AI brand misinformation in the United States, what is measured today and what nobody has measured yet

Ask an assistant about a company and it will tell you things. Some of them will be wrong. The question every US business owner actually asks is how often, and the honest answer in October 2026 is that the rate is not published anywhere with a method you can check.

This page does three things. It states the figure the subject needs and says plainly that we do not have it. It gives the measurement we do hold, which is about naming and citation rather than accuracy. And it sets out, in full, the study design that would produce a real US error rate, so that anybody who runs it can be held to it, including us.

The number this subject needs

A usable headline for AI brand misinformation is a verified material-error rate per 1,000 evaluated business claims, carrying four things with it at all times, the total number of claims tested, which platforms and model versions were covered, the dates of testing, and a confidence interval.

The denominator in that sentence is claims, not answers. This matters more than it looks. One answer about a company can contain no checkable claim at all, or twenty. Counting errors per answer produces a number that moves when question wording changes and tells a reader nothing. Counting per claim produces a number that can be compared between platforms and across months.

We do not hold a material-error rate. Our measurement programme has never scored a business claim for truth. Printing a rate would mean inventing one, and a fabricated accuracy figure about accuracy is a special kind of worthless.

What we did measure, and what it cannot tell you

On 22 October 2026 we scored 635 readings across two engines, ChatGPT and Gemini, taken through their product surfaces by API. The questions were 324 US buyer questions grouped into 19 batches. Of the 635 readings, 550 ran a live web search and the rest answered from what the model already held. Every citation in those answers was collected, which came to 19,793 citation URLs resolving to 3,307 distinct hosts.

That dataset answers who gets named and which sources get pulled. In 451 of the 635 readings a rival was named and we were not. In 184 of them no company was named at all. Across a roster of 24 vendors there were 1,053 namings, counted once per reading in which a name appeared rather than once per occurrence. Our own domain appeared in 0 of the 19,793 citation URLs and in none of the 3,307 hosts.

None of that is an accuracy measurement. A reading where an assistant names Semrush, which it did in 259 readings, is evidence about recommendation share. It says nothing at all about whether the sentences around that name were accurate. Visibility and accuracy are two different studies with two different denominators, and the quickest way to get a brand-misinformation page wrong is to borrow a number from one and spend it on the other.

The five things that get called an error and are not the same thing

Before a rate can exist, the taxonomy has to exist, and this is the part a business can use today without any study at all. Five categories, kept separate.

A factual error

A statement that was never true. Wrong founder, wrong headquarters, a product you have never sold, a certification you do not hold. This is the only category where the word false is safe.

An outdated fact

A statement that was true and is not now. Last year's price, a discontinued plan, an office you closed. The assistant is not wrong about the past, and the fix is almost never a report. It is a dated page on your own site saying the current thing clearly.

An unsupported assertion

A claim with no source behind it, favourable or unfavourable. These are the ones worth logging most carefully, because they tend to appear and disappear between runs.

A subjective opinion

A judgement about quality, value or fit. Not an error, however much it stings, and a correction request built on one gets nowhere.

A misleading omission

Everything said is true and the shape of the answer is wrong anyway, usually because the assistant answered about a competitor's strength and never reached yours. This is the hardest category to score and the most common complaint.

What a real US error-rate study would have to contain

We put this question to the assistants themselves as part of the audit, and the design they asked for is more demanding than anything currently published. Reproduced here as a standard, not as a claim that we have met it.

• A sampling frame of around 500 US businesses, stratified across at least ten industries and more than one company size, with the distribution disclosed.
• At least 20 neutral prompts per business, which gives a minimum of 10,000 business-prompt pairs. Neutral means the prompt does not contain the answer, and the wording is published.
• At least three separately identified assistants or search surfaces, with exact model and version detail where the product exposes it, and with logged-in state, location and whether live search ran recorded per reading.
• At least two independent human reviewers per sampled claim, a reported agreement statistic, and a written procedure for resolving disagreement.
• The five-way taxonomy above, applied before results are counted rather than after.
• Error counts and rates with denominators broken out by platform, industry, severity and testing date.
• A longitudinal subset retested after 7, 30 and 90 days, so persistence can be reported separately from incidence.
• Public method, anonymised claim-level data, prompt templates, exclusion rules, and a funding and commercial-interest disclosure.

