Ask an AI assistant for your price and something counter-intuitive can happen. The model reads the number straight off your brochure, sees it perfectly, and still will not put your name on the answer. A vision model can look at a picture and make out a figure. The step that decides who gets credited is a separate one, and it runs on text the engine retrieved and indexed, not on what the model's eye can see in a graphic.
When the Crane Index™ scored UK websites for AI readability in October 2026, the readable text layer came out strong, averaging 86 out of 100, so where a fact is live text on the page an engine lifts it without trouble. The gap sits elsewhere. Fewer than three sites in a hundred publish what they actually sell as structured data a model can select, a Product, a Service or an Offer. The facts that decide a sale are the ones most often left where the citation step cannot reach them, and a price set inside an image is the hardest case, because there the fact is not plain text. Ranking is no longer proof the fact is working.
Ranking a file is not the same as being quoted from it
Google's Search Central documentation lists PDF among the file types it indexes and can rank, and that documented position is the foundation of the standard SEO view: if the file turns up in results, it is doing its job. That logic held while search returned a list of links and the buyer clicked through to read the document. It breaks the moment the answer, not the link, is what the buyer sees. Ranking means Google will return the file somewhere in a list. Being quoted means an answer engine can lift your fact and hand it back with your name on it. A file can pass the first test and fail the second, and the second is the one that now closes the sale.
Retrieval runs on text, and that is where your name is attached
AI Overviews and AI Mode do not return documents to open. The engine reads what it can retrieve, answers in a sentence or two, and attributes that sentence to the source it lifted it from. It cannot credit a figure it could not retrieve as text, so when your price sits inside an image the engine reaches past you to the nearest page that states the number in plain words.
The obvious objection in 2026 is that multimodal models can see an image directly, so surely they can read the price off the picture. They often can, at the moment they compose an answer. Attribution is a different moment. It runs on what the engine retrieved and indexed as text, not on what a vision model made out in a rendered graphic, so a number that is legible to the model's eye but absent from the retrieved text still earns no citation. There is no indexed source string to attach your name to. The question is not whether a model could read your price if you handed it the picture. It is whether the engine can retrieve that price as text in the first place.
How the pipeline turns a document into text
Google does not index the file on your screen. It stores a text-based version, extracting the words from the document and, where text is missing, applying OCR to the images. Anything that still does not resolve to text is absent from that copy. A tariff typeset into a graphic, a branded price table, a menu laid out as a designed image all look right to a person and return nothing to the index, because those words are pixels, not characters. This is the same line our scan draws: the readable-text score is high because engines lift what resolves to characters, and a fact that stays a picture sits on the wrong side of it.
A fair follow-up is whether AI Overviews even bothers with a file, or prefers the HTML page regardless of its text layer. In practice it leans towards readable HTML, and that reinforces the fix rather than weakening it. Whether the engine skips the file for your page or reads the file's text layer, the deciding factor is the same: the fact has to exist as real text somewhere the engine can read. The page is the safe home either way.
A document is not the problem; a fact trapped in an image is
This is not an argument against PDFs. An answer engine can read and quote a file whose text is real, selectable text. The failure is narrower than the format. The deciding figure is set as artwork: a price typeset into a designed table, a menu rendered as a graphic, a scanned tariff. The document ranks, and the number inside it does not survive as text. If your price is live text you are fine. If it exists only as an image you are exposed, and the structured-data gap in the scan says most sites leave the commercial facts that matter most out of a machine-readable form.
The fix is small, so the real gap is the audit
The remedy is a single edit: put the figure on the page as live text in your CMS, and keep the file as the downloadable brochure for anyone who wants one. So why does a deciding fact still end up trapped in a download? Because ranking hid the failure. The file appeared in results, the conventional scorecard read green, and no one checked whether the number inside it survived as text a machine could repeat.
The check is quick and needs no tools. Open the page a buyer lands on and try to highlight the price with your cursor. If it will not select, it is a picture, not text, and an answer engine cannot retrieve it either. The free Crane Index readability read does the same job across a whole page at once, scoring how much of your content resolves to text a machine can lift. Run either over the facts that decide a sale, the price, the current menu, the key specification, the opening hours, and confirm each one exists as readable text on an ordinary page, not only inside a file and not inside an image in that file.
Will the engines close this gap for you? Attribution may well improve, and models already read images well at the point of answering. The retrieval and citation layer is the slower half, and it still rewards a clean, retrievable text string. The fix costs nothing, takes minutes, and also narrows the structured-data gap the scan measures, so it is worth doing now rather than waiting on a change you do not control. The fact you cannot find in plain text is the one an AI answer hands to someone else.
