What is Synthesis prompt?
The synthesis prompt is the engine-side prompt that takes the union of reranked candidate chunks across every sub-query in the fan-out tree and instructs the LLM to compose a single coherent answer. The prompt is not published by any engine in mid-2026, but its structure is inferrable from the rendered output: it always names a citation convention (numbered chips, bracketed footnotes), a paraphrase-vs-quote policy (some engines prefer quotation on factual claims, others prefer paraphrase), an answer-shape directive (bullet, paragraph, comparison table by intent), and a 'further sources' overflow rule for chunks the answer didn't render inline. The synthesis prompt rewards chunks that are easy to compose with — chunks shaped like citable claims with explicit numbers, named entities, and self-contained context survive composition at materially higher rates than chunks that need supporting context from the surrounding paragraph the prompt no longer has.
How it relates to AI UGC
The multimodal slot of the synthesis prompt rewards images shaped for composition — persona-locked, schema-described, caption-anchored to the text chunk the answer will cite. Single-persona discipline across the page set is the visual equivalent of a citable-claim shape: it makes the synthesis prompt's image-selection step cheap rather than ambiguous.
Key statistics
- Inferred synthesis prompts route 60–80% of the answer's rendered text from chunks ranked in the top 3 surviving slots per sub-query — chunks that survive rerank into slot 4–8 are rendered into the answer at materially lower rates (synthesis-prompt position audits, 2026).
- Self-contained chunks — those whose load-bearing claim does not require the surrounding paragraph for context — survive synthesis at 1.6–2.0× the rate of equivalent chunks whose claim is split across two paragraphs (self-containment audits, 2026).
- Programs that brief editorial against the inferred synthesis prompt shape (citation convention, paraphrase policy, answer shape) ship 20–30% higher verbatim citation rates than programs that brief only against rerank survival (synthesis-aware briefing cohort, 2026).