This is a Heading

The answer kept a version of the company that had ended

Meta description: Before correcting a misleading AI answer, diagnose what it said, why it might have said it, and which sources it could trust. More on his work: his consulting practice.

The general counsel forwarded it with a short note: is this actionable? The AI answer stated, neutrally and without emphasis, a fact about the company that had stopped being true — a settled lawsuit presented as current, a brand absorbed in a merger described as independent, a market exited years ago listed among the company’s activities. Nothing in the answer was an attack. That is what makes the pattern dangerous, and it is why I am writing about it: a hostile error triggers the company’s defenses, but a calm, reasonable, wrong description triggers nothing. It simply informs buyers, partners, and journalists, incorrectly, at scale, with the tone of settled fact.

When an AI assistant presents an incomplete or misleading account of a company, the immediate reaction is often to write another page telling the model what it should say. I prefer to diagnose first. What, exactly, did the assistant say? Which alternatives did it mention? Does the error concern a product, price, service area, professional qualification, or broader reputation? A vague complaint that the AI does not like a brand is not yet a workable brief.

What wrong fact is stable across answers?

The first diagnostic I run is persistence: is the wrong fact stable across engines and across phrasings? A one-off error is noise — models improvise. A stable error is a record problem: somewhere, a source states the wrong thing, and the engines have agreed to trust it. Stable errors have a genealogy. I have traced them to a single high-authority article from the wrong year, to a database entry nobody owned, to a Wikipedia edit war that ended badly. The stability is the finding: the company is not being mis-imagined. It is being mis-recorded, and the engines are faithfully repeating the record.

Why does owned-site correction fail to move it?

The distortions worth unwinding first are the ones a neutral reader would also act on: an outdated price, a wrong service area, a qualification the company does not hold, a reputation claim sourced from somewhere untraceable. I work backward from the visible result. The answer is the observation; the public information environment is the field I can inspect — the website’s technical accessibility, the content available about the company, how its entities are described, the competitor pages visible in traditional search, and the off-site publications associated with those competitors.

The company’s instinct is to correct its own surfaces: update the site, issue a statement, fix the About page. I have watched these corrections stand, correct and current, while the engines continued repeating the old fact. The mechanics are not mysterious once seen: the engine’s account rests on the corroborating record, and the company’s own site is a single, interested source. The corrected page says “we are no longer in that market.” The uncorrected record — directories, old coverage, archived profiles — says otherwise, in many voices. The company corrected the document it controls. The distortion lives in documents it does not.

What tone makes the error more believable?

The tone problem is the part leadership underestimates. The engines do not repeat the old fact with the breathless framing of the original coverage — the lawsuit’s dramatic headline, the merger’s uncertainty. They repeat it calmly: a plain declarative sentence, no urgency, no slant. Calm reads as settled. I have watched executives dismiss the distortion because it lacked hostility, not seeing that the lack of hostility is the mechanism of harm. A buyer who reads an alarmed error discounts it. A buyer who reads a calm error files it as background knowledge — and background knowledge does not get checked.

What pattern marks distortion rather than absence?

Absence is legible: the company is missing, the gap is measurable, the remedy is presence. Distortion is illegible in the company’s instruments: the company is present, mentioned, described — the mention metrics are healthy — and the description is wrong. I have learned to separate the two questions leadership tends to merge: are we visible, and is what is visible true? The first has dashboards. The second has, in most companies I visit, no instrument, no owner, and a forwarding email from the general counsel.

I have described the method as regression-like reasoning: an analogy to looking for the factors that might explain an observed outcome, not a claim that I have fitted a statistically validated model to a proprietary ranking system. Sometimes the intervention is on the site; sometimes it is outside it. After the change, I revisit the target question and related questions, retaining the before-and-after wording, the date, and the test conditions. Even if the account improves, attribution should be modest. I cannot guarantee that an assistant will repeat a corrected answer. I can help a company improve the accuracy and availability of the information from which people and systems may form an answer.

[AWAITING MIKLOS — scene: second scene]

What question locates the source the model is trusting?

The company’s question is “how do we correct the answer?” and it leads to frustration, because the answer is an output, and outputs do not hold still to be corrected. The productive question asks about inputs: which record, trusted by whom, is feeding this description? That question has a findable answer — the genealogy always exists — and finding it converts an atmospheric grievance into a specific piece of work. The distortion is not in the model. It is in the record the model reads.

The question that keeps the correction honest: which specific public source would have had to change for this answer to change — and has it?

PR excerpt (80–120 words, Miklos’s own words): When an AI assistant presents an incomplete or misleading account of a company, I diagnose first: what exactly did it say, which alternatives did it mention, does the error concern a product, price, service area, qualification, or reputation? I work backward from the visible result — the answer is the observation, the public information environment is the field to inspect. After intervening, I revisit the target question and keep the before-and-after wording, date, and test conditions. Attribution should be modest: I can improve the accuracy and availability of the information from which systems form an answer.

Miklos Roth — founder of CRS AI Marketing & SEO Agency

© Copyright gdpr adatvédelmi tisztviselő