A claim that can lose a deal on its own. Wrong pricing quoted to a buyer, a compliance claim you cannot make, a capability you do not have.
Brand-fact recommendations from the recommendation engine execute here. See how the queue decides what runs here.
Hallucination Detection
Catch the things AI says about you that are simply not true
Models state wrong prices, invent features and name the wrong founder, all with total confidence. Targetlytics compares every answer about your brand against the facts you have declared, and logs each contradiction with its severity, its source and the response it came from.
The problem
A confident wrong answer costs more than no answer
Being left out of an AI recommendation is a missed opportunity. Being described incorrectly is an active liability, and it scales to every person who asks.
Language models fill gaps. When a model has partial information about your pricing, it produces a plausible number rather than declining to answer. When your team page changed two years ago, it names whoever it learned about first. Nothing in the answer signals uncertainty to the person reading it.
You cannot correct what you have not seen, and nobody is going to tell you that ChatGPT quoted a price you retired eighteen months ago. Detection has to be systematic, because the failure is silent by design.
How it works
Ground truth, then comparison
Detection needs something to detect against. That is what the brand constitution is for.
- 1
Declare what is true
Your brand constitution holds the checkable facts about your company: pricing, features, leadership, metrics and the claims you stand behind.
- 2
Read what models say
Every answer from your tracked prompts, across every model, is scanned for specific claims about your brand rather than general sentiment.
- 3
Log every contradiction
Each conflict is stored with the model claim, your ground truth, an explanation of the conflict, a severity, and a link to the response it appeared in.
This is the part most monitoring tools skip. Without a declared ground truth, all you can report is that a model said something about you. With it, you can report that the model said something wrong, and show exactly what the correct answer was.
Triage
Sorted by how much it can cost you
A log that treats every error equally is a log nobody works through.
A claim that misrepresents you in a way a prospect will notice. An outdated feature set, the wrong person named as a founder.
A claim that is wrong but survivable. A stale metric, a rounding error, an old office location.
Claims flagged as widespread appear across more than one model provider. Those are rarely one model misfiring. They usually mean a source on the open web is teaching several models the same wrong thing, which makes the source itself the fix.
Claim types
Grouped so you can see the pattern
Six categories, because models getting your pricing wrong and models getting your team wrong are different problems with different fixes.
Pricing
Retired plans, wrong tiers, invented discounts, a number that was never yours.
Features
Capabilities you do not ship, integrations you do not have, limits stated wrongly.
Leadership
The wrong founder, a departed executive, a role attributed to the wrong person.
Metrics
Customer counts, funding, growth figures and market share stated from stale sources.
Facts
Founding date, headquarters, ownership, certifications and other checkable record.
Other
Anything checkable that does not fit the categories above but still contradicts your record.
Working through it
Every row ends in an action
Open a detection and you get the full response it came from, the model that produced it, how many times it has recurred, when it was last seen, and the possible sources behind the claim. From there the work is ordinary: correct the source, publish the right answer, or take it up with the platform.
Mark it resolved and it leaves your active list without leaving the log. If a model starts repeating a claim you already fixed, it comes back with its occurrence count intact rather than arriving as a brand new problem.
Related
Where this sits next to red teaming
Both protect your brand in AI answers. They answer different questions.
Hallucination Detection
Watches the prompts you already track and catches claims that contradict your declared facts. Continuous, and grounded in something objectively checkable.
Brand Risk Red Team
Deliberately asks the hostile questions you would never track, then reports what came back and which sources the models leaned on. Adversarial, and about reputation rather than fact.
Frequently Asked Questions
Hallucination detection, answered
A specific, checkable claim an AI model made about your brand that contradicts your brand constitution. Not a matter of tone or opinion. Wrong pricing, a feature you do not ship, the wrong name in a leadership role, a metric that is off. Each one is stored with the claim, your ground truth, and the explanation of how the two conflict.
From your brand constitution, the structured record of facts about your company you maintain in Targetlytics: pricing, features, people, metrics and the claims you stand behind. Detection is a comparison against that record, which is why it produces a specific contradiction rather than a general warning that a model might be wrong.
Pricing, features, leadership, metrics, general facts and an other category for anything that does not fit. Each detection is tagged by type, so you can see at a glance whether models are mostly getting your pricing wrong or mostly getting your team wrong. Those two problems have different fixes.
Critical, major or minor, based on how much commercial damage the claim can do. Wrong pricing quoted to a buyer is critical. A slightly stale metric is minor. Severity is carried by a written label as well as colour, so the triage list stays readable for anyone who cannot rely on colour alone.
The same false claim showing up across more than one model provider. That matters because it usually means the error is not one model misfiring, it is a source on the open web that several models have all learned from. Widespread claims are the ones worth chasing to the source.
Each detection records the prompt, the run, the model and the possible sources behind the claim, and links straight to the full response it was found in. That is what turns a false claim into a task, because you can go and correct the page that taught it.
Mark it resolved and it moves out of your active list while staying in the log. Occurrences and last-detected dates keep updating, so if a model starts repeating a claim you already fixed, you see it return rather than assuming it stayed fixed.
Hallucination monitoring and alerts are on the Professional and Agency plans.
Find out what AI is getting wrong about you
A free visibility audit shows how AI engines currently describe your brand, before you connect anything.
Get a free visibility audit
