The mistakes are rarely random, which is the useful part.
An AI tool writing about your business tends to get the general shape right and the particulars wrong. It gives a product an attribute that belongs to a different product in your range. It quotes a price from last season. It describes a feature you retired two years ago. It writes with total confidence about something that is true of your competitor.
If you have seen that pattern and concluded the tool is unreliable, the diagnosis is close but the cause is worth knowing, because it changes the fix entirely.
Why does it invent a detail instead of saying it doesn’t know?
Because saying “I don’t know” scores badly.
OpenAI published research on exactly this in Why Language Models Hallucinate, and the finding is more mundane than the word “hallucination” suggests. Models are trained and graded against evaluations that mark an answer right or wrong. Under that scoring, a model that guesses beats a model that abstains, because a guess is sometimes right and an abstention never is. The behaviour is not a glitch in the model. It is the strategy the scoring rewarded.
Read that back as a practical rule and it is almost reassuring: an AI tool will fill any gap you leave. So the question stops being “how do I make it more accurate” and becomes “which facts have I actually given it?”
This is the magnifier principle in its least glamorous form. Put in slop and you get out bigger slop. Put in nothing at all and you get out something confident and invented.
So what does the fix look like?
Not a better prompt. A better source.
The most reliable pattern we build for clients is boring on purpose: one reference file that is the single source of truth for the facts the AI keeps getting wrong, and a system that reads it every single time rather than remembering it.
For a business with a product range, that file usually holds one row per product with every attribute the content depends on: the official name, the category, the current price, the correct reference photo, the year the packaging changed. For a service business it is the service list, the areas covered, the current turnaround times, the claims legal has approved. It is not sophisticated. It is a spreadsheet.
What makes it work is that it is the answer rather than an answer. When there is one file, and everything reads from it, updating a price is one edit rather than a hunt through six prompts and a document nobody remembers writing.
Why do the exceptions matter more than the rules?
Because a general rule is where the wrong details come from.
Give a system a rule and it applies the rule everywhere, which is exactly what you asked for and almost never what you meant. “Show the product in a glass” becomes a glass on every item including the ones that do not come in one. “Mention the warranty” becomes a warranty claim on the one product line that has never carried it.
So the exceptions go in the file, explicitly, next to the items they apply to. Not in a paragraph of guidance somewhere else. A column that says which rule does not apply here, and why.
We treat this as a design step rather than a cleanup step. In the discovery conversation, once someone describes how their content is supposed to work, the follow-up question is always: where does that not hold? The answers to that question are worth more than the rule itself, because those are the errors your customers would actually notice.
What about all the old material?
Move it out of reach, rather than asking the system to ignore it.
Most businesses have years of retired assets living alongside current ones: the old logo, last year’s packaging, a photo of a product that has been reformulated, a price list nobody deleted. If those files are in a folder the AI can read, they will eventually come out in something you publish, and the instruction “use only current assets” will not reliably stop it.
The fix is access, not instruction. Keep an active folder and an archive folder, point the system at the active one, and move things across as they retire. It takes an afternoon once and then it holds by itself, which is the test of whether a control is real: does it survive a busy week when nobody is checking?
The same logic applies to the reference file. If the AI can read the live sheet directly, it is current by definition. If somebody has to export it and paste it in, it is current until the first week they forget.
Does the file have to be anything special?
No, and it is better if it isn’t.
A spreadsheet, a plain document, a folder of ordinary files on storage you control. Formats anything can open, in a place you own. The system reads and writes to those, rather than keeping what it knows about you inside a tool you rent. That way accuracy and portability turn out to be the same piece of work, which is a rare thing to get for free. There is more on the portability half of that in what happens to your data if the AI company shuts down.
The pattern scales further than people expect. In one engagement, a small studio’s entire content system runs off a single structured document holding voice, audience, offers and brand rules, and every part of the system reads from it before writing anything. Owner time on social content went from six hours a week to about forty-five minutes of review. The hours did not come from the drafting being fast. They came from not having to correct it.
How do you know it’s actually working?
Check the exceptions, not the easy cases.
Anything will get your best-known product right. Pick the three facts that are unusual about your business, the ones a new hire gets wrong in their first month, and ask the system about those. That is where you find out whether the file is being read or the model is guessing.
And expect a round of correction. Very rarely does any of this land exactly right on the first attempt. You try it, you find the two places it still invents, you write those two facts down where the system can see them, and the error class disappears rather than the individual error. That loop is short, and it converges.
The wider point
“AI gets things wrong about my business” usually turns out to mean “nobody has written down what is true about my business in a form a machine can read.”
That is a much better problem to have. It is finite, it is one afternoon of work for most small businesses, and the artifact you produce at the end is useful to your team whether or not any AI ever touches it. The new person can read it. The agency can read it. You stop being the only person who knows that one product is the exception.
If you want to see what your current setup is actually reading from, and where the gaps it is filling in are, book a free readiness call. Twenty minutes, no pitch. It pairs well with reading what AI already says about your business, because the wrong details usually come from the same place: a fact you know and never published.