What to do when an AI assistant describes your business wrongly

You can't edit an AI answer. What you can change is the material AI assistants read. Five steps: record exactly what was said, on which engine, with which sources cited; trace the wrong fact back to where it's published; correct it at the source and on your own site in identical wording; check the assistants can actually read your site; then re-run the same prompt on a schedule until the description settles. None of that forces an outcome. It removes the conflict that produced the wrong answer.

Written for owners and managers of Klang Valley SMEs with 2 to 50 staff and nobody in-house who owns marketing. Disclosure: it comes from Geo One, a Generative Engine Optimisation (GEO) and Answer Engine Optimisation (AEO) agency in Kuala Lumpur. GEO is the work of improving the signals AI systems use to understand a business. AEO is structuring content so an assistant can lift a clean, correct answer out of it. If both terms are new, start with the plain-English explanation of GEO.

Why do AI assistants describe a business wrongly?

Your website says one thing. Your Google Business Profile says another. A directory you signed up to years ago still lists the old office address and a phone number nobody answers. An assistant weighing all three may pick one, or blend them into something that was never true anywhere.

Three patterns come up most often:

• Stale facts. An address left behind after a move, a branch that closed, a partner who left, a service you stopped offering two years ago.
• Conflicting facts. The same detail written five different ways: trading name versus registered name, "Level 3" versus "3rd Floor", one phone number on the site and another on a listing.
• Missing facts. Nothing anywhere states your opening hours, your languages or which suburbs you cover, so the assistant fills the gap with something reasonable and wrong.

Gemini and Google AI experiences may use Google's local and business data alongside web sources, so an accurate Google Business Profile is still worth maintaining. Treat that as a reason to get your Google Business Profile exactly right, not as a reason to ignore the rest.

Step one: record the error before you fix anything

A screenshot in a WhatsApp group isn't a record. Log six fields, one row per attempt: engine, the exact prompt you typed, the language, the date, the incorrect wording word for word, and any sources the engine cited. Do it twice more on different engines, because an error in ChatGPT often doesn't appear in Gemini, and the difference tells you which source is at fault.

Run the same prompt in Bahasa Malaysia too. The method for that is in our guide to tracking Bahasa Malaysia prompts in GEO reporting.

Without the log, you can't tell later whether a fix worked or the engine simply phrased things differently that day.

Step two: where did the wrong fact come from?

Ask the assistant. Perplexity shows its citations by default, and other assistants may list sources when searching the live web. Open every link. One of them is normally carrying the error.

When no sources are shown, go hunting yourself. Put the wrong detail into Google in quotation marks: the old phone number, the former address, the outdated price. You'll usually find it sitting on a business directory, an old press mention, a job advert you posted on a portal, or a supplier's "our partners" page you forgot existed. Write down every URL where the wrong fact appears. That list is your work order for step three.

Step three: correct it at the source, in identical wording

Before touching anything, write down one canonical version of each fact and don't deviate from it:

• Registered name exactly as it appears on your SSM registration, plus the trading name you actually use, if they differ.
• The full address in one format. Pick "Level 3" or "3rd Floor" and never write the other.
• One phone number in one format.
• Opening hours, languages spoken, and the areas you serve.

Then push that exact string everywhere: your website footer and contact page, Google Business Profile, Facebook, LinkedIn, your professional body's public listing, and every directory URL from step two. You know "Jln" and "Jalan" are the same street. Don't make the machine decide.

The fix that most owners miss is what to do about a source you cannot edit. Some directories carrying outdated information may have no working contact address, making direct correction impossible. You can't delete the contradiction, so resolve it instead. Put a dated, plainly worded correction on your own site where a crawler will find it: "We moved from [old address] to [new address] in March 2024. Our phone number is unchanged." Name the old detail explicitly. A model that encounters both versions now has a stated reconciliation rather than a straight conflict, which may make the corrected version easier to select or cite. Post the same line as a Google Business Profile update.

The same logic applies to the page on your own site that carries the old information. Update it. Don't delete it — a deleted page takes the correction path with it and leaves the stale copies unanswered.

Step four: check the assistants can actually read your site

You can correct every fact and still be ignored, because the engines never get in. Four checks, all of which take minutes:

1. Open yoursite.com/robots.txt. Look for Disallow rules aimed at GPTBot or OAI-SearchBot, ClaudeBot, PerplexityBot, or Google-Extended. Bingbot matters for Microsoft Copilot.
2. Check your CDN or security settings. Check whether your host, CDN or security plugin is blocking AI crawlers; Cloudflare, for example, provides AI crawler controls that site owners should review.
3. Make sure the facts are text. An address inside a header image, a phone number that only appears after a chat widget loads, or hours held in a third-party booking script are invisible to most crawlers. Right-click the page and select the text. If the text isn't selectable, crawlers are likely to miss it too.
4. Add structured data, a LocalBusiness block carrying the same canonical name, address, phone and hours. It gives machines an unambiguous version of the facts you just standardised. Regulated practices have extra fields worth getting right; we cover that separately for clinics.

Step five: re-run the same prompt on a schedule

Weekly for the first month, then monthly. Same prompts, same engines, both languages, logged into the same six-field sheet from step one. How often each engine refreshes its view of a given business is not publicly documented, which is exactly why you measure it rather than assume it.

Two things to expect. Engines will move at different speeds — it's normal for Gemini to reflect a corrected Google Business Profile well before ChatGPT does. And the description often gets partly right first: correct address, stale service list. Log the partial win; don't call it a failure.

Use three consecutive clean runs on each engine as a practical monitoring threshold, not a guarantee that the error cannot return. If the error is still there a month after the source was fixed, the fix wasn't complete — go back to step two and find the copy you missed.

Three things that don't work

Telling the chatbot it's wrong. Correcting it inside a conversation changes that conversation and nothing else. It will not propagate.

A one-off press release. A single new mention doesn't outvote a dozen listings that still say something else.

Fixing your website only. Your site is one source among many, and for local facts it isn't always the one the assistant trusts most.

The pattern underneath all three: these systems respond to the weight and consistency of published evidence, not to requests. That is genuinely frustrating, and it's also the good news — consistency is something a small business can control completely, without a budget.

Where this stops being a Saturday morning job is when the wrong fact is load-bearing: a closed branch still taking enquiries, a competitor's name attached to your services, a price that hasn't been true for years. At that point you're not correcting a listing, you're rebuilding a set of signals across four engines and two languages, and it's worth having someone run the log for you.