Ask fourteen frontier models the same question and you will get fourteen different relationships to the source material. Some quote. Some paraphrase to the edge of recognition. Some assert your finding as settled fact with no attribution at all. If we are going to optimise for citation, we first have to classify it.
Four behaviors
Across roughly 44,000 model responses, citation behavior sorted cleanly into four classes:
- Verbatim cite — the model reproduces a span and attributes it. Rare, high-value.
- Paraphrase-with-source — the claim is restated but the source survives.
- Silent absorption — the claim is reproduced as fact; the source is gone.
- Contradiction — the model surfaces a competing claim and yours loses.
The distribution is not uniform across vendors, and — more usefully — it is not uniform across document types. The same claim, published as research prose versus as a marketing page, lands in different classes at very different rates.
The durability premium
The headline result: statements written as defensible research — hedged, sourced, quantified — are roughly 3.4× more likely to be attributed than the identical fact written as a confident marketing assertion. The model treats epistemic humility as a trust signal. This inverts twenty years of conversion-copy instinct, and it is the single most actionable finding in the audit.
Methodology, briefly
The corpus is controlled: 3,200 documents, six verticals, matched claims across document types, fourteen models queried under fixed decoding parameters, three runs each to measure reproducibility. Full method and the prompt set are in the methodology note; the dataset is available on request. Everything here should be replicable — that is the point.
