Skip to main content
An answer can name you, recommend you, and still get your price wrong. Attensira pulls the checkable statements out of every answer and checks the ones about you against the facts you confirmed. A wrong one is flagged once, and an automation can act on it.

​
What a claim is

A claim is one factual statement an answer makes about a brand: what it costs, what it does or lacks, what it integrates with, a certification, where it is available, or a fact about the company. Opinions are not claims. “Acme starts at $49 per seat” is a claim; “Acme is the easiest to use” is a stance, and is scored by sentiment instead. Every claim is a receipt, not a paraphrase:
  • The quote is the answer’s own words. A claim is kept only when its quote appears in the answer’s text, with spacing normalized. A statement reworded by the model that reads the answer is dropped.
  • It names who it is about. You (your brand, or a part of your brand family you confirmed), a tracked competitor, or a domain the answer cited. A claim about a brand that resolves to none of these is not kept, because nothing could ever check it.
  • It carries its sources. The links the answer attached to the sentence the claim is in, when it attached any.
Claims are pulled from an answer in the same read as its brand order, for answers from the last 90 days. An answer that has not been read yet has no claims list, which is different from a list that is empty: an empty list means the answer was read and made no claim we could check.

​
Which facts a claim is checked against

Only claims about you are checked, and only against facts a person confirmed in your brand profile:
  • fields in What you sell, Differentiators and Proof, plus any field whose name is about pricing, wherever it is filed;
  • that someone on the workspace confirmed. A field Attensira drafted from your site is our guess, not your word, and is never used;
  • that are not stale. A confirmed field whose source page has changed since you confirmed it may no longer be true, and is set aside until it is confirmed again.
With no confirmed fact to check against, nothing is asked and every claim about you is not_covered. Confirm your pricing, your integrations and your certifications first: those are the facts an answer gets wrong in a way that costs a sale.

​
The verdict

Each claim about you gets one of three verdicts: A claim not checked yet shows Not checked yet. Nothing to check against is not a wrong claim, and it is not a right one either. Your facts are silent on it, and silence is not a contradiction. If a claim keeps landing here and it matters, confirm the fact that settles it. The verdict is a decision, made with a confidence. A contradiction is only treated as a wrong claim when that confidence is at least 80%. Below it, the verdict is kept with its confidence and nothing is raised.

​
When claims are checked

The check runs every night at 04:15 UTC, newest answers first, over the claims whose verdict is missing or out of date. A verdict goes out of date when:
  • the answer is read again and its claims change;
  • you confirm, edit or un-confirm a fact. Every claim checked against the old facts is checked again against the new ones, starting the next night and newest answers first. You never re-run anything by hand.
When a workspace is left with no confirmed fact to check against, every verdict is settled to “Nothing to check against” without asking again, so a contradiction of a fact nobody stands behind any more does not stay on the list.

​
The claim_wrong event

A wrong claim raises one claim_wrong event. It is raised once per distinct wrong statement against a fact: the same sentence, said by three engines on ten days, is one event. It is raised again only if the fact it contradicts is edited, because then it is a different question. The check only flags. It never edits a fact, never files work and never approves anything. What happens next is up to whoever listens for the event.

​
Acting on it with an automation

claim_wrong is recorded in the same event log as the other event types, but it is not one you can pick yet: the dashboard, MCP and the REST API offer only the five listed there, and the built-in Event responder does not listen for it. Only an event automation saved with an empty event list, which listens for every type, acts on it. Such a run’s cause reads: “Because an AI answer said something about you that your confirmed company facts contradict.” The event carries what a run needs without reading anything else: When an automation does act on it, the event limits apply as for every event: at most 10 events per run and 3 runs per automation per rolling 24 hours, and events older than 48 hours are not acted on.

​
Where you see claims

  • Over MCP and the REST API, answer_history and answer_slice rows carry each answer’s extracted claims, each with kind, subject (self for you, otherwise a domain), quote and citation_positions. null means the answer has not been read for claims yet; [] means it was read and made none. Verdicts are not on the answer rows.
  • query_evidence counts wrong claims with the wrong_claims metric, by day or week, prompt field, engine and surface. See MCP tools.
  • In the dashboard, the prompt page’s claims section and the analytics overview’s wrong-claims list do not show verdicts yet. Each says so in place of a list, rather than showing an empty one that would read as “nothing wrong”.

​
What it gets wrong

The verdict is a judgement a model makes against the facts you confirmed, and it can be wrong. It errs toward silence: a contradiction it is unsure about is kept, not raised. It can only be as right as your facts: a confirmed fact that is out of date makes a correct answer look wrong. Keep the facts in your brand profile current, and read the quote before you act on a flag.