Severity, per model
Findings are graded critical or watch, and attributed to the model that produced them. Models disagree with each other far more than most teams expect.
Brand Risk Red Team · New
We ask AI the questions your buyers ask behind your back
Your visibility tracking covers the questions you want to win. This covers the ones you cannot afford to lose, and shows you exactly which sources the models leaned on when they answered.
The questions nobody asks you
Every buyer runs this test. You just never see the results.
Before a shortlist gets to a call, someone types your name into an assistant followed by a word you would rather they did not. That answer is formed from sources you may never have looked at, and it is influencing deals you will never know you lost.
The kind of prompt we run
- “Is [brand] a scam?”
- “What are the biggest complaints about [brand]?”
- “Why do customers leave [brand]?”
- “Is [brand] safe to use with sensitive data?”
- “[brand] vs [competitor]: which should I avoid?”
- “Has [brand] had any security incidents?”
Prompts are generated for your category and competitive set, across trust, complaint and comparison themes.
What a finding contains
The answer, and the sources behind the answer
Knowing a model said something damaging is only half of it. The half you can act on is where it was looking when it said so.
Cited sources
The URLs each model referenced. These show influence, not truth. A cited page is where the claim came from, so you have something to work on.
One exposure score
A 0–100 summary with critical and watch counts, so you can tell whether exposure is improving between runs rather than re-reading every finding.
Monthly, plus on demand
Scheduled runs give you a trend. Trigger an extra run after a news cycle, an incident, or a competitor campaign.
From finding to fix
A cited source is a source you can do something about
A model repeating a damaging claim is usually leaning on a specific thread, review or article. That is a target, and it connects straight into the off-page reputation workflow.
- 1
Test
Adversarial prompts run across multiple models on a schedule, capturing responses and citations.
- 2
Trace
Each finding names the sources the model used, so you can see what is shaping the answer.
- 3
Respond
Address the source, or publish the better answer for the model to reach for next time.
The product does not decide whether a claim about you is fair. It reports what was said and what was cited. Judging it is your call, but you cannot judge an answer you have never seen.
Where it fits
Reputation monitoring watches the sources. This watches the answers.
Off-page monitoring tells you what people are saying about you on Reddit, LinkedIn and Quora. The red team tells you what the models have already concluded from it. You want both, because a quiet forum does not guarantee a clean answer.
Runs happen in the background and can take a few minutes, since several prompts run across several models is real work, not a cached lookup.
Frequently Asked Questions
Brand risk testing, explained
A recurring adversarial test of how AI models answer the hostile questions buyers ask in private: is this company a scam, what are the complaints about them, why do people leave them, who is better. It runs those prompts across multiple models and captures what each one said, along with the URLs it cited.
Because those are the questions that decide deals, and nobody asks them where you can see. Your visibility tracking covers the questions you want to win. This covers the questions you cannot afford to lose, and the answers are frequently very different.
The prompt that was asked, which model answered, the risk category, an excerpt of the response, and the source URLs that model cited. The citations show where the model was looking. They are evidence of influence, not confirmation that the claim is true.
A single 0–100 summary of how much reputational risk the test set surfaced, alongside counts of critical findings and items to watch. It is a tracking number for whether exposure is improving between runs, not a verdict on your company.
Monthly by default, so you have a trend rather than a snapshot, and you can trigger a run on demand, for example after a bad news cycle, a product incident, or a competitor campaign.
Start with the cited sources. A model repeating a damaging claim is usually leaning on a specific page, thread or review, and that source is addressable: through the off-page reputation workflow, a correction at the source, or content that gives the model a better answer to reach for.
Not necessarily, and the product does not pretend to adjudicate. It reports what was said and what was cited. Whether the claim is fair is your call, but you cannot make that call about an answer you have never seen.
Find out what AI says about you when you are not in the room
Start with a free audit of how AI engines describe your brand today.
Get a free visibility audit
