Opening hours are the fact AI gets wrong most often
In the reports we have run so far, the error that turns up most is opening hours. Not prices, not services, not the address: hours. A restaurant that opens at 13:00 is described as opening at noon. A clinic that closes on Fridays is recommended for a Friday appointment. A shop that switched to summer hours in June is still on winter hours in August.
Trace it back and the story is nearly always the same. The model is not inventing anything. It is repeating a directory listing that was right once and has never been touched since. Those directories syndicate to each other, so one stale entry can turn up on a dozen sites — and to a model reading the web, a dozen sites saying the same thing looks a lot like consensus.
Your own website usually has the right hours. But if they are embedded in an image, loaded by a script, or written as "see Google for hours", the model has nothing to read there, and the directory wins.
What to do
Put your hours in plain HTML text on your contact page and in LocalBusiness structured data, and keep the two identical.
Use the cited-source section of your report to find the directories that models actually rely on for your business, and fix those first. It is rarely more than three.
Set a calendar reminder for every seasonal hours change. The models will not notice until the sources do.
NOTE
Mention rate is not the same across models, and that's useful
A pattern we keep seeing: a business is named in around half the relevant answers from one model and a fifth from another. The first question is always "which number is right?" Both are. The gap is the interesting part.
Each model builds its picture of your industry from a different mix of sources and with different weightings. A model with strong web grounding tends to reflect what review platforms and recent press say this month. A model relying more on training data reflects the web as it looked a year or two ago. A business that has recently improved its reviews and rewritten its site will show the gap clearly: strong on the grounded model, weak on the other.
So read the gap as a diagnosis. Weak on the grounded model? Look at your reviews and listings as they stand today. Weak on the model leaning on older data? That is history, and the only fix is to put out enough consistent, current signal that the next snapshot of the web tells a different story.
What to do
Do not average the four mention rates into one score. Compare them.
Check which sources each model cites for your competitors. The sources they cite most are the ones you should appear in.
Track the gap over time. On monitoring plans, a narrowing gap is usually the first sign that fixes are working.
NOTE
Six website changes worth making before you blame the model
Take a worked example — a composite, not a named client, but built from the kind of readiness report we produce most often. A physiotherapy clinic scores badly for sports-injury questions in its own city despite being one of the biggest providers there. The website is attractive and fast. It is also almost silent: the homepage says "Move better. Live better." and very little else in actual text.
Six changes follow from that report. None of them needs a redesign.
Added a plain-language sentence at the top of the homepage: who they treat, for what, and in which city.
Created one page per specialism, each with the specialism in the title, heading and first paragraph.
Added MedicalBusiness and Service structured data with address, hours and services.
Replaced an image-based price list with an HTML table.
Corrected two directory listings that still showed the old address.
Added an FAQ page answering the questions patients actually ask, which happened to overlap heavily with the report's question set.
Why do these matter? Because each one turns something the clinic knows into something a model can quote. That is the whole mechanism. We cannot promise a number, and any agency that does is guessing — but a site that never states its specialisms in text has no chance of being recommended for them.
None of this is because structured data is magic. It is because models can only recommend what they can read, and most sites written for humans in a hurry never bother to say the obvious things out loud.
Find out where you stand.
A report gives you your own numbers to work from, not someone else's case study.