How to stop AI inventing things about your business

Ask ChatGPT to write the services page for your business and it will hand you something that reads perfectly well. Somewhere in the middle there will be a sentence you never told it. A price. A response time. A guarantee. A number of years you have been trading. It arrives with the same calm confidence as the parts that are true, which is the entire problem.

The short version: a model asked to write about your business will fill the gaps rather than leave them. That is not a bug you can prompt away. It is what the thing does when it does not have the information.

Why it makes something up instead of leaving a blank

The model is not looking your business up. Even with browsing turned on, the bulk of what it writes about you is not retrieved from anywhere. It is predicted. It has read an enormous number of pages written by and about businesses like yours, and learned in extraordinary detail what those pages tend to say. Ask it about a specific electrician in Papakura or a physio in Geelong and it has no record of that business. What it has is the shape of an electrician's page. So it writes that shape and puts your name on it.

There is no internal flag that lights up and says "I do not know this one". A blank is not really a move available to it. A sentence like "with over 15 years serving the local area" is statistically ordinary in that spot on that kind of page, so out it comes, in the same voice as your actual phone number.

It is also why the invented details are believable. It will not tell your customers you have traded for 400 years. It will say twelve. It will not invent a guarantee nobody offers. It will invent the one your closest competitors offer.

The invented details are never absurd. They are always plausible, and plausible is exactly what makes them dangerous.

This is the same behaviour that makes AI copy sound like nobody in particular wrote it. The model reaches for the average whenever it has nothing specific, whether the gap is a fact or a personality. I have covered that in why AI content sounds generic.

The things it invents most reliably

After enough drafts the same categories keep turning up. These are the ones I check first, every time:

• Numbers of any kind. Prices, call-out fees, hourly rates, "from $X" figures, response times, discounts.
• Dates and durations. Years in business, "established in 2009", turnaround times, how long a job takes.
• Credentials and qualifications. Licence classes, registrations, "fully qualified", training a staff member is said to hold.
• Accreditations and awards. Industry body membership, certified installer status, "award winning", association logos it suggests you add.
• Guarantees and warranties. Lifetime warranties, satisfaction guarantees, money-back promises, "we will beat any written quote".
• Service lists. Services you do not offer, sitting between two you do.
• Coverage areas. Suburbs and towns you do not travel to, pulled from whatever sounds adjacent.
• Team and capacity claims. Staff numbers, after-hours availability, languages spoken, weekend cover.

Notice the common thread. Every one is a specific, checkable claim that a customer reads and then acts on. The model is not inventing adjectives. It is inventing the load-bearing bits.

Why this is worse than embarrassing

Most owners hear "AI makes stuff up" and picture an awkward correction. The real costs are more concrete, and they land on you rather than on the tool.

A price you now have to argue about

Someone rings and quotes a number back at you from your own website. Two options, neither good: honour a price you never set, or tell a prospective customer your website is wrong. The second is not a neutral conversation. Thirty seconds in, you have told them the information you publish about yourself cannot be relied on.

Enquiries you cannot fulfil

An invented line in a service list generates real enquiries from people who wanted that exact thing. You burn your time and theirs, and end the call looking like you do not know what your own business does. Some of them were referred to you, and they go back and report that you do not actually do it.

Guarantees you never agreed to

A guarantee published on your own site is a representation you have made to a customer. Consumer law on both sides of the Tasman takes a dim view of claims a business cannot back up, and "the AI wrote that bit" is not a position I would want to argue from. You do not need legal training to see that an unbacked lifetime warranty is a liability rather than a typo.

Credentials that are not yours

This is the one that genuinely worries me. There is a real difference between a page overstating your experience and a page claiming a registration you do not hold. That is not sloppy copy. It is a false claim about a regulated status, published under your name, on a site you control.

If you run a clinic or a psychology practice, this is a different problem

For a tradie or a retailer, an invented fact is an accuracy problem with a commercial cost. For a clinic, a psychology practice, a physio or a dental practice, it is a professional standards problem, and the people who decide whether you have one are not your customers.

