What AI search actually reads on your site
A growing share of the questions people used to type into a search box are now asked of an assistant, and answered without anybody opening a website at all. Whether your business appears in that answer, and whether the answer is correct, comes down to something unglamorous: how much of your site could be read as text.
There is a lot of noise about this. Most of the practical advice reduces to things that were already good practice and are now simply less optional.
The plain mechanics
A model answering a question about your business is working from some combination of what it absorbed in training, what a retrieval system fetched from your pages just now, and what other sites say about you. In all three cases the input is text and structure. Not layout, not design, not anything that only exists once JavaScript has run.
Which puts a very old list back at the top of the priorities.
What tends to be unreadable
- Prices inside images. A price list exported as a JPEG is invisible. So is a menu, a tariff, a size guide. This is extremely common and completely self-inflicted.
- Content that only exists after a click. Accordions are fine if the text is in the HTML and merely hidden. They are not fine if the answer is fetched when you expand it. FAQ sections fail this constantly, and FAQs are exactly what gets quoted back.
- Critical facts only in the header or footer image. Opening hours, service areas and phone numbers set as graphics because they looked better that way.
- Forms with no accessible names. If a booking form's fields have no labels, nothing parsing the page can tell what the form does — the same failure that makes it unusable with a screen reader. One defect, three separate consequences.
What tends to get quoted
Passages that answer one specific question, completely, in one place, in plain sentences.
This is a real change in emphasis. Content written to rank was often written to be long — accumulating keywords and internal links, burying the answer in the middle of the ninth paragraph. Content that gets used in a generated answer needs the opposite shape: the question as a heading, the answer immediately underneath, self-contained enough to survive being lifted out of context.
"Do you cover Broughty Ferry?" as an H2, followed by "Yes — we cover Dundee and everywhere within 25 miles, including Broughty Ferry, Monifieth and Carnoustie. There is no additional call-out charge inside that radius." That is a paragraph that can be repeated back to a customer correctly, by anything, without a link.
Structured data still earns its keep
Schema.org markup for your organisation, your products, your opening hours and your location is not a ranking trick. It is the difference between a machine inferring your address from a footer and being told it unambiguously. It is cheap, it is stable, and it is one of the few things in this area that has not changed in a decade.
The uncomfortable bit
You now have less control over the sentence a customer reads about your business than at any point since you bought the domain. You cannot design that sentence. You can only make sure the facts it will be assembled from are present, current, and machine-readable on your own site rather than being reconstructed from a directory listing that has your old phone number.
Start there. Check what your site says about your hours, your prices and your service area — as text — and check that it is right. Most of what passes for AI search optimisation is that, done properly.