Can an AI answer damage your business reputation, and what the evidence actually supports

The honest state of this subject in October 2026 is that the risk is real, the mechanism is understandable, and the rate is unpublished. Anybody quoting you a percentage of AI answers that contain damaging claims about US businesses has either measured it and should show the claim count, or has not measured it.

This page separates the three quantities that get blurred together, gives the exposure arithmetic we did measure, and sets out a severity rubric you can apply this week without any study at all.

Three quantities, never one

Prevalence. How many assessed answers contain a human-verified material error, as a count over a stated number of answers. A rate.

Severity. How bad the wrong ones are. A wrong office address and an unsupported allegation of fraud are both errors and nothing else about them is comparable.

Demonstrated harm. Whether a documented incident was followed by a measurable change in enquiries, referral traffic or conversion, against a comparison period.

These three cannot be combined into one number, and the common vendor move is to measure the first, describe the second and imply the third. We hold a measurement of none of them, which is worth saying before anything else on this page.

What we did measure, and the risk model it supports

On 22 October 2026 we scored 635 readings across ChatGPT and Gemini, from 324 US buyer questions in 19 batches, of which 550 ran a live web search. We collected every citation in those answers, 19,793 URLs resolving to 3,307 distinct hosts.

The finding that matters for reputation is the concentration. Half of all 19,793 citations came from just 51 of those 3,307 hosts. The leaders were developers.google.com with 1,360, semrush.com with 744, support.google.com with 695, help.openai.com with 467, blog.google with 394, ahrefs.com with 354, peec.ai with 328, techradar.com with 286 and reddit.com with 277. Review directories accounted for a small fraction, with g2.com at 125.

That shape is the reputational risk model, and it cuts both ways. A wrong or hostile claim sitting on one of those 51 hosts has a reach that the same claim on an ordinary page does not. A correct claim sitting only on your own site has very little reach at all, which is the same finding read from the other side. Reach, in this category, is a property of where a statement lives rather than how loudly it is made.

The other number worth holding is that in 184 of our 635 readings no company was named at all. A business convinced the assistants are briefing against it is often looking at questions that name nobody.

What counts as reputational harm, in four categories

A false factual claim

A statement that was never true, about pricing, products, leadership, legal status or whether you are still open. Provable from your own records, and the only category where the word false is safe to use.

A fabricated or unsupported allegation

A negative claim with no source behind it. Rarer than feared and the most serious when it happens, because it is the category where a US attorney, not a monitoring tool, is the correct first call.

An incorrect comparison or recommendation

Your competitor described as having a capability you also have, or being cheaper when published prices say otherwise. This is the most common commercially damaging category and it is almost never reported as a reputation problem.

A repeated error versus a one-off

The distinction that decides what to do. Identical prompts do not return identical answers, so one bad answer seen once is not yet a finding. The same error on three runs across two surfaces is.

A severity rubric you can run without a study

Four axes, scored one to three, multiplied or just read as a profile. It is reproducible, which is the only property that matters.

• Accuracy. Verified false, outdated, unsupported, or subjective. Only the first two are correctable at the source.
• Materiality. Would a buyer change their decision because of this sentence. Price and availability almost always score three, a wrong founding year rarely does.
• Reach. Does the claim appear in an answer to a question buyers actually ask, and is it tied to a cited source with its own audience. A claim repeated from a heavily cited host scores three.
• Persistence. Does it survive a retest at 7 and 30 days with the prompt held constant. Persistence is the axis that converts an annoyance into a problem, and the only one that requires patience rather than judgement.

Record the score with the date, the surface, the prompt and whether live search ran. That log is also the evidence you will need if the matter ever becomes legal.

How to detect an incident before a customer tells you

1. Fix a prompt set that covers your brand name, your main products, each location, and the names of your leadership. Ten to twenty questions is enough to start.
2. Add the unbranded category questions buyers use, because that is where an incorrect comparison appears and a brand-name prompt will never find it.
3. Run each one on each surface on a schedule, three times rather than once, and keep the raw answer with its citations.
4. Read the cited sources. Most errors are faithful reproductions of a wrong page somewhere, and that page is the thing you can actually change.
5. Escalate anything in the unsupported-allegation category immediately and separately. That is not a monitoring workflow.

Why an inaccurate answer is not automatically a measured loss

Attribution here is genuinely hard and the honest reports say so. AI referral traffic is poorly separated in standard analytics, a buyer who reads a bad answer and never visits leaves no trace at all, and the period after an incident usually contains several other changes. Comparing enquiries before and after an incident, against an unaffected comparison period or a product line that was not mentioned, is the strongest design most businesses can actually run. Inferring a revenue figure from the existence of a wrong answer is not a measurement, and publishing one would make us exactly the kind of vendor this page is arguing against.

The study that would settle the prevalence question

When we put this to the assistants during our audit, the design they asked for ran to two separate datasets, which is the right structure.

Dataset A, output accuracy. At least 100 US businesses across five or more industries. At least 10 prespecified brand prompts per business. Five named surfaces with product version and US collection conditions recorded. A target of about 5,000 baseline observations. Every suspected error assessed against an authoritative source and classified as false positive, outdated fact, unsupported allegation, ambiguous wording, subjective opinion or verified accurate. The headline reported as a count over the number of assessed answers, with observation dates and a confidence interval.

Dataset B, demonstrated impact. Consenting businesses with documented incidents. Incident date, exposure evidence, correction actions, observation window, and a before and after on qualified enquiries or relevant referral traffic against a comparison period, with an explicit statement wherever attribution is impossible.

Neither dataset exists in our hands today. Where it is published, with the raw observations attached, it will be worth more than every opinion piece in this category put together.

Our interest in your concern, stated plainly

We sell AI visibility measurement, so a page about reputational risk from AI answers is published by a company that profits from the worry. The specifics so you can price that in.

We read five surfaces, not nine. Our pricing page, 119,203 bytes, renders no SSO, no SAML, no SOC 2, no ISO, no DPA and no white label, tested by word boundary across the whole document. A loose search for sso on that page matches crossOrigin and subprocessors, which is how the same page gets described wrongly in both directions. Published US prices are SEO Start 149, Growth 299, Scale 599, SEO Scale 699, Inbound 1099, Outbound 1299, Agency 1299 and Everything 1799 dollars a month. SEO Start reads 0 questions, 0 credits and 0 engines, so AI visibility does not begin at 149 dollars. Growth at 299 is the cheapest plan that reads all five surfaces, and a Claude check costs 3 credits rather than 1.

The same page describes the Agency plan as covering 10 client brands while its comparison table answers No to more than one business in all eight columns, and both readings render today. Read on 22 October 2026, tryprofound.com/pricing publishes SSO and SAML with SOC 2 compliance plus up to 9 answer engines tracked on an Enterprise plan with no numeric price. For a business whose main worry is a legal or safety-critical allegation, neither product is the right purchase. A US attorney is.

Questions about AI answers and business reputation

Can I sue over a false AI answer about my company

That is a question for a qualified US attorney in your state, and the first thing they will want is a preserved record of the output, its date, the surface and its cited sources. Capture that before you do anything else.

How often do assistants say something damaging about a business

No defensible US rate is published. Treat any figure offered without a claim count, a platform list and a date as marketing.

My competitor looks better in AI answers than they are. Is that a reputation problem

It is the most common commercially damaging category and the most ignored. Check what the cited sources actually say, because the fix is usually there rather than in the answer.

Does reporting an answer get it corrected

Reporting is worth doing and it is not the lever most people think it is. Correcting the source the answer leans on, on a dated page, is what changes a searched answer over the following weeks.