Skip to main content
Private access is opening up — request an invite
askFinz
All guidesExplainer · Explainer

Why most search can't read a PDF properly

Slide decks, scans and reports hold some of the most useful information on the web — and most search engines link to them without reading a word.

Sep 6, 2026
Published
5
Sections
1
Figure
4 min
Read

Try searching for a specific fact from a research report, a regulatory filing, or a company's investor deck, and you'll notice something odd: search engines will often find the document — they know it exists, they can show you the filename and the link — but they can't tell you what's actually on page 14. They found it. They didn't read it.

This is one of the most underserved gaps in how search generally works, and it matters more than it looks like it should, because a huge share of genuinely useful information lives in exactly this kind of file rather than in an ordinary web page.

Why documents get treated as an afterthought

Web pages are the easy case for a search system — they're built to be read by machines, with a predictable structure, and there are billions of them, so most indexing effort is naturally aimed there. A PDF, a Word file, or a slide deck is a different kind of object. It doesn't have the same predictable markup a browser can parse cleanly. It might contain columns, tables, footnotes, and images that carry meaning on their own. And some of the time, it's not even text at all — it's a scanned photograph of a page, and there's no text to extract without a further step of actually reading the image.

Faced with that extra difficulty, a lot of systems take the cheap way out: they index the filename, maybe a short snippet from wherever plain text is easiest to grab, and call it "indexed." The result is a document that shows up in search results but can't actually answer a question about its own content — you get pointed at a 40-page report and left to find page 14 yourself, which is exactly the work a search engine was supposed to save you.

What it actually takes to index a document properly

Reading a document well means treating it the way you'd treat a web page — read all the way through, understood, and broken into passages that can each be found on their own — rather than as a link with a filename attached.

Page by page, not as one blob. A 60-page report shouldn't become one giant chunk of text where a question about page 3 and a question about page 55 both retrieve the same undifferentiated mass. Reading a document properly means the specific page — or the specific passage on that page — is what gets matched, so a citation can point at the right spot rather than "somewhere in this file."

Scans included. A scanned page has no underlying text at all — it's an image. Reading it requires recognising the text in the image itself, not just checking whether a text layer exists. Skipping scanned documents means skipping a meaningful share of older reports, archival material, and anything that was never natively digital.

Structure preserved where it matters. Tables, headings, and figures carry information that a naive text extraction can flatten into nonsense — numbers that lose their row and column context, headings that get mashed into the paragraph below them. Reading a document properly means keeping enough of that structure that the extracted content still means what it originally meant.

WHAT IS ON THE PAGEWHAT PLAIN TEXT EXTRACTION RETURNStwo columnsRevenue rose by a The board met onthird in the second 14 May and askedhalf of the year, for a second drafttwo documents, welded together line by linea table across a page breakpage ends hereRegion 2024 2025 North 12 14South 9 11 West 4 7 Total 25 32every row now says the same thing: nothinga scanned pagea photograph of some wordsnot one character — there is no text in the file
All three files open perfectly in a viewer. That is what makes the failure expensive: nothing looks broken until someone searches for a figure that is sitting in a table the extractor flattened.

A traditional search result that just points at a PDF still leaves the reading to you — a minor inconvenience, but not a broken promise, since a human can open the file and skim to the relevant page. An AI answer that's supposed to be grounded in a source has a much bigger problem if the document behind that source was never actually read: the model either can't answer the question at all, or worse, it fills the gap with something plausible-sounding that isn't actually in the document. Properly indexed documents are what let an AI answer cite "page 14 of the report" instead of gesturing vaguely at a filename it never opened.

Where this fits in a real index

In askFinz's web index, non-HTML files aren't filtered out and quietly dropped — they're read the same way a web page is: page by page, with scanned pages read too, rather than skipped for lack of a text layer. That's what lets a question about a specific report or filing get answered from the actual content, with a citation that points at the page it came from, rather than a link to a document nobody actually opened. askFinz's collections group documents like standards, filings, and research alongside ordinary web pages precisely because they're read with the same seriousness, not treated as a lesser category of content.

Further reading

Now onboarding

Put this to work.

See how askFinz fits the way you already work, then ask for access.