How we verify citations
The accuracy claim stated precisely: what each model does, when the second one runs, and what happens when they disagree.
This is the page to send a sceptical client or a procurement team. It states what the pipeline does, without rounding it up.
What you'll learn
- What each model does, and when
- What happens on disagreement
- What the confidence score means, and what it does not
The claim, precisely
- Answers with no signal of your business at all are recorded as not cited without a model call. There is nothing to weigh.
- Every remaining answer is read by a judge model, which decides whether you are genuinely named and captures the sentence that says so.
- Where the judge is not certain, an independent verifier re-reads the original answer and reaches its own verdict.
- If the two disagree, the citation is not counted.
- If the judge was certain and no verifier ran, the check is recorded as decided by one model and keeps the judge's own confidence. It does not inherit the higher confidence an agreement would have earned.
That fifth point is the one most tools would leave out. It is in the data, on the record, and in this article for the same reason: a claim you cannot check is not a claim worth making.
What is excluded, and why
| Excluded | Reason |
|---|---|
| A generic term that matches your name | The material is not the company |
| A mention that recommends against you | Being named as a warning is not a citation. It is recorded as a negative mention |
| A citation the two models dispute | Fails closed |
| An absent AI Overview | Google drew no panel, so there was no answer to be cited in |
What the confidence score measures
How sure the models are that the citation is real and substantive. On an agreement it is the average of the two. On a single-model check it is the judge's own.
It is not a quality score, not a ranking, and not a statement about the future. See how a citation is counted.
Why rates differ across surfaces
The engines are built differently. Some retrieve live pages before answering, some lean on what they were trained on, and one is a panel on a results page rather than a chat model at all. A spread between surfaces is the normal state, not an error, which is why each surface is reported on its own row rather than averaged.
What we do not publish
An industry-wide citation rate, or a table of per-engine averages presented as fact. Nothing in the product produces those, and a benchmark nobody can source is worse than no benchmark.
Check us
Ask for the sentences and read one against the engine yourself. If you find a counted citation you disagree with, send it to support@outercite.com with the workspace and the date. Disagreements about individual judgements are how the prompts behind the pipeline get better.
Try this in Outercite
Ask the assistant "Show me the exact sentences AI wrote about us", pick the lowest-confidence row, and decide for yourself whether you would have counted it.
Related
How verification works
The same pipeline, for a reader who is new to it.
Why fewer citations than a manual search?
Why the strict definition produces a smaller, more useful number.
Trust overview
The three commitments.
Last updated on