An Answer-Engine Citation Evidence Log Template

Sooner or later someone important asks 'can you show me?' — a customer, a boss, your own future self wondering if March was really better. Screenshots in a Slack thread don't survive that question. A log does, if it's built to be audited rather than admired.

The template fields

• event_id — unique, boring, sequential
• captured_at — timestamp with timezone
• engine + mode — which engine, browsing on or off, any relevant settings
• prompt_id + set_version — ties the event to the exact measuring instrument
• cited_url — verbatim, including parameters; plus whose it is (yours / competitor / third party)
• answer_excerpt — the verbatim sentence(s) around the citation, copied not paraphrased
• full_answer_ref — where the complete answer text or screenshot lives
• logged_by — a person, because provenance has a name

Storage rules

Append-only: corrections are new rows referencing the old event_id, never edits. Full answers live in dated files the log points into. And the log outlives dashboards — it's the raw material any future tool can recompute from, which is precisely what makes it worth the discipline.

A filled row — fictional, for shape only

LD-0042 · 2026-09-10T14:22+05:30 · engine A, browsing on · prompt LD-P2, set v1 · cited https://luminadesk.example/pricing (ours) · excerpt: 'Pricing for round-robin tools varies; Lumina Desk lists per-seat rates on its pricing page.' · full answer at runs/2026-09-10/LD-P2-r2.txt · logged by MK. Fictional brand, illustrative row — but every field is the kind an audit accepts.

A failure worth checking

The paraphrase failure: a log whose excerpt column says 'engine recommended us for scheduling'. Paraphrase records the logger's optimism, not the answer. When the claim matters — a board slide, a case study, a dispute — only verbatim text with a stored source stands up. If the excerpt column contains no quotation marks, treat the whole row as unverified.