How Fidenta Verify compares
There are a lot of ways to check an AI's answer. Most of them ask you to copy the answer somewhere else. Here is how the common approaches differ — and the questions worth asking of any of them, including ours.
The four questions
1. Does it work where you already are?
If checking an answer means copying it into another tab, you will stop doing it by Thursday. The check has to be in the conversation.
2. Is the check actually independent?
A model reviewing its own output is not a second opinion. Independence means different evidence and a different model lineage — otherwise you inherit the same blind spots.
3. Can you see the evidence?
A confidence score you cannot audit is just a second opinion with a number attached. You should be able to open the sources and disagree.
4. Does it distinguish "false" from "not found"?
The most consequential distinction in the category, and the easiest to get wrong. A tool that collapses them will eventually call something true a lie.
Approaches side by side
| Fidenta Verify | Other verifier connectors |
Ask the model to re-check itself |
Browser fact- checker apps |
Fact-check inside a writing tool |
Free hallucination flaggers |
|
|---|---|---|---|---|---|---|
| Runs inside your assistant | Yes — a connector | Yes | Yes | No — separate app | Only in their editor | No — separate app |
| Evidence from outside the model | Two independent retrievers | Usually one search leg | None | Varies | Varies | Typically none |
| Verdict audited by a different lab's model | Yes | Not that we have seen | No — same model | Rarely disclosed | Rarely disclosed | No |
| Shows the sources behind each verdict | Yes | Usually yes | Sometimes | Varies | Varies | Gives you a list to check yourself |
| Separates “contradicted” from “not found” | Four distinct verdicts | Some do | No | Varies | Varies | Flags risk, does not rule |
| Re-checking a claim you already checked | Free while still current | Usually charged again | Costs a new turn | Counts against your quota again | Counts again | Free, but nothing was checked |
| Priced by | Claims needing a fresh check | Per call | n/a | Scans per day or month | Bundled seat | n/a |
| Cost shown before it runs | Yes — quoted per claim | Rarely | n/a | Subscription | Bundled | Free |
On “other verifier connectors”: there are now a good number of these, several of them decent, and some cost a fraction of what we do. If price per call is the deciding factor, one of them is probably your answer. What we have not found in that group is a second model from a different lab auditing the verdict, or a price shown before anything is spent. Judge that trade for yourself.
“Varies” means it differs by product and is not consistently disclosed — check the specific tool rather than trusting a table, including this one. Claims in the Fidenta column describe our own product and are verifiable in the documentation.
Flagging is not verifying
Worth checking on any tool you consider, because the two get sold under the same words. Some tools tell you which sentences look risky and leave the checking to you. That is a triage tool, and a useful one — but you still have to go and settle every flag yourself, which is the work you were trying to avoid.
The tell is in the vendor's own words. Look for a line saying the results are a starting point for your own research, or that a flag does not mean something is wrong. Several tools in this category say exactly that, on the same page as the demo, and they are being honest when they do. It just means the job is not finished when the tool stops.
Fidenta returns a verdict per claim — supported, contradicted, unsupported, or unverifiable — with the sources it used and an audit by a model from a different lab. You can disagree with it, which is the point. But you are disagreeing with a conclusion, not receiving a to-do list.
On the numbers tools publish about themselves
Accuracy percentages are common in this category and almost never come with a method. A figure like "94% accurate" tells you nothing without knowing what was measured, on which claims, and against whose ground truth — and that applies to any figure we might put here too.
So the useful question to ask a vendor, including us, is not what is your accuracy but how did you measure it and can I see the test set. Ours is a scope test of 18 cases run three times each; the case that matters most to us is that a true claim is never called false, and we track it separately because it is the failure that costs a customer their credibility. We would rather tell you that than print a percentage.
Where we would not pick us
Objections worth raising
Because a model grading its own answer tends to agree with itself. It re-reads what it already believes, finds it plausible, and tells you so — with the same confidence it had the first time.
Fidenta pulls evidence from two independent retrievers, assesses the claim strictly against what came back, then has a model from a different lab audit whether that verdict is actually carried by the evidence. Different training data, different failure modes. That is what makes it a check rather than a second opinion from the same source.
Some free tools are worth using. Read what they claim carefully, though: several describe themselves as flagging risk and producing a list of things for you to go verify — which is a to-do list, not a verification.
The question to ask any tool, free or paid: does it show you the actual sources behind each verdict, and does it distinguish "I found evidence this is false" from "I searched and found nothing"? If it collapses those two, it will eventually tell you something true is false.
Those check what you have already written, in their editor, after the fact. Fidenta checks what your assistant just told you, in the conversation, before you have pasted it anywhere.
That ordering matters more than it sounds. The expensive error is the one that makes it into the document, the proposal or the client email — not the one you catch while editing.
You pay for what you check. A sentence costs one credit; an article might cost fifteen, and you are told the number before anything is charged. Credits do not expire.
Watch out for per-word pricing anywhere you see it — it means your longest and most important documents cost the most to verify, which is backwards.
You are not charged twice. Verifications are saved to your account, and asking again
returns the saved result for free. Pass force when you want a fresh check
against current sources — useful when the underlying facts may have moved.
It says so, and marks the claim unsupported — explicitly not the same as false. This is the single most important behaviour in the product and the one most often got wrong elsewhere. Treating absence of evidence as evidence of absence is how a checker starts inventing failures that were never there.
40 free claim checks. No card, two minutes to connect.
Start free