Five hundred businesses is a budget, not a threshold. A ten-business pilot is citable if it is labelled exploratory and its denominator is printed next to every number. What is not acceptable is presenting a convenience sample as representative of US business, which is the shape most vendor research in this category takes.

What you can run this week instead of waiting

You do not need a 10,000-pair study to find out whether an assistant is wrong about you. You need a fixed prompt set and a log that does not lie.

1. Write between 10 and 20 brand-fact questions whose answers you can prove from your own records. Pricing, leadership, locations, what you sell, what you do not sell, any licence or registration.
2. Ask each one on each surface you care about. Record the date, the product, the account state, and whether the answer searched the web. For ChatGPT and Gemini you can ask the assistant directly which pages it read, and the answer changes how you treat the result.
3. Split every answer into claims. Mark each claim with one of the five categories. Keep the count of claims, because that is your denominator.
4. Prove the correct fact from a dated page you control before you report anything to anybody.
5. Repeat the same prompt set after 7 and 30 days. An error that disappears on its own and an error you fixed look identical unless you held the question still.

Two cautions from our own runs. Repeated identical prompts do not return identical answers, so a single before and a single after is not a measurement of anything. And the US targeting controls on the route we used were accepted and then ignored, measured on our own traffic, so a US reading has to be fixed in the prompt and verified in the answer rather than trusted to a country parameter.

Where our commercial interest sits

We sell AI visibility measurement, so a page about AI getting your facts wrong is published by a company that benefits from your concern. The specifics, so you can discount them properly.

We read five surfaces, not nine. Our pricing page, which runs to 119,203 bytes, renders no SSO, no SAML, no SOC 2, no ISO, no DPA and no white label, checked by word boundary over the whole document. A loose search for sso on that page matches crossOrigin seven times and subprocessors twice, which is how this gets reported wrongly in both directions. The published US ladder starts at 149 dollars for SEO Start, which reads 0 questions, 0 credits and 0 engines, so it is not an AI visibility plan at any price. The cheapest plan that reads all five surfaces is Growth at 299 dollars. Scale is 599, SEO Scale 699, Inbound 1099, Outbound 1299, Agency 1299 and Everything 1799. A Claude check costs 3 credits rather than 1.

That page also contradicts itself three ways, and both readings are published here rather than the flattering one. The Agency plan is described as covering 10 client brands while the comparison table answers No to more than one business in all eight columns. The main listing names five apps including Claude while the Scale description names four without it. The FAQ says SEO Start gets 12 pages a month and no AI app checks, beside a table that lists five engines for the same plan.

On the competitive side, read 22 October 2026 on tryprofound.com/pricing, Profound publishes SSO and SAML with SOC 2 compliance on its Enterprise plan and up to 9 answer engines tracked, against our five, and publishes no numeric price for it. If a procurement checklist with SOC 2 on it is what decides your purchase, that is a reason to buy them and not us, and we would rather you knew now.

Who should not buy a monitoring product for this problem

If the claim against your business is defamatory, or a customer has already acted on it, monitoring is the wrong first call and a US attorney is the right one. If you have one known error on one surface, fix the source page and retest by hand. If your own site does not state the correct fact on a dated page, no amount of monitoring will help, because there is nothing for a model to find.

Questions businesses ask about wrong AI answers

Can an assistant publish false information about my company

Yes, and it happens in all five categories above. What does not yet exist is a defensible rate for how often, which is the whole point of this page.

How often do assistants get company facts wrong

Unknown at a national scale. Any figure quoted at you without a claim count, a platform list and a date was not measured.

Is a wrong answer evidence that the model is biased against me

No. In our own 635 readings, 184 named no company at all. Absence is usually a retrieval and corroboration result, not a judgement.

Should I report the answer or fix my site first

Fix the source first and prove it with a dated page. A report about a claim your own site still contradicts is unlikely to go anywhere.