Health advertising in both Australia and New Zealand is governed by rules about what a regulated practitioner may claim: no promises of outcomes, care with testimonials, no misleading comparative or superlative claims, nothing creating unrealistic expectations of a treatment. Those rules are not obscure. They are invisible to a model trained on a great deal of ordinary marketing copy and none of your professional obligations.

So it writes marketing sentences. "Proven results for anxiety." "The best psychologist in the area." "Overcome your symptoms for good." Every one is a normal sentence on a business website, and every one is the kind of claim that puts a registered practitioner in front of a complaint they did not need.

Then add the invented specifics: session lengths, fees, how many funded or subsidised sessions someone is entitled to, referral pathways, wait times. Get one wrong and you are not just annoying somebody. You are affecting whether an unwell person seeks help.

My rule for regulated practices is blunt. The model can help with structure, plain-English explanation and readability. It does not get to write a clinical claim at all. Anything describing what a treatment does, who it suits or what it achieves comes from the responsible practitioner and goes back to them before publication. The fuller version is in marketing a clinic without breaching the rules.

The fix is supply, not prompting

The natural instinct is to solve this with a better instruction. "Do not make anything up. Only use verified information about my business." I have tried every version of that line. It helps at the edges and does not fix the problem, because the model cannot reliably tell which of its own sentences came from you and which it generated. Recalling and inventing are the same operation to it.

What works is changing the job. Stop asking it to recall your business and start asking it to arrange facts you supplied.

Write your facts down once

Build one plain-text page of everything a writer would need: exact service names as you sell them, current prices or the reason you do not publish them, the towns you actually cover, the year you started, the qualifications you genuinely hold, what you guarantee in the precise words you would use in a dispute, team size, hours. Paste it in with every request. It takes an hour once and kills most of this permanently. The broader argument is in the context AI needs before it writes anything.

Never let it write a number you did not give it

If I could keep only one rule, this is it. Every digit in an AI draft is wrong until you have matched it against your own brief. Not "does that look about right", which is how wrong numbers survive, but "did I supply this exact figure". If you did not, it came from a page about somebody else.

Give the blank somewhere to go

Models follow an instruction that gives them something to do far better than one telling them to stop. So instead of "do not guess", try "where you do not have a fact, write [TO CONFIRM] and carry on". You get a draft with visible holes, which is exactly what you want, and they take two minutes to fill.

Make it audit its own draft

Then send one more message: "List every factual claim in that draft about my business that I did not give you." It is surprisingly good at this, because you have turned a generation task into a sorting task. Not a replacement for reading the draft yourself, but it surfaces the obvious ones in ten seconds.

The five minute check before anything goes live

This applies to a service page, a Google Business Profile post, a newsletter, a quote follow-up. Anywhere a model wrote words going out under your name.

• Find every number and check each one against your brief. If you did not supply it, cut it or confirm it. No exceptions for numbers that look harmless.
• Read every sentence that names a service. Do you offer exactly that, under exactly that name, right now?
• Search for the trigger words: guarantee, warranty, free, certified, accredited, licensed, registered, qualified, award, approved, official. Cut anything you could not evidence in writing.
• Check the geography. Every suburb and town named, and whether you would genuinely take a job there.
• Check the team claims. How many people you have, what they are qualified in, your availability.
• Read the opening and closing paragraphs aloud. Sweeping claims congregate there, where the model reaches for a strong finish.
• In a regulated field, the responsible practitioner reads it before publication. Not the person handling the marketing, the person holding the registration.

That is about five minutes on a 900 word page. Far less than you will spend on the phone explaining why your website said something you never said.

This is not an argument against using AI

I use these tools constantly. They are excellent at structure, at turning rough notes into something readable, at drafting the boring eighty percent of a page. What they are not is a source of information about your business. Think of it as a fast, confident junior who has never met you, never seen your price list, and would rather write a smooth sentence than admit a gap.

So the judgement stays with the only person who knows the facts. That is the same split I apply across my own work, set out in human in the loop SEO and in the wider guide to using AI content without wrecking your site.

The model's job is to arrange your facts. Yours is to supply them and check nothing else crept in. Get that division right and the problem largely goes away. Get it wrong and you are publishing a machine's best guess about your own business, in your voice, on your domain, with your name at the bottom.