Menus, Photos, and PDFs: What Machines Can and Can't Read on Your Site

Here's a quiet tragedy playing out on thousands of small business websites. The owner paid for a nice design. The information is all there: the menu, the price list, the service details, the hours. And the machines that now answer customers' questions can't reliably read any of it, because it's locked inside images and PDFs.

The result is a business that looks complete to humans and half-empty to the systems doing an increasing share of the recommending. It's worth twenty minutes to understand what machines actually parse, because the fix is usually cheap.

The hierarchy of readability

At the top: plain text on a normal web page. HTML text is effortless for any crawler or engine. Every fact stated this way is fully available for answers.

In the middle: PDFs. Engines can often extract text from them, but inconsistently. Multi-column layouts scramble, scanned documents may read as blank pictures, and PDF content tends to rank as second-class. A fact that lives only in a PDF is a fact the machine might have.

At the bottom: text inside images. The stylized menu JPEG, the price list your designer exported as a graphic, the photo of the sandwich board with this week's hours. AI vision has improved, and some engines can read some image text some of the time. But "some, some, some" is a terrible foundation for the facts your business depends on. Treat image-locked text as invisible and you'll make better decisions.

Where this bites hardest

Restaurants are the classic case. The menu is the single most-asked-about document, "does this place have gluten-free options," "how much is dinner for two," and it's overwhelmingly published as a PDF or a photo. An engine asked those questions about your restaurant either finds a competitor whose menu it can read, or answers from a third-party site whose copy of your menu is two years stale.

But it's everywhere. Salons with service lists as Instagram screenshots. Contractors with project galleries and not a sentence of describable text. Gyms with schedules as weekly graphics. In each case, the customer-facing question the business gets asked most often is exactly the information machines can't retrieve.

The fix: publish facts twice

You don't have to abandon the pretty version. Keep the designed menu, the branded graphic, the downloadable PDF. Just make sure every fact in them also exists as plain page text. A simple HTML menu page alongside the PDF. A text service list under the screenshots. Hours typed into the page, not just photographed on the door.

While you're at it, use the supporting rails: descriptive alt text on images, real headings on sections, and structured data for the facts that have schema types, like menus, prices, and hours. Each layer makes extraction more certain.

Keep the text version current

One warning from the trenches: businesses that maintain a pretty version and a text version sometimes update only the pretty one. Now the machine-readable copy is the wrong copy, which is worse than none, because engines will confidently repeat your outdated prices. Whatever your update routine is, the text version belongs in it. Same fact, same day, both places.

A good test for any page: select the text with your cursor. If you can't highlight a fact, a machine probably can't lift it. Walk your key pages with that test once, fix what fails, and you've removed one of the most common and least visible reasons good businesses go missing from AI answers. The information was never the problem. The format was.