How a citation is counted
What has to be true before a check becomes a counted citation, what the confidence score measures, and how to use it in a decision.
Every citation in your workspace carries a record of how it was established: which models looked at it, whether they agreed, and how confident they were. This is how to read that record.
What you'll learn
- The four ways a check can be resolved, and which ones count as cited
- What the confidence score is, and what it is not
- How to use it without over-reading it
The four resolutions
| How it was resolved | What happened | Counted as cited |
|---|---|---|
| Two models agree | The judge and an independent verifier reached the same verdict. Confidence is the average of the two | Yes, if both said cited |
| The models disagree | They reached different verdicts. Confidence is the lower of the two | No |
| Judged by one model | The judge was confident enough that no verifier was needed. Confidence is the judge's own | Yes, if the judge said cited |
| No signal | The answer contained nothing that could be your business, so no model was asked | No |
Two things follow, and they are the whole reason this table exists.
A disagreement fails closed. When two models do not agree, the citation is not counted. The platform would rather understate you than hand you a number that falls apart when somebody reads the answer.
A single-model judgement is labelled as one. It keeps the judge's own confidence and does not borrow the higher confidence that an agreement would have earned. One model's opinion is one model's opinion, and the record says so.
This is worth knowing before you quote an accuracy claim to a client. The honest sentence is: a judge reads every answer, an independent verifier audits the ones the judge is not certain about, and a disagreement means the citation is not counted.
What the confidence score is
How sure the models are that the citation is real and substantive.
- High. Clear, unambiguous, substantive.
- Middle. Genuine, with something uncertain about the context or prominence.
- Low. Present but weak. Read the sentence yourself in Proof.
What it is not
- Not a quality score. A high-confidence citation can be a passing mention in an answer that recommends somebody else more warmly.
- Not a ranking. Where you sat in a list is a separate field, and it only exists when the answer made a list.
- Not a promise about the future. It describes one answer on one day on one surface.
Fields that are often empty, and legitimately so
Not every citation carries every field. Position exists only when the answer ranked things. Sentiment exists only when the answer expressed one. Descriptive phrases exist only when the answer used them. An empty field means the answer did not contain one, so do not read it as a zero, and do not build a report that assumes every citation has a position.
How to use it in practice
- Filter low-confidence rows out of a client report, and say that you did.
- Read the low-confidence rows yourself. They are where the interesting edge cases live: near-misses, ambiguous product naming, mentions of a subsidiary.
- Never average confidence into a headline metric. It is a property of one check, not a score for your brand.
Try this in Outercite
Ask the assistant "Show me the exact sentences AI wrote about us." Then compare a high-confidence row with a low-confidence one and see whether you agree with the machine.
Related
How verification works
The pipeline end to end.
Why fewer citations than a manual search?
Why the strict definition gives you a smaller and more useful number.
Data definitions
What counts as a check, and what counts in a rate.
Last updated on
What is a citation?
A citation is your business being named inside the answer an AI engine writes, whether by name, by URL, by phone number or by product.
The six-level stack
The six levels that decide whether AI names you: four on your site, two in the loop you run. They fail bottom up, so fix them in order.