AI states your facts with confidence. Some of them are wrong.
Hallucination Detection for AI Search
Hallucination detection checks every AI answer in your audit against the facts in your brand profile: your price, founding year, location, CEO, ownership and features. When an answer contradicts one, you see what the AI said, what is true, and how serious it is.
99Visibility is an AI SEO tool built around a question most reports skip: is what AI says about you true? A wrong price in an AI answer is read by a buyer who may never reach your pricing page to see the right one.
Detection runs on every answer in every audit, on every platform in your plan. What it finds feeds the accuracy part of your score (30%) and a fix that quotes the sentence the error came from.
Accuracy is a rate, not a tally. The same error rate scores the same on a ten-prompt audit and a hundred-prompt one, and the single worst flag still pulls the number down, so one serious error is not averaged away.
Errors cluster around change. In our audit of four sportswear brands, every answer to "who is the CEO" named the right person, and the wrong statements were about a sale and a start date: facts with a before and an after.
Accuracy Issues (2)
1 moderate
AI claims
$89/month
Actual fact
$120/month
The AI stated $89/month but the actual Pro plan price is $120/month
AI claims
founded in 2015
Actual fact
2018
The AI stated the company was founded in 2015, but it was actually 2018
AI quotes the wrong price for Acme
What it checks
Prices
A dollar price in the answer that does not match the price you entered. Small differences, like rounding, are ignored. A wrong price is marked moderate.
Founding year
A year given for when your company was founded, launched or established that differs from the year you entered. It is read only when the sentence is about your company, not a competitor or a product line, and a wrong year is marked minor.
Location
Where the answer says you are headquartered or based, checked against a fact keyed "Location" or "Headquarters". A wrong location is marked moderate.
Leadership
Who the answer says runs the company now, checked against your CEO fact: a former CEO on your record named as current is marked outdated, anyone else incorrect, both moderate. If the fact carries a start date, a wrong start year is marked minor.
Features
An answer that says you lack a feature you listed, in phrases like "does not offer" or "lacks". Marked moderate.
Ownership
An answer that says you still own something your facts record as sold is marked outdated and moderate. A claim that you were acquired, which your facts contradict, is the only flag marked critical; an acquisition simply not on file is marked minor, as unverifiable, and is left out of the accuracy score.
Headcount
A specific employee count when you have none on file. We cannot check it, so it is marked minor and the flag says it could not be verified.
How it works
- 1
Enter your facts
Add them in onboarding or later in brand settings: a price, a founding year, a location, your CEO, anything you have sold, the features buyers ask about. Put a date in the value where it matters, as in a "CEO" fact of "Arthur Hoeld (since 2025-07-01)" or a "Divested" fact of "MyFitnessPal (sold 2020-12-18)", and the check reads it.
- 2
Run an audit
Every answer from every platform in the audit is read against those facts. A claim is compared only when your company is what the sentence is about, so a competitor’s founding year in the same answer is not counted against you.
- 3
Read the flags
Each flag shows what the AI claimed beside your fact, with a severity and a one-line explanation. On Free you see how many issues were found; what each one says is on Starter ($49/mo) and up.
- 4
Fix the source
The fix it triggers tells you to state the correct fact once, in visible text, and remove every copy on your own pages that says otherwise. If the answer cited third-party pages, check those too and ask them to update.
- 5
Hear about the serious ones
When an audit finds a critical flag, we email you. The alert is on by default and can be turned off in notification settings.
What a flag looks like
A brand whose Pro plan costs $120 a month enters that as a pricing fact. One answer says otherwise.
Example output
- AI claims
- $89/month
- Your fact
- $120/month
- Severity
- Moderate
- Type
- Incorrect
- Explanation
- The AI stated $89/month but the actual Pro plan price is $120/month
- Fix it triggers
- "AI quotes the wrong price for [brand]", priority critical, about 30 minutes
Illustrative brand and prices. The wording, severity and fix are what the detector and the recommendation rules produce for a price that does not match.
From a real audit: what Claude got wrong
On 15 October 2026 we audited Nike, Adidas, Puma and Under Armour and checked the answers by hand against primary sources. All 16 answers to "who is the CEO" named the right person. The errors were elsewhere, and both came from Claude (claude-sonnet-5) answering from memory with web access off.
