Blog · Diagnosis

Content locked in images and PDFs

Published 14 September 2026 · Updated 12 October 2026 · 9 min read

Content locked in images and PDFs
Photograph: Markus Spiske / unsplash

The short answer: A sentence rendered inside a PNG or a scanned PDF is a picture to a retrieval system, not a sentence. Retrieval indexes are built from text, and none of the engines document running optical character recognition over the images they crawl. Anything that would need OCR may never be recovered at all, and where it is, the structure around it is gone. Your pricing, your spec table and your best explanation can be on the page in full view of a human and absent from the index the answer was written from.

This is the cheapest class of AI visibility problem to find, and one of the most common in the audits we run. It is also the one people argue about, because assistants can read an image you paste into a chat window. That is a different code path from the one that builds the index.

Two different machines, one confusing demo

Paste a screenshot of your price list into a chat and the model reads it. Ask the same assistant, in a fresh session, what your product costs, and it will go through retrieval: a query, a set of candidate documents pulled from an index or a live fetch, a chunking step, and a handful of passages kept. Nothing published about that loop says vision runs over every image on every candidate page, and none of the engines document the step either way. The part you can see, the text an extractor pulls from a page, is what reaches the index.

So the demo proves the model can see. It proves nothing about whether the crawler that filled the index ever turned your diagram into words. Those are separate systems with separate budgets, and the retrieval one is the one that decides whether your name appears. If you have read why AI never mentions your brand, this is the most mechanical version of that failure.

What travels with an image, and what does not

Three strings can accompany a picture into a text index: the alt attribute, the caption or nearby prose, and the file name. Google's image SEO documentation, checked on 12 October 2026, calls alt text the most important attribute for providing metadata about an image, and lists the formats Google Search supports in an img source at all: BMP, GIF, JPEG, PNG, WebP, SVG and AVIF. That is Google describing Google. No other engine publishes a comparable list, so do not read it as a rule the category follows. Everything else about the picture, including every word drawn inside it, is pixels to a text extractor.

Now add chunking. Retrieval pipelines cut a page into passages before they index it, and none of the engines publish the size they use, so treat the few hundred words usually quoted as an illustration rather than a specification. If your figure sits at a boundary, the caption lands in one chunk and the surrounding explanation lands in another, and the passage that gets retrieved has neither. A table rendered as an image is the worst case of this. Where OCR does run, plain extraction returns a flat run of characters with no column headers bound to the numbers, so "2,400" arrives with nothing telling the engine it was the annual price.

The scale of the plain version of this is measurable. The HTTP Archive's 2024 Web Almanac, accessibility chapter, reports 15 percent of desktop images with no alt attribute set at all and 30 percent with an empty one, and 69 percent of images passing the Lighthouse audit for images with alt text. Those figures come from the Almanac's 2024 crawl of home pages, not from any sample of ours, and they say nothing about the inner pages where pricing and specifications usually live. They are accessibility numbers. They are also retrieval numbers, because the same string is doing both jobs.

There is a standards answer here that predates all of this. WCAG success criterion 1.4.5, Images of Text, at level AA, asks you to use real text wherever the technology can produce the same visual result, with logos and essential presentation as the exceptions. Its stated goal is that users can adjust how text is presented, and it shipped with WCAG 2.0 in 2008. It is load-bearing for engines in 2026 as a side effect, not by design.

PDFs fail in a different way

A PDF is not one thing. A born-digital export from a word processor carries a text layer and extraction mostly works. A scan, or a slide deck flattened to images, carries none, and without OCR there is nothing to extract. Google lists PDF among the file types it indexes, and that page, checked on 12 October 2026, does not mention optical character recognition at all. That is an absence in one engine's documentation, not proof that the step is missing everywhere. It is still the gap worth assuming is real until you have evidence otherwise.

Tagging is the second trap. An untagged PDF has no declared reading order, so a two-column layout can extract as interleaved half-sentences: column one line one, column two line one, and so on down the page. The words are all present. The meaning is shredded.

Then the smaller indignities. A PDF has no standard place to state a publication date on the page the way an HTML article does, and plenty carry no visible date at all. We have not counted how many, so take that as an impression from the client files we open rather than a rate. Where the date is missing, nothing tells an engine whether the file is current, which is a real problem for anyone thinking about how often to refresh content for AI search. A PDF has no title element either. The nearest equivalent is the Title field in the document metadata, which word processors fill with the source file name, which is how a listing ends up reading "Microsoft Word - whitepaper_final_v3.docx". Google's title link documentation does not say how that field is used, so we cannot tell you which engines read it. And a PDF has no place to put a robots meta tag, so the only control surface is the HTTP response header.

