Quick answer: first establish whether the characterisation is accurate. If it is, no amount of source work fixes it — the answer is fixing what customers are complaining about, and everything else is postponement. If it isn't accurate, find what's feeding it: usually a small number of detailed negative accounts, an unresolved public complaint, a confusion with another company, or simply an absence of information that produces caution by default. Then work the source layer, the same way as any other factual correction.
There's a category of AI answer that costs more than a wrong price or an outdated service list: the one where an assistant tells a prospective buyer to be careful with your company.
It arrives in different forms — a warning to check credentials, a note that reviews are mixed, a suggestion to consider alternatives, sometimes a flat statement that the company has a poor reputation. The buyer usually never mentions it. They just don't get in touch.
Start by establishing whether the AI characterisation is true
This has to come first, and it's uncomfortable enough that it usually gets skipped.
Assistants don't invent reputations. They summarise what's publicly written, and when the summary is negative it's most often because the underlying material is negative. If several detailed customer accounts describe missed deadlines, unresponsive support or work that didn't match what was promised, the assistant is doing its job accurately.
In that situation, source-level work is the wrong tool. Corrections don't apply to accurate statements, and suppression tactics carry more risk than the original problem. The answer is to fix what customers are describing and let the record change from there — slowly, because it will take as long to rebuild as it took to accumulate.
That's not a satisfying answer, and it's the honest one. Everything below applies to the case where the characterisation is not accurate, or is accurate about a period the business has already moved past.
What feeds a negative characterisation in AI answers
Where the claim isn't fair, it's usually one of five things.
A small number of detailed negative accounts. Length matters more than volume here. A long, specific, calmly written complaint is far more useful to summarise than twenty five-star ratings with no text — so a handful of detailed negatives can outweigh a large number of positive ratings.
An unresolved public complaint. A complaint with no response from the company reads differently from one where the company replied, explained and resolved it. Absence of a reply is itself information, and it's read as agreement.
Confusion with another company. A similarly named business with genuine problems, or a former partner, or a company sharing a founder's history. This is the entity problem covered separately in this cluster, and it produces some of the most unfair characterisations because none of the underlying material is about you.
Stale information from a bad period. A business that had a genuinely difficult year three years ago and has since changed can carry that year's record indefinitely, because nothing newer replaced it.
Absence of information. The one most companies don't anticipate. When there's little public material about a company, assistants frequently default to caution — advising the buyer to verify credentials, check references, be careful. It reads as a warning and originates in nothing more than thin data.
How to find what's driving a negative AI characterisation
Same method as any other error in this cluster, with one addition.
Run the prompt and read what's cited. Where sources are shown, the material is usually right there. Where they aren't, search your company name alongside the terms the assistant used — "reviews", "problems", "complaints", "scam" if it went that far — and see what ranks.
The addition: ask the assistant directly what its assessment is based on. Phrased neutrally — "what sources describe this company's reliability" — it often names them. Verify whatever comes back rather than trusting it, but as a lead it saves time.
Then check the entity question. Ask whether the assistant is distinguishing you from any similarly named company. If several negative claims trace to material that isn't about your business at all, this is a confusion problem, and the fix is entity separation rather than reputation work.
How to respond when the AI characterisation is unfair
The response differs by source, and applying the wrong one wastes months.
For detailed negative accounts that are inaccurate: respond publicly on the platform, factually and without heat. A calm, specific reply from the company is read as part of the record and changes how the whole exchange summarises. Arguing, or replying defensively, makes the summary worse.
For unresolved complaints: resolve them, publicly, and say so. This is the highest-value action available in most cases, and it's also just good practice — the reputational effect is a side benefit of doing the right thing.
For accurate-but-outdated material: you can't remove it, so it has to be outweighed. That means new material — recent, detailed, credible — describing how the business works now. Volume of thin content doesn't do this; a small number of substantial, dated pieces does.
For confusion with another company: entity separation, covered in the article on misattribution. Reputation tactics won't touch it.
For thin data producing default caution: the fix is publishing enough verifiable substance that caution is no longer the safest summary. Credentials, real client work, professional memberships, dated content that demonstrates rather than claims.
What not to do about negative AI characterisations
Everything in this list is actively sold, and every item carries more risk than the problem.
Don't gate reviews. Filtering who gets asked so only satisfied customers are invited is against most platforms' terms, detectable, and when discovered produces a far worse story than the reviews it was meant to prevent.
Don't generate positive reviews. Bulk positive reviews in a short window are recognisable to platforms and to readers, and the pattern itself becomes the reputational problem.
Don't threaten platforms or reviewers legally. For a business selling trust, a legal threat over a review is the single fastest way to convert a small negative into a large one. And it doesn't remove the underlying material from what assistants have already read.
Don't publish attack content about whoever complained. It produces more content pairing your name with the complaint, which is the mechanism you're trying to unwind.
Don't ask an assistant to change its answer. There's no channel for it, and the conversation doesn't persist beyond your session.
How long reputation changes take in AI answers
Longer than factual corrections, and worth knowing before starting.
A wrong price is a discrete fact that can be replaced. A characterisation is a synthesis of many sources, and changing it requires changing the balance of that material rather than editing one page. Where the response is a public reply to a complaint, the effect can appear within weeks of a recrawl. Where it requires new material to outweigh old, it's months. Where the characterisation sits in a model's training data, it persists until a retrain regardless of what you publish.
The realistic goal for the first few months isn't reversal but dilution: making sure that anyone reading a summary also sees the response, the resolution and the current evidence.
How I work on negative AI characterisations of a company
The part I do first is the honest assessment. If the characterisation reflects something real, I'll say so — spending months on source work while the underlying problem continues is expensive and doesn't hold. Where it's unfair, the work is the same source-layer discipline as any other correction, plus the public responses that only you can make.
What I don't do is suppression tactics. Beyond the ethics, they're the highest-risk option available and the one most likely to produce a worse outcome than the original problem.
Where to start if AI describes your company negatively
Reduced to one principle: establish whether it's true before deciding what to do, because the two situations have nothing in common except the symptom.
A practical step for today: ask an assistant, neutrally, what it can tell a buyer about working with your company, and read the answer carefully. Then find whatever it's summarising. If the material is accurate, you've found your priority list. If it isn't, you've found your correction list.
If you'd like help establishing which of the two you're looking at, write to me and we'll go through it.
