Do LLMs Read PDFs?

Created by

(opens in a new window)

Ask five people in investor relations whether ChatGPT can read a PDF and you might get five different answers. Some say no. Some say yes, but badly. Some say they have no idea. The confusion is fair, because the honest answer has two parts, and most people only ever hear the first one.

Part one: yes. If your PDF sits at a public URL with no login, no form, and no gated download, an LLM like ChatGPT or Claude can retrieve the PDF and pull text out of it, cite the content for chatbot users, and even share the link. Nothing about the format per se blocks access. The exception matters, though. If your annual report or earnings deck is gated, crawlers won’t fill out forms. Content behind a form is content that does not exist as far as an LLM is concerned.

Part two: the part almost nobody talks about, and it is the part that changes how you should think about your disclosure stack. Being able to read something is not the same as reading it well. Or choosing to read it at all.

It’s also worth noting that accessible PDFs can improve machine readability by adding structure such as headings, reading order, and tagged tables. But even then, there is still limited evidence that they perform as well as semantic HTML in AI search.

The Evidence Is Growing

Researchers from HHL Leipzig Graduate School of Management (PDF, opens in a new window) ran a controlled experiment on exactly this question. They submitted more than 2,500 standardized prompts to ChatGPT about 20 publicly listed companies, covering financial performance, executive compensation, strategy, risk, and sustainability disclosure.

The sample had been structured as a matched pair. Ten companies had published a full HTML annual report. Ten had published as a PDF. Same kinds of companies, same kinds of questions, different format.

Three findings stand out.

  1. Corporate reporting were at the center of answer. Of the 24,662 citations analyzed, 59% pointed directly at the company's own polished annual report. Add the IR website and the corporate site and roughly 85% of citations were to company-owned content. The LLM wants to use the company owned materials. That is genuinely good news, and it undercuts the assumption that LLMs are off somewhere ignoring primary disclosure.
  2. Format decides whether it can be cited or not. Companies with HTML reports were cited three times more often for annual report content than their PDF-publishing peers.
  3. This is the finding with real consequences: when the LLM couldn’t cite the company's own document, it went somewhere else. Answers about companies reporting by PDF contained roughly 2.7 times more third-party citations. The substitutes were largely financial data aggregators: MarketScreener, Bloomberg, Reuters. External summaries of your numbers, with a narrative you did not write, are being used in place of your own numbers. Accuracy also moved with it. In the sample researchers checked for factual correctness.

Why the Format Matters At a Technical Level

None of this is a mystery once you know what a PDF actually is.

A PDF was designed for printing. It stores content as positioned glyphs on a canvas (opens in a new window), not as structured text. There is no inherent concept of a heading, a paragraph, a table row, or a reading order. Text is not even stored in the order it appears on the page. A two-column layout, a footnote, a sidebar, and a table can all end up interleaved into one jumbled string.

Tables are problematic. A table in a PDF is frequently just text positioned near other text, sometimes with lines drawn around it and sometimes not. The relationship between "Revenue" in the left column and "Q3 2026" in the top row is something a human eye reconstructs instantly and an LLM has to guess at.

Charts are worse still. A bar chart in a PDF is a set of colored rectangles. A line chart is a series of path segments. There is no data in there. The LLM is not looking at your revenue climbing up and to the right. It cannot see the shape of the line at all. It can only read whatever text is available on the page.

Remediating a PDF for accessibility can improve this significantly by adding a logical reading order, tagged headings, table structure, alt-text descriptions and other semantic cues. That makes the document easier for machines to interpret, but it still does not guarantee that every LLM or AI bot will use that structure correctly.

An HTML page gives an LLM everything it needs for free. A heading is tagged as a heading. A table has rows and cells. Structure is declared rather than inferred.

Two Audiences, One Piece of Content

This is the shift worth internalizing: Every piece of IR content now serves two readers. A human, who scans for the story. And an LLM, which needs the story to be told in text it can lift, understand, and recite back.

Which reframes something IR teams have done for years without thinking about it. If your best chart is a beautiful image of five-year revenue growth with no caption, no title, and no supporting text, you have published that insight for one audience only. The LLM can’t see it; it reads only the file name, the alt text, and the caption. That is all.

Our research with Brodeur Partners keeps surfacing the same pattern. The same principle applies to official SEC filings. They are authoritative sources, but they are dense regulatory documents rather than content designed for LLMs to easily read and recite. Even when a filing is available in HTML, important facts can be buried across hundreds of paragraphs, tables and disclosures. That is why we recommend pulling the most important investor messages into dedicated HTML pages with clear headings, Q&As and key takeaways.

The fix is not to abandon design. It is to make sure that every claim carried by a visual is also carried by a sentence. A caption that says what the chart shows. A line of body copy that states the number. A data table under the graphic.

What This Actually Means

Nobody should read this as "stop producing PDFs." Filings have to be filed. Investors download decks. Auditors sign documents. The PDF is not going anywhere and it should not.

The argument is narrower and more practical. The PDF should stop being the only place a piece of information lives.

If your earnings release exists only as an attachment, if your CEO letter exists only inside a 100-page report, if your sustainability targets exist only in a chart on page 74, you have made a choice about who gets to explain your company. When the primary source is hard to reach, something else fills the space. The research is clear on what that something is: aggregators, secondary summaries, and third-party interpretation.

There is a small irony in all of this that is hard to miss. The original Digital Investor Relations study on why HTML outperforms PDF was published as a PDF.

Old habits are stubborn. The data is not.

This article was created in collaboration with Brodeur Partners (opens in a new window), a full-service integrated communications agency with public relations at its core. Brodeur Partners (opens in a new window) offers a complete range of positioning, branding, and marketing communications capabilities, including strategic planning, research, media relations, digital and social media, content creation, brand development, and creative execution.