Prompt file contract

The output is one file the user uploads into their AI-visibility platform's prompt importer. Exact column names, engine codes, and region codes are defined by your platform — treat the shape below as a representative contract and adapt the enums to whatever your importer accepts.

Columns

A row is one monitoring prompt. A workable column set:

Column	Purpose
prompt	The question, written the way a real buyer asks an assistant, in the row's language. No fabricated brand names in non-branded rows.
description	Short human summary, e.g. <persona> · <source>.
topic	One of the few category topics, or the single branded topic.
is_branded	true only for brand-pass rows.
awareness_stage	Awareness / consideration / decision. Skew to consideration + decision.
platforms	Which assistants to monitor this prompt on — a subset of the engines your platform supports.
regions	Region codes supported by your platform, and only codes whose primary language matches this row's language.
persona, icp, key_source, reason, language	Documentation columns: who asks, the segment, honest provenance, why it matters, and the prompt's language.

Keep the operational columns (prompt, topic, branded flag, stage, engines, regions) aligned to what the importer actually reads; the rest document the research for human review.

Provenance labels (be honest)

social:<community> · keyword:<term> · sales_transcript · competitor:<name> · brand_page (brand pass only). Every row carries one.

Language and region rules

• Generate the prompt text once per language, in that language — translate, don't machine-localize spelling.
• A row's regions may only contain codes whose primary language equals the row's language. English text can group all English-primary regions; German text is German-market only; and so on.
• Bilingual B2B markets (e.g. tech audiences that search in English even in a non-English country) are a judgment call — confirm with the user, and if they want native coverage, add native-language rows in addition to, not instead of, the English ones.
• If a requested region isn't supported by the platform, say so and drop or remap it.

Engine rules

• Assign 2–4 engines per prompt, varied across the set so monitoring isn't skewed to one.
• Lean technical personas toward research-style engines; business/exec personas toward mainstream assistants and search-embedded answers; decision-stage "best / alternatives" queries toward the search-embedded engines that surface them heavily.
• Only include engines you actually monitor — don't add one you can't track.

Validation before handoff

Parse the file and assert: header matches the contract; valid UTF-8 (non-ASCII languages are expected); no empty prompt; branded flag and awareness stage in their allowed sets; every engine and region in the platform's allowed enums; each row's regions all share its language; no duplicate prompt text; a small number of category topics plus at most one branded topic. Then print the distributions (per topic, persona, awareness stage, language/region, branded vs non-branded) and show the user before finalizing. Keep the total reasonable (roughly 60–200 prompts) unless the user asks for more — authenticity beats volume.