That last point has a sharp edge. Google's robots documentation, checked on 12 October 2026, says X-Robots-Tag is how you control non-HTML resources such as PDF files, and that nosnippet also stops the content being used as a direct input for AI Overviews and AI Mode. That is Google's documented behaviour for Google's surfaces. What other engines do with the same header is not documented anywhere we can cite. We have found blanket nosnippet rules applied to .pdf at the server level on client sites, inherited from an old legal instruction nobody had reread. Check yours before you spend a quarter rewriting the files themselves.

Four checks, about twenty minutes

Run these against your own site before you believe any of the above applies to you.

Extract the PDF and count. pdftotext -layout yourfile.pdf - | wc -w. A twelve-page document returning a handful of words almost certainly has no text layer. Read the output too, not just the count, and look for the interleaving that signals missing tags.

Try to select the text. Open the PDF, drag across a paragraph, press copy. If nothing highlights, no extractor is getting it either.

Fetch the HTML without a browser. curl -sL yoursite.com/pricing | textutil -stdin -format html -convert txt -stdout. What you see is roughly what arrives at a text index when nothing renders the page first. If the price is not in that output, it is not in the text a crawler that does not render would get. Some crawlers do render, so this is the worst case rather than every case.

Check the headers on a PDF. curl -sI yoursite.com/file.pdf | grep -i robots. Anything containing nosnippet or noindex is the whole explanation.

The self-check walks through ten of these end to end, including the ones that have nothing to do with file formats.

What to move, and what to leave

Move anything a buyer would quote at you: prices, specifications, compatibility matrices, eligibility rules, turnaround times, supported regions. These belong in HTML, as real text, in a table or a list. Keep the image as the pretty version if you like, beside the text rather than instead of it.

Leave photographs, brand assets and genuine diagrams as images, and write the alt attribute as a sentence that states what the diagram shows, not a keyword. A chart of two lines crossing in March deserves "monthly support tickets fell below self-serve resolutions in March 2026", not "chart".

For PDFs that must stay PDFs, publish an HTML summary page beside each one, with the key numbers as text and a link to the file. That page is what gets retrieved, dated and cited. The same instinct makes FAQ content that is not filler work: put the answer in retrievable prose, close to the question it answers.

When you are ready to be systematic about it, the file-format pass is one row in a self-audit checklist for AI visibility.

Honest limits

We cannot tell you which engines run OCR over indexed images, or how often, because none of them document it, and an undocumented step is not the same as a missing one. Our working assumption, that a fact which exists only inside a figure rarely surfaces in an answer, comes from running audit prompt sets against client pages rather than from a controlled study, and we have no frozen prompt set, sample size or date to attach to it, so take it as a prior and not a measurement. Each sourced engine claim here rests on Google's own documentation, which describes Google; for ChatGPT, Perplexity, Gemini, Copilot and Grok we have no comparable published statement about image text, PDF extraction or nosnippet handling. The alt attribute figures are the Web Almanac's 2024 crawl of home pages, so they do not describe the deeper pages where the locked content usually sits. And we cannot tell you what a multimodal crawl will do in a year: the direction of travel favours engines getting better at pictures, and the fix costs a day and helps screen readers regardless.

FAQ

Can ChatGPT read text inside an image on my website? It can read an image you upload into a conversation. Whether the pipeline that indexed your site extracted that text is a separate question, which none of the engines document. In our own audit work the fact inside the image usually fails to surface, which is an observation from client pages rather than a measured rate. Do not reason from the chat window to the crawler.

Should I delete my PDFs? No. Publish an HTML page carrying the same key facts, dated and linked to the file. The PDF stays for people who want to print or circulate it.

Does alt text help AI answers or only accessibility? Both, and the mechanism is identical. Alt text is one of the few strings about an image that reaches a text index, so a descriptive sentence is worth more than a keyword.

How do I know if my PDF is tagged? Extract it with a command line tool and read the first page of output. Scrambled order across a multi-column layout, or headings that arrive in the middle of body text, points at missing tags. Most PDF editors will also report tag status in document properties.

Sources

Related reading

Why AI never mentions your brandA self-audit checklist for AI visibilityFAQ content that is not fillerHow often to refresh content for AI searchInformation architecture for answers, not sessionsInternational and multilingual AI visibility