Notes

Can I check whether AI tools get my business facts right?

AI tools can state your opening hours, prices, address, products or customers wrongly, and each tool may be wrong in a different way. You can check this by hand in an afternoon by asking every tool the same fixed questions in the same words and writing down what each one said.

Fix the questions before you ask anything

The check only works if every tool receives identical wording, so write the questions down first and do not improvise them in the chat window. Each question should name the business and ask for one fact, because a broad question such as "tell me about this company" lets a tool choose which facts to mention and hides the ones it would have got wrong.

Five facts cover most of what a customer needs before they act: when you are open, what you charge, where you are, what you sell and who you serve. The wording below is a template to adapt, and the exact words matter less than keeping them unchanged for every tool and every repeat.

One fixed question for each basic fact
FactQuestion to ask, word for wordWhere to find the true answer
Opening hoursWhat are the opening hours of [business name] in [town] on weekdays and on Sundays?Your own contact page and your business profile
PricingHow much does [business name] charge for [your main product or service]?Your current price list
LocationWhat is the street address of [business name] and which suburb or town is it in?Your contact page
What you sellWhat products or services does [business name] provide?Your home page and product pages
Who you serveWhich customers is [business name] for, and which areas does it serve?The page where you describe your customers
Write the true answer beside each question before you start, so that marking an answer later is a comparison and not a judgement made from memory.

Ask several tools and keep a record

Use at least three different AI tools, including one that searches the web and one that answers from what it already knows, since the two can disagree about the same fact. Open a fresh conversation for every question so that an earlier answer cannot shape the next one, and ask each question three times in separate conversations. Repeating matters because the same tool can give different answers to identical wording, so a single reply tells you what one draw looked like and little more.

For each answer, copy the text exactly as given and note the tool, the date, whether it showed sources and whether it searched the web. Paraphrasing at this stage destroys the evidence, because the difference between "open until five" and "open around five" is the kind of detail you are checking for. A spreadsheet with one row per answer and a column for your verdict is enough.

Sort each answer into wrong, out of date or vague

An answer that differs from your records can differ in three ways, and they lead to different repairs. A wrong answer states something that was never true of your business, an out of date answer states something that was true once, and a vague answer avoids committing to the fact at all.

How to sort one answer against your records
VerdictTestTypical example
WrongNo version of your business, past or present, matched itA suburb you have never traded in, or a service you have never offered
Out of dateIt matched your business at an earlier dateLast year’s prices, or an address from before you moved
VagueIt is true only because it commits to nothing checkable“Hours vary, check their website”, or “a range of services”
CorrectIt matches your records in every detail a customer would act onThe same hours, price and address you publish today
Ask of each answer whether it ever matched the business. A yes points to an old source somewhere, and a no points to a confusion with another organisation or an inference from thin evidence.

The out of date test needs your own history, which is why a dated note of when you last changed each fact is worth keeping. The reasoning here is that stale facts usually come from a source that was accurate when published, whereas an invented fact has no such origin, so the two call for different searches. When a tool cites a page, open it and see whether the page itself carries the old fact.

Treat a vague answer as a finding and not as a pass. A customer who is told to check the website may do so, but the tool has still failed to carry the fact, and for a question about pricing or hours that usually means the facts are hard to find in a readable form on your own pages.

What to fix first

Order the repairs by what a customer loses when the answer is wrong. An incorrect closing time or address sends someone to a locked door, a wrong price creates a dispute at the counter, and a loosely described product range costs a sale nobody sees. On that reasoning, hours, location and price come before the description of what you sell and who you serve.

Within each fact, fix the sources you control before the ones you do not. Start with your own website, then the business profiles and directory listings you manage, and make every one of them state the same fact in the same words. Only then contact the owners of other sites that repeat the old version. Where several tools repeat the same wrong detail, an old public page is the more likely origin, and the post on what to do when AI describes your company wrongly explains how to trace an error to unclear pages, a retrieval failure or stored memory.

Leave vague answers until last unless they cluster on one fact. If every tool hedges on pricing, publish the price as plain text on a page that a crawler can read, because a figure that sits inside an image or an interactive widget may never reach the tool at all.

What the check cannot tell you

Nine answers per fact, from three tools asked three times each, is a small sample, and it shows what is likely to appear without measuring how often it appears across everyone who asks. A tool that gave the right hours three times may still give wrong ones tomorrow, and a tool you did not test may behave differently. The result also covers only questions that name your business, so it says nothing about whether you are mentioned when a customer asks for options without a name.

Corrections do not arrive on a schedule you control. Any tool that searches the web may reflect a corrected page within days, while an answer drawn from stored knowledge can lag for months, a gap described in the post on how long a website change takes to show up. Once you have fixed your pages, repeat the same questions in the same words on a fixed date and compare the new record with the old one, which is the only way to tell whether anything moved.

AnswerFit measures how AI models describe, compare and recommend one organisation, then re-tests after a change to see whether the answer moved. The methodology page sets out how that measurement is put together and what it deliberately refuses to claim.