Segments and Topics: How Results Are Grouped

The five axes of a segment

Provider, model, tool version, geography and language, all five stored on every sample. Two samples are comparable only when all five match. A share of voice for a whole run that mixed two engines would be an average across two different measuring instruments, so it is never emitted; the metrics tool returns one set of numbers per segment and says so in its own description.

Geography and language stay empty when a sample carried no locale. Defaulting them to something plausible would assert a locale the measurement never had, and locale is one of the facts the disclosure exists to state honestly.

What you see: engine and instrument token

Customers never see a model id. Each segment is presented as the engine name, ChatGPT, Claude, Gemini or Perplexity, plus an opaque instrument token, a short digest over provider, model and tool version. The token is stable for one instrument and different for another, so when a provider changes its model the trend shows a new token, and a brief can say two weeks are not comparable rather than silently merging two lines.

The exact snapshot is stored, because it is what lets re-verification pair a baseline with a follow-up correctly. It is redacted on the way out, at the presentation boundary, never in the identity the pairing uses.

Topics: your grouping

A topic is a label on a prompt. On Questions, or with answer_prompts_update through MCP, select prompts and add or remove topics; every topic you add goes onto every prompt you selected. A topic exists only because prompts carry it, so there is nothing to create or delete. To rename one, list its prompts with the old topic, and send them back with the new name added and the old one removed.

Metrics and trends accept a topic filter. The numbers are then computed over the admissible samples of the prompts carrying that topic, still per segment. Assigning a topic does not approve a prompt.

Using both together

Segments answer 'which engine' and topics answer 'which part of my business'. A common layout is one topic per product line and one per funnel stage, read per engine: your pricing questions on Perplexity, your comparison questions on ChatGPT. Trends group the same six metrics per day, week or month by segment, so each engine is its own line, and a bucket with nothing to measure is reported as unusable with a reason rather than as zero.

• Filter metrics by topic to isolate a product line.
• Read each engine's line separately; never average them by hand.
• When a token changes mid-series, treat the two halves as two series.

FAQ