How often does AI cite PDFs?

An identifiable PDF is a cited page whose address ends in `.pdf` or whose observed response type was PDF. Identifiable PDFs supplied 16,075 of 3,954,548 citations, or 0.4065%. That is about one PDF citation for every 246 citations. Google confirms that PDFs can be indexed, but indexability does not show how often the format appears in AI citations.

PDF indexability should not be mistaken for common citation use. Keep important facts in a searchable web page as well as any downloadable document. This result describes the formats we observed and does not show that the file format caused the difference.

A PDF appeared in 13,380 of 398,743 answers, or 3.3555%. At least one PDF appeared for 5,185 of 17,035 prompts, or 30.4373%, across repeated answers in the window.

PDFs are rare within an answer but spread across many tracked questions. Audit PDF accessibility on the questions where formal documents matter instead of treating the format as a universal citation strategy.

Google AI Mode used 13,253 PDF citations among 3,018,395 citations, or 0.4391%. ChatGPT Search used 2,822 of 936,153, or 0.3014%. The per-citation rate was 1.46 times as high on Google AI Mode.

A PDF audit should cover both engines. The combined rate hides a measurable engine difference, but this comparison only describes what we saw and does not identify the cause.

Government publishers supplied 2,549 PDF citations among 30,163 citations, or 8.4508%, and education publishers supplied 1,656 of 40,521, or 4.0868%. Academic publishers reached 2.2484%, documentation publishers 2.0156%, reference publishers 1.7510%, and SaaS publishers 0.2245%.

The format appears most where formal reports, rules, and instructional documents are normal. Compare a site with its publisher category rather than the 0.4065% overall rate.

IEEE Xplore supplied 356 PDF citations across 314 pages and 241 prompts. Contentful-hosted assets supplied 106, the U.S. Department of Energy 94, SSRN 91, and the Federal Aviation Administration 88.

The leading domains are research, government, and document-hosting sources. The count identifies where PDFs already work; it does not show that a hosting domain or publisher type caused the citation.

The California Energy Commission's Pipeline Right-of-Way Monitoring and Notification System led with 72 citations. ArborMetrics' Integrated Vegetation Management on Pipeline Rights-of-Way had 35, the U.S. Department of Energy's Vegetation Management Resilience Investment Guide had 27, Petco's Small Animal Parent Guide had 23, and Marriott's Pet Policy had 22.

The leading files answer narrow operational or policy questions. Use the page list to inspect how a formal document serves a specific buyer question, not as proof that every PDF needs the same treatment.

Of 11,419 identifiable PDF pages, 9,472, or 82.9495%, appeared in one answer. The top 10 pages supplied 268 citations, or 1.6672% of all PDF citations, and the top 10 domains supplied 1,044, or 6.4946%.

There is no small set of PDF files that explains the result. Marketers should audit documents tied to their own questions instead of copying a global leaderboard.

URL endings identified 15,818 PDF citations. Observed response types added 257 citations whose addresses did not end in `.pdf`, for a combined total of 16,075. A response type was available for 1,391,327 of 3,954,548 citations, or 35.1830%.

The 0.4065% result is a conservative identifiable-PDF rate, not a complete count. Files without a `.pdf` address and without an observed response type can be missed. The study counts format, not document quality, accessibility, or whether the file shaped the answer.

PDFs supplied 16,075 of 3,954,548 citations, Google AI Mode used them at a higher rate than ChatGPT Search, and government and education publishers had the highest category rates.

Keep a searchable HTML page for facts you want cited broadly. When a PDF is necessary, give it searchable text, a descriptive title, and a stable address, then link to it from an HTML landing page. Track the page separately by engine and rerun the same study over a window of the same length next quarter.