Fixing what AI gets wrong about your hours

If four sources say you close at five and only your website says nine, the answer says five. Here is how to work out which source to fix first.

Why it happens

Engines blend several sources. If four of them agree you close at five and only your own site says nine, the answer says five. It says it confidently, to a customer who will never see your correction, and who simply goes elsewhere.

This is the part of the problem that feels unfair. You did update your website. The website was never the deciding vote.

Why hours specifically

Of all the facts an engine can get wrong about you, this is the one that costs the most per occurrence, and it is treated as the most expensive class of error for that reason.

A wrong service description loses you a job you might not have won anyway. A wrong phone number is at least visibly broken — the customer notices and often looks you up again. A wrong closing time produces a customer who was actively choosing you, went somewhere else on the strength of a sentence, and never learned they were misinformed. You do not get a complaint. You get nothing, which is indistinguishable from never having been asked.

Hours also rot faster than anything else on the list. An address changes once a decade. Hours change seasonally, at Christmas, when a member of staff leaves, and every time you decide Saturdays are worth trying again.

Worth saying plainly, because it is the objection every owner raises: none of this is your fault, and none of it is fixed by being annoyed about it. The records were built from feeds you never saw, by companies you have no relationship with, and the engines are averaging them in good faith. The only lever anybody has is to make the sources agree, starting with the one that is actually being read.

The fix, in order

Correct the most-cited source first, then work down until the sources agree. Your report names which ones came back for your prompts; if you are doing this by hand, ask the engines your customers’ questions and note what they cite.

The order matters more than the effort. Fixing five obscure listings while the one an engine actually reads stays wrong changes nothing at all.

Finding the source that is actually wrong

Do not start by editing things. Start by asking, and write down what comes back.

Ask an assistant when you are open, in the way a customer would — whether you are open now, whether you are open on a Sunday, what time you close today. Ask each question more than once and in more than one engine, because answers vary between runs and one wrong answer might be one unlucky draw rather than a pattern in your data.

Then look at what was cited. Usually one of three things is happening. The citations point at a specific listing that carries the old hours, which is the easy case. Or the engine cites nothing in particular and is recalling something stale, which usually means the correct hours are not stated anywhere machine-readable and it is falling back on an old impression. Or the citations are right and current and the answer is still wrong, which happens and is covered further down.

The listings nobody cites are not your problem this week. Working the whole internet alphabetically is how this turns into an infinite job instead of an afternoon.

Your profile first, then the rest

For most local businesses the Google Business Profile is the single most consulted record, and it is also the one you can correct in about ten minutes.

Set the regular hours, then set the special hours — public holidays, the week between Christmas and New Year, the day you close for a family event. Special hours are the ones that decay silently: a holiday entry from last year that has expired leaves the profile asserting normal hours on a day you will be shut, and the profile is exactly the record everything else copies.

If you are temporarily closed, say so there rather than leaving the hours in place and hoping. An explicit temporary closure is a fact an engine can state correctly. Silence gets filled with the last thing anybody recorded.

Then work down your own citation list — the directories, the aggregators specific to your trade, and the industry bodies you are listed with. The mechanics of finding and claiming those records are in the sources that keep coming back.

Say it twice on your own site

Once in prose, where a person reads it, and once in structured data, where a parser reads it without having to interpret a footer.

Both matter and they matter differently. The visible text is what a human checks before driving over. The markup is what stops a machine guessing whether the numbers next to your address are opening hours, a phone extension or a price. The field and its quirks are covered in the schema walkthrough — use the array form, one entry per pattern, and set a reminder to revisit it when the clocks change your trading pattern.

Make sure the two agree. A site whose visible page says nine and whose markup says eight has published a contradiction about itself in the one place designed to remove ambiguity, and an engine reconciling that will pick one without asking you.

What to check while you are in there

Hours are the obvious one, but they travel with company:

| Field | The common failure | | --- | --- | | Name | A trading name in one place, the limited-company name in another | | Address | An old unit number left behind after a move | | Service area | Towns you left, or towns you now cover and never added | | Services | A list written years ago, before half your revenue existed | | Phone | A number that forwards, on a listing nobody claimed |

Each of these produces the same failure as the hours do: a confident, specific, wrong answer. And each is fixed the same way — find the sources, make them agree, starting with the ones that keep coming back.

When the sources are right and the answer is still wrong

This happens, and it is worth knowing what it means before you go back through every listing a second time.

Engines cache. An index entry built three weeks ago does not update because you edited a record yesterday, and a model answering partly from recall can produce a detail that no current source supports. There is nothing to fix in that case except to have made the correction and to wait for the next crawl.

What is worth doing is checking the claim carefully before assuming it is wrong. Wrong-fact detection is deliberately conservative for the same reason you should be: a claim only counts as an inaccuracy when it asserts a value we actually hold and that value confidently disagrees. An answer saying you close at five is a checkable assertion. An answer saying to ring ahead because hours vary is not asserting a time at all, and treating it as an error would be manufacturing a problem.

The same conservatism applies to the always-open claim, which turns up more often than you would expect. It is checked against whether any of your listed periods actually runs the full day, because an engine describing a business as round-the-clock when it closes at six sends people out at midnight.

Anything vague, partial or ungrounded is skipped rather than counted as clean — skipped is not the same as correct, and the report says which is which.

The habit that stops it coming back

Fixing this once is an afternoon. Keeping it fixed is a habit, and the habit is smaller than people expect.

Keep a list of every place your hours are published. Write it down somewhere dull — a note, a spreadsheet, the back of the same document that holds your logins. Most local businesses have between five and a dozen, and the value of the list is that it turns a vague anxiety into a checklist you can work through in twenty minutes whenever something changes.

Then attach the list to the events that actually change hours. Not a calendar reminder every quarter, which everybody ignores by the third one, but the moment you decide something: taking on a Saturday, dropping an early opening, closing for a fortnight in August. The decision is the trigger, and the list is what you open when you make one.

Two occasions are worth putting in a calendar anyway, because they catch everyone. The run-up to the winter holidays, when special hours need setting and last year’s have expired. And whichever seasonal change your trade actually has — the month the garden work starts, the week the schools go back, whatever it is for you.

What good looks like afterwards

You will know it has worked when the answers get boring.

Ask the same questions a month later and the times that come back match the times you keep. No contradiction between one engine and another, no confident wrong closing time, no phrasing that hedges because the sources disagree. Hedging is itself a signal worth noticing: an engine that says hours vary, check before visiting is frequently telling you it found conflicting records rather than none.

That is the whole outcome, and it is deliberately unexciting. Nobody chooses a business because its hours are correct. They just fail to choose the one whose hours are wrong, and you never find out that it happened.

How long until it changes

Longer than feels reasonable. Records propagate on their own schedule, engines re-read when they re-read, and a fix made on a Tuesday shows up whenever the source that carries it is next pulled into an answer. Weeks rather than days is the honest expectation.

Two things make the wait shorter. Correcting the most-cited source first, because it is the one most likely to be re-read soon. And making sure every other source agrees, so that when the re-crawl happens there is no contradiction left to reconcile.

Then leave it alone and re-measure later. Accuracy is one of the pillars where a single fix moves the number visibly — how much, and why it is weighted the way it is, is worth reading before you judge whether the afternoon paid for itself.