Why can't AI read PDFs?

PDFs are effectively invisible to AI for two reasons. First, they carry no semantic structure: headings, sections, tables and questions all flatten into a format built for print. Second, AI crawlers favor structured HTML when they pick which sources to cite. Those same crawlers do not execute JavaScript either, so monthly commentary published as a PDF — or on a JS-rendered site — never makes it into the answers ChatGPT, Perplexity, Gemini and Claude give. According to IDC Innovators, 91% of asset manager content is locked inside PDFs.

What does an AI crawler see when it finds a PDF?

A stream of bytes built for print. Text-positioning instructions, embedded fonts, images with no alt text, and tables that dissolve into loose cells with no header and no unit. The hierarchy a human reader takes for granted — title, summary, argument — is not structure at all. It is font size. And without structure, the model cannot pull out a quotable answer with any confidence, so it goes and cites someone else.

The problem is not that the text is unreadable. It is that the text has no address. In well-built HTML, the model knows that this paragraph answers that question, that a named author wrote it, and when. In a PDF, all it has is a block of characters: it knows where they sit on the page and nothing about what they are doing there.

What gets lost in conversion

Source: ReBo PDF extraction analysis, 2026.

This is why "but our PDF has selectable text" settles nothing. Selectable is not the same as structured. The text is there. What is missing is any information about what each passage is for.

Do AI crawlers execute JavaScript?

No. That is the second kind of invisibility, and it lands hardest on the newest websites — the JavaScript-framework builds many firms commissioned in recent years, convinced they were modernizing.

Sources: Vercel analysis of AI crawlers; SearchOptimo/Lantern, 2026.

The distinction between downloading and executing is the whole point. Downloading a JavaScript file yields no content at all — it is code, not text. Only execution would turn it into paragraphs, and execution is exactly what never happens. A React site without server-side rendering hands the crawler a shell wrapped around an empty div.

The difference between CSR, SSR and SSG

• CSR (client-side rendering). The HTML arrives empty and the browser assembles the page. The worst case: invisible to AI crawlers.
• SSR (server-side rendering). The server returns the HTML fully assembled, with the text inside it. Readable.
• SSG (static site generation). Pages are generated at publish time and served as finished files. Readable, and the fastest of the three.

For an asset manager's editorial content, SSG is usually enough and the cheapest to maintain. No monthly commentary needs to be assembled inside the reader's browser.

Why did PDF become the industry standard?

Because it solved the problems the industry actually had. A PDF holds its layout on any device, which matters when the material carries carefully designed charts and tables. It attaches to an email, still the main channel for reaching investors. And compliance likes it: the file is immutable, versionable and archivable exactly as approved.

None of those reasons has stopped being true. Everything around them changed instead. When investors searched Google and clicked through links, the PDF at least showed up in the list. Now they ask an AI and get a synthesized answer built from three to five sources, and the PDF is not even in the running. The format did not get worse. The channel moved.

A real case: the same content, opposite outcomes

In June 2026, ReBo compared one Brazilian asset manager's monthly commentary with that of two firms of comparable size that ChatGPT was already citing on macro questions. The manager's material was excellent: dense fiscal analysis, hard data, primary sources, clearly argued views. It was also published as a PDF sitting in a list of links.

The diagnostic showed it was not losing on one factor. It was losing on all five at once. Here is the before and after.

Source: ReBo analysis, June 2026, of public material from three Brazilian asset managers. Names withheld.

Same content, opposite outcome. It stops being an invisible PDF and becomes a page AI can read, understand and cite. That is the move from Level 1 to Level 3, on the way to becoming Level 4.

The lesson is uncomfortable for anyone who invests in editorial quality. The contest for the three to five citations in each answer is not won by the firm with the best analysis. It is won by the firm with the readable analysis. A boutique with well-structured content beats a large house that keeps everything in PDF.

Which formats can AI actually cite?

Static HTML with a heading hierarchy, an executive summary at the top, headings phrased as questions, tables with headers and units, an FAQ and structured data in JSON-LD. That is precisely what the publishing stage of the ReBo method produces from each monthly commentary.

Source: ReBo AI-friendly checklist, 2026.

Does the PDF have to go?

No. The two formats coexist, because they serve different audiences. The PDF still goes out to investors, still gets attached to the email, still gets archived by compliance and still gets designed with the same care. The HTML version exists for the machines — and for anyone who arrives through search.

In practice, nothing changes for the investment team. The content gets written once, in the format it is already written in, and the conversion happens afterward. What changes is that the thinking now has a permanent address where anyone can read and cite it, instead of living in a file only existing clients ever receive.

One caveat on regulated content. The public version deals in analysis and education, not recommendation. It draws on the firm's market views and general expertise — the material that builds authority — and leaves out specific recommendations, performance figures without disclaimers, and client data.

How do you test whether AI can read your content?

Run three tests. None of them needs a paid tool, all three take about five minutes, and together they cover the most common points of failure: text that only exists after JavaScript runs, a firewall turning the crawler away, and a page no assistant ever surfaces.

Test 1: the page source

Open the page holding your analysis and choose "view page source". Look for a sentence from the middle of the text. If it does not appear in the HTML that comes back, it does not exist for GPTBot, ClaudeBot or PerplexityBot. This test alone separates the readable site from the invisible one.

Test 2: the bot at the door

Run curl -A "GPTBot" https://yoursite.com/your-page. If the response is a 403 or a JavaScript challenge, a WAF or CDN is blocking AI crawlers — almost always through a default setting nobody consciously turned on. One 403 in the log settles the argument faster than any amount of speculation.

Test 3: the real question

Ask ChatGPT, Perplexity and Gemini something your analysis already answers: "What is your firm's view on interest rates for 2026?" Note whether your firm shows up, and who shows up in its place. Repeat across ten to twenty questions and you have an informal baseline for your citation rate.

What does staying in PDF cost you?

The cost never shows up on an expense line, which is exactly why it goes unnoticed. It shows up as absence: in the questions about your market where another firm gets cited, in the screening an allocator does before the first meeting, in the shortlist your name did not make.

The numbers give the urgency its scale. 58% of investors already use AI to research investments, according to Swissquote, and AI-driven searches grew 4.2x in twelve months, according to Evolve Media. Meanwhile, 91% of asset manager content remains in PDF, per IDC Innovators. Demand has moved. The supply of readable content has not.

There is also a compounding effect that makes waiting more expensive than it looks. The models are forming their reference set for this market right now, and sources that get cited tend to accumulate mentions, which reinforce the citation. Getting in early costs less than dislodging a competitor who is already entrenched.

To measure where your firm stands today, the ReBo diagnostic tests the real answers from all five engines and shows you who is occupying your place.