How We Verify Citations

A citation in your Outercite dashboard means something specific. It is not a keyword match and it is not the output of a single model. It is a result that two independent models agreed on. This page explains why that matters and what it means for the data you see.

What you'll learn

• Why a single-model approach produces unreliable citation data
• How the two-model consensus design keeps false positives near zero
• What the confidence score tells you about each individual citation
• Why a low citation rate on some engines is a feature, not a flaw

The problem with simpler approaches

The most basic way to detect a citation is a keyword search: check whether the AI response contains your brand name. This is fast and cheap, but it is unreliable. An AI response might mention your brand in a completely irrelevant context, in a sentence that says you were not recommended, or in a list of things to avoid. A raw keyword match would count all of these as citations.

A slightly better approach is to use a single model to classify whether a mention is a meaningful citation. This is more accurate than keyword matching, but a single model can be wrong in ways that are hard to predict. It might over-count in certain sentence structures or under-count in others, depending on quirks in its training or architecture.

Outercite uses a two-model consensus design to get past both of these limitations.

The four-stage pipeline

The full pipeline has four stages. The how the pipeline works article covers all four in detail. The trust-relevant stages are stages two, three, and four.

Stage 2: Deep analysis

A high-performance language model reads the full AI response and identifies every business mentioned. It assesses whether each mention is substantive, what the sentiment is (positive, neutral, or negative), and what position the citation holds in the answer. The output is a structured list of citations with metadata.

Stage 3: Cross-verification

A second, independent model receives the original AI response and audits the first model's findings. Critically, it does not see what the first model concluded. It makes its own independent assessment of which citations are present and meaningful.

Stage 4: Consensus

The pipeline compares the two models' outputs. Only citations that both models identify are counted. Where the models fully agree, the confidence score is high. Where there is partial agreement, the confidence score reflects that uncertainty.

What near-zero false positives means in practice

Every counted citation passes through both models and emerges with consensus agreement. Across the prompts we track, brands are cited in about half of all checks (a 48% citation rate), and each of those citations cleared the two-model bar.

Near-zero false positives means: when Outercite says you were cited, you can rely on it. You are not looking at a noisy feed of questionable mentions. You are looking at verified data.

The inverse is equally valuable. When Outercite records a check with no citation, that is accurate data too. It tells you your brand was genuinely absent from that AI response, which is a useful signal for understanding where you need to improve visibility.

What the confidence score tells you

The confidence score on each citation is a 0-to-100 value that reflects the strength of the two-model agreement. Real example values you will see in the dashboard are in the range of 80 to 90.

A high confidence score means both models saw a clear, substantive citation with no ambiguity. A lower (but still positive) score means both models agreed a citation was present, but there was some ambiguity in context, prominence, or framing.

Why citation rates vary across engines

Per-engine citation rates range from 38% (Google AI) to 58% (Perplexity). This variation is real and accurate. It reflects how each engine behaves, not a gap in Outercite's verification.

Google AI's lower citation rate means the verification pipeline is confirming that Google AI cites businesses less frequently, not that Outercite is failing to detect them. The same two-model standard applies to every engine.

Transparency as a design choice

Outercite publishes the pipeline design publicly because transparency is part of the trust model. You do not have to take the accuracy of your data on faith. You can read exactly how the pipeline works, inspect the proof snippet on each citation, and see the confidence score that tells you how clearly both models agreed.

If you have questions about the methodology that go beyond what is covered here, reach out to the team.

Try this in Outercite

Open any citation in your dashboard and inspect the proof snippet and confidence score. The snippet is the actual excerpt from the AI response where your brand was mentioned. The confidence score tells you how clearly both models agreed on the citation.

Related