From a real audit
| What Claude said | What is true | Source | Flag the check raises now |
|---|---|---|---|
| "UA owns popular fitness-tracking apps", naming MyFitnessPal and MapMyRun | Under Armour sold MyFitnessPal in December 2020 and MapMyFitness, which includes MapMyRun, in August 2024 | Under Armour’s SEC filing; Outside Inc.’s release | Outdated, moderate |
| "Arthur Hoeld took over as CEO in 2023", and 2024 in a second answer | He became PUMA’s CEO in July 2025 | PUMA’s management page | Incorrect, minor, in each answer |
Neither error was flagged when the audit ran: we found both by hand. The accuracy check was fixed afterwards and now compares CEO and ownership claims with the brand facts you enter; re-run on the saved answers, it raises the flags in the last column.
Claude also knew better. Asked who runs Under Armour, the same model wrote that MyFitnessPal "was sold in 2020", so one correct answer proves little about the next.
Answers recorded on the models the product used that day; Google AI Overviews was not part of the run. The fixed check sees errors like these only when the CEO and the sale are in your brand facts.
Read the AI visibility audit of Nike, Adidas, Puma and Under ArmourFix it with Agent V
Agent V is the assistant beside your dashboard. When an answer gets a fact wrong, ask V to fix it: it drafts the page copy that states the right fact in its first sentence, ready to paste on the page that should carry it.
It writes only facts you gave us or that came back as the correct fact in a flag, and it does not name the engine that got it wrong, because the page is for your buyers.
Anything that changes your account, like running an audit or adding a keyword, comes to you as a card to confirm. Nothing changes until you do, and V never edits your site. It runs on Anthropic's Claude models and is on every plan.
Who it is for
- Brands that changed a price, a plan or a product name and need to know which engines kept the old one.
- Teams hearing "I read that you cost…" from prospects and wanting to find where it came from.
- Anyone whose company has been through an acquisition, a sale, a merger, a rename or a change of CEO.
What it does not do
- It checks only the facts you enter. With none on file, accuracy is not checked and stays at 100.
- It has no check for other claim types. A wrong product spec or an invented partnership is not compared with anything, so read the answers as well as the flags.
- It compares dollar prices only. A price in euros or pounds is not checked.
- It does not judge opinions. "Not the best choice for small teams" is a view, not a fact we can check.
- It reads text patterns, so it can miss a wrong fact phrased an unusual way, or flag a price that belongs to another product. Every flag puts the claim beside your fact so you can confirm it before you act.
- It does not change what an engine says. A corrected page reaches an answer given from memory only through a newer model; an answer that searches the web can read it the next time the engine fetches the page.
Questions
Will it catch every wrong thing AI says about my brand?
No. It checks the claim types on this page against the facts you entered. A claim with no matching fact on file has nothing to be compared with.
Why is a wrong price "moderate" and not critical?
Severity describes the claim; priority describes the fix. A wrong price is a moderate flag, but the fix it triggers is ranked critical, because buyers act on a price.
What do I get on the Free plan?
The accuracy score is still calculated and you see how many issues were found. What each flag says, and the step-by-step guide under each fix, start on Starter ($49/mo).
How do I make a wrong fact go away?
State the right fact once, in visible text, on the page an engine would read: usually About, pricing or the product page. Remove every other copy on your site that disagrees, then re-run the audit to see which engines changed.
Does AI still hallucinate in 2026?
Yes. In our 15 October 2026 audit of four sportswear brands, Claude credited Under Armour with apps it sold in 2020 and 2024, and gave two different wrong years for PUMA’s change of CEO. Both came from answers given from memory, which no change to a website reaches until the vendor ships a newer model.
Which AI hallucinates the most?
One audit cannot rank them. In ours, all three wrong statements we confirmed came from Claude, but three are too few to compare platforms, and a model can be right about a fact in one answer and wrong in the next.
How do I check whether AI is wrong about my company?
Ask the platforms about your company and check every factual claim against a primary source: a filing, a press release, your own pricing page. Start with what changed in the last few years, such as prices, leadership, ownership and discontinued products. Enter those facts in your brand profile and every audit repeats the check.
Explore More Features
See What AI Sees
Start on the Free plan. No card needed.