What AI assistants get wrong about your business — and how to correct it

Most conversations about AI search focus on absence — you asked ChatGPT for a supplier in Dubai and it named three companies that were not you. That is the obvious failure. The quieter one is being named and described incorrectly: opening hours from a previous branch, a price you stopped charging two years ago, a service you dropped, or a confident sentence about a completely different company that happens to share half your trading name.

This costs more than it looks. Someone who reads that you close at 4pm, or that your minimum project is double what it actually is, does not enquire and does not tell you why. There is no bounce rate for a conversation that happened somewhere you cannot see.

Where AI assistants get their facts about you

An answer about your business is usually stitched together from four kinds of source, and knowing which one is wrong is most of the work:

• Your own website — every page that is still live, including the ones you forgot about. This is the source you fully control, and the one most often at fault.
• Your Google Business Profile — hours, category, address, phone, plus whatever your reviews say about you.
• Directories and aggregators — UAE business listings, marketplaces, chamber and free-zone directories, old portfolio pages on an agency's site.
• Third-party mentions — press coverage, partner sites, forum threads, LinkedIn. Weak individually, influential when several repeat the same claim.

Some assistants search the web live and quote what they find; others answer from training data captured months earlier. That is why the same question can produce a current answer in one tool and a two-year-old one in another. The mechanics of which sources get picked up are covered in our guide to how AI search engines choose which websites to cite.

The five errors we see most often

1. Details from a previous life. An address from before the move, a landline that was disconnected, hours from the era when you opened on Saturdays. Almost always traceable to a listing nobody updated after the change.

2. Services you no longer offer — or don't offer yet. Businesses narrow their focus and leave the old service pages live because deleting pages feels like losing SEO. The pages keep describing you, and the AI keeps quoting them. The opposite also happens: your newest and most profitable service is invisible because it is mentioned only in a slide deck and one social post.

3. Prices that were never meant to be permanent. A launch offer, an old package, a figure quoted in an interview. Price is the fact customers most want and the one they are least likely to verify before deciding you are out of budget.

4. Mistaken identity. Common in the UAE, where trading names repeat across free zones and a similar company in another emirate can absorb your description entirely — including its reviews and its complaints.

5. Confident invention. When evidence is thin, models fill gaps plausibly: a founding year nobody published, staff numbers, an industry you have never worked in. The fix is not to argue with the model but to make the real facts easy to find.

Why you cannot simply request a correction

There is no listing to edit, because there is no listing. Each answer is generated on the spot from what the model has learned and, sometimes, what it retrieves at that moment. Nothing is stored as your record, so there is nothing to log into and amend.

Which leaves one honest route: change what the sources say, then wait. Assistants that browse live can reflect a corrected page within days. Facts absorbed into a model during training may persist until it is next updated, and stubborn errors often need the same correction visible in several places before they stop resurfacing. Anyone promising to scrub an AI answer on a fixed date is selling something they cannot deliver.

A correction routine that works

Write down the error precisely. Which assistant, which question, and the exact wording of the wrong claim. "ChatGPT says we are closed on Saturdays" is actionable; "AI says wrong things about us" is not. Test the questions customers actually ask — service plus area, price, comparison with a named competitor — not just your company name.

Fix it on your own site first. Find the page still carrying the old fact and remove it, redirect it, or update and date it. Then make sure the correct fact is stated somewhere plainly, in a sentence a machine can lift without context: opening hours, service areas, what you do and explicitly what you no longer do. Facts implied by a photograph or a PDF do not count.

Then the profile and the listings. Your Google Business Profile carries disproportionate weight for anything local, and directory entries are where dead phone numbers live longest. Work through them until name, address, phone and hours agree everywhere. Consistency is what makes a fact look reliable to a model.

Give the correction somewhere to be repeated. A single corrected page is one voice against several stale ones. An FAQ that answers the question directly, a structured-data block stating the same facts in machine-readable form, and a couple of third-party pages carrying the current version shift the balance of evidence. This is the same groundwork behind getting recommended by ChatGPT in the first place.

Re-check on a schedule. Monthly is enough for most businesses, and more often around a rebrand, a move or a price change. Ask the same questions each time and keep the answers, so you can see whether the error is fading or entrenched.

The version of this that catches most businesses

Nurseries and schools feel this hardest, because the facts that go stale are the ones parents filter on: fees, age ranges, term dates, whether you have a waiting list. A model quoting last year's fee structure removes you from a shortlist before anyone visits your site — which is why we treat fee and admissions information as maintenance work for education clients rather than something written once.

If you want to see what the engines currently believe about your business, our free AI visibility check runs a handful of real customer questions and reports what comes back — including the uncomfortable parts. Correcting and monitoring the record is part of the AEO work included in every monthly plan we run. What none of it can do is guarantee an AI answer will change on a particular date; what it reliably does is remove the wrong information from the places these systems are reading.

Common questions

Can I contact ChatGPT or Gemini to correct my details? No — there is no desk to write to and no entry to edit. Corrections have to be made in the sources the answer is built from.

How long until AI answers reflect the fix? Weeks rather than hours, and no honest fixed timeline. Live-browsing assistants update fastest; facts held in training data can persist until the model is refreshed.

Why does AI confuse us with another company? A shared name plus thin evidence. If your own pages never state clearly who you are, what you do and where you operate, the better-documented business wins the description.

Does old content on our site really matter? Yes — a retired service page or an old pricing page still live is a published statement, and it will be quoted next to your current pages. Clearing it out is usually the fastest win available.