | 23 (0.06%) | 17 | 6 |
| October | 88,494 | 209 (0.24%) | 137 | 72 |
| November (1 to 9) | 18,086 | 84 (0.46%) | 62 | 22 |
The Markdown share rose from under 0.1 percent in July to September to 0.46 percent in the first nine days of November. The rise starts in the month the Perplexity collection changed and its cited-URL volume more than doubled. This page did not test whether a change in the questions sampled explains part of it.
llms.txt: three events #
An llms.txt file was cited three times in 201,266 events, all by Perplexity: customer.io/llms.txt, docs.perplexity.ai/llms.txt and mastra.ai/docs/rag/retrieval/llms.txt. No other engine cited one. That is a count of a file listed as a source. It does not say whether engines read llms.txt files while answering, which this ledger cannot see.
What this changes for how pages are published #
- In marketing-side verticals the citation record for PDFs is thin. In AI visibility and GEO, 30 of 13,237 cited-URL events were PDFs; in HR and talent, 29 of 8,626. This is a count of listed addresses. It does not measure how many download-only documents exist in those verticals or what happened to them.
- In hardware and industrial questions PDFs are a real share of citations. There, 5.7 to 8.1 percent of cited-URL events are PDFs, from manufacturers, government bodies, research labs and trade bodies. A PDF published there sits in a category where PDFs are cited; state what is inside on a web page that links to it.
- A Markdown twin is access for agents, not a proven citation route. Of 201,266 cited-URL events, 115 were a publisher's own Markdown page; 97 of those 115 (84 percent) came from Perplexity and 69 (60 percent) answered AI infrastructure or AI security questions. Serve the twin for agents that ask for it; do not expect it to be what an engine lists.
- Three events in 201,266 is not evidence for or against llms.txt. The ledger cannot see whether engines read the file.
- Check the engine before the format. PDF and Markdown citations are common on Perplexity (and PDFs on Gemini) and nearly absent on Claude and ChatGPT.
Machine Relations serves a Markdown copy of every research page. In this window machinerelations.ai appears in 7 of the 201,266 events and authoritytech.io in 58, and none of those addresses was a .md, .pdf or llms.txt file. Those counts are too small to say whether a twin of ours would or would not be cited.
What this page does not show #
- Cause. A PDF or Markdown address may be cited because of its content or its publisher, not its format. No page was read.
- Reading. The ledger records the address an engine listed, not the file it fetched. An engine may read a Markdown or llms.txt copy and cite the HTML page. This page cannot see that.
- Preference. The counts compare citations. They do not control for how many PDFs and Markdown files exist for each question.
- Buyer demand. The 1,008 questions are the Index's own panel, not a sample of what buyers type.
- Consumer apps and Google AI Overviews. Not included. The five engines are API routes.
- A trend. The monthly Markdown rise coincides with a collection change on the engine that supplies most Markdown events. This page did not separate the two.
- Suffix as format. A PDF served at an address with no .pdf suffix, or a page that ends in .pdf but serves HTML, is counted by its address. The "any other suffix" row includes numbers and domain names inside a name, such as version strings.
How the address is read for other signals is in the URL slug study, the site sections study and the year-in-the-URL study.
Method note #
Source: the Index answer ledger's citation-events table, daily export of 2026-11-10 08:10 UTC, read with DuckDB and joined to the run table for the question and its shape. Filters: runs on the buying-question panel (mri_v2) that succeeded, run dates 2026-07-06 to 2026-11-09, events not flagged as brand-prompted, a non-empty cited URL, engines ChatGPT, Claude, Gemini, Google AI Mode and Perplexity. Unit: distinct (answer run, cited URL) pairs. The path is the URL lowercased with scheme, host, query string and fragment removed. A suffix is the one to eight letters or digits after the final dot of the last path segment. PDF is .pdf; Markdown is .md, .mdx or .markdown; llms.txt is a path ending in llms.txt or llms-full.txt; HTML is .html or .htm. A publisher's own Markdown page is a Markdown address on any root domain other than github.com and huggingface.co. The sibling check strips .md, .mdx, .markdown and a trailing /index, and a trailing slash, and looks for an identical lowercased address cited in any run in the window. Government domains end in .gov, .gov.* or .mil; university domains end in .edu, .edu.* or contain .ac.; the research list is arxiv.org, aclanthology.org, usenix.org, nature.com, ieee.org and acm.org (acm.org had no PDF events). Intervals resample the 1,008 questions, 1,000 draws, fixed seed, with every cited URL of a question kept together. Engine, vertical, shape and monthly rows report pooled counts and no interval. Run shares use the runs that cited at least one address (23,729) as the denominator.
Frequently asked questions #
Do AI engines cite PDFs? #
Rarely, and unevenly. In the Machine Relations Index, 1.4 percent of cited URLs (2,762 of 201,266) were PDFs: 2.0 percent on Perplexity, 1.8 percent on Gemini and 0.24 percent on Claude. On hardware questions they were 8.1 percent of cited-URL events, and on AI-visibility questions 0.23 percent.
Do AI engines cite Markdown files or markdown versions of web pages? #
Seldom. Markdown addresses were 0.17 percent of cited URLs (352 of 201,266), and 322 of them came from Perplexity. Two thirds were files in GitHub or Hugging Face repositories. A publisher's own Markdown page was cited 115 times, most of them (60 percent) on AI infrastructure and AI security questions, and 84 percent of them on Perplexity.
Do AI engines cite llms.txt files? #
Three times in 201,266 cited-URL events, all on Perplexity. The ledger records the address an engine listed, not whether the engine read an llms.txt file to answer.
Should I publish a PDF or a web page to be cited by AI? #
The ledger cannot say which earns a citation; it shows where each was listed. PDFs were under 1 percent of cited-URL events in AI visibility (0.23), sales and CRM (0.35), HR and talent (0.34) and martech (0.68), and 8.1 percent in deep-tech hardware and 5.7 percent in review-sparse industrial. It does not show that the format caused any citation.