What is Citation-vs-paraphrase decision?
The citation-vs-paraphrase decision is the per-chunk step the synthesis stage runs over every surviving reranked candidate — does this chunk get rendered as a verbatim quote in the answer (with a numbered source chip and quoted span the user reads as the engine's evidence), or as a paraphrased composition in the engine's voice (with a numbered source chip but no quoted span), or dropped from the rendered answer entirely (source relegated to the 'further sources' panel). The decision is not random — it correlates with five chunk-side properties: claim specificity (numeric statistics survive as verbatim quotes at 2.4× the rate of vague claims), self-containment (chunks whose claim does not require surrounding paragraph context survive verbatim at 1.7×), rationale shape (chunks that open with a citable claim survive verbatim at 1.9× the rate of chunks that open with a topical introduction), source authority (Article schema with author and credentials lifts verbatim survival by 1.3×), and lexical distinctiveness (chunks with phrasing the engine cannot easily compress into the answer's voice survive verbatim at 1.5×). Engineering the decision toward verbatim is the highest-leverage downstream investment a chunk-level rewrite makes.
How it relates to AI UGC
Verbatim text citations on multimodal-active sub-queries pull the engine's image-selection toward the visual asset paired with the cited paragraph — caption-anchored persona-locked imagery surfaces in the carousel at 2.6× the rate of decorative imagery on the same page. The text citation-vs-paraphrase decision and the visual carousel decision are not independent in mid-2026.
Key statistics
- Verbatim-cited chunks earn click-through at 2.3–3.1× the rate of paraphrased chunks at equivalent answer-position because the quoted span anchors the user's eye to the cited source (CTR-by-citation-type audits, 2026).
- Mid-2026 cohort: of chunks surviving rerank, 22–34% are cited verbatim, 28–40% are paraphrased into the engine's voice, and 32–48% are dropped from the rendered answer and relegated to 'further sources' (citation-disposition audits, 2026).
- Chunks engineered against the five-property checklist for verbatim citation lift the verbatim-citation rate 1.6–2.2× over rerank-survival-optimized baselines without adding new pages (citation-engineering cohort, 2026).