Start with the pipeline, not the rank
Generative engines do not return a ranked list. They select candidate pages, retrieve passages, synthesise an answer, and cite evidence. Optimisation fails when you improve the wrong stage. Lumar GEO Studio is built to mirror that pipeline in software.
| Stage | What happens | In Lumar GEO Studio |
|---|---|---|
| 1. Availability | Page can be fetched and parsed | Cited-page crawls; content eval requires a reachable URL |
| 2. Candidate selection | Page is allowed to compete for this query | Prompt runs + cited-page lists — if absent, the failure may be upstream of content quality |
| 3. Chunk retrieval | Passages are embedded and matched | Content Precision, Semantic Recall, chunk-level eval |
| 4. Synthesis & citation | Content is quoted and linked in the answer | Answer logs, citation quality, brand mention quality, visibility score |
| 5. Brand positioning | Brand is named vs competitors | Brand mentions, mention quality (role, placement, prominence) |
Two measurement layers in the product
Prompt-level (AI Visibility) — tracks how AI models answer the queries you care about.
- Scheduled prompt runs per project and AI platform
- Visibility score — composite of how often and how well you appear (citations + brand mentions)
- Cited pages — URLs the model linked to when answering
- Brand mentions — named references to your brand and competitors in the answer text; competitor tracking is built in
Page-level (content evaluation) — assesses whether a specific URL is structurally fit to be retrieved and cited.
- Evergreen Health: headline composite (0–10) across Content Precision, Semantic Recall, Content Freshness, Content Uniqueness, and Questions Answered
- Content Precision and Semantic Recall: the two core content-quality signals
- Runs on cited pages (and pages you send to the copywriting workspace)
These layers connect: a prompt run tells you that you lost a query; page evaluation tells you why a URL may not be citable.
What visibility score actually reflects
Visibility score is not a rank position. It combines:
- Citation quality — when your owned URLs are linked, how strong those citations are
- Brand mention quality — when your brand is named, how well that mention serves you (role, placement, prominence, sentiment)
Both components are dampened by consistency — appearing in 1 of 10 runs with a perfect quality score ranks lower than appearing reliably with good quality. A one-off spike is weaker signal than sustained presence across runs. Mentions carry more weight than citations in the composite (roughly 3:1), so brand positioning in the answer text matters even when your URLs are not linked.
Use visibility at topic and prompt level for prioritisation. Use per-model breakdowns when platforms diverge — different AI providers often behave differently for the same prompt set, especially across informational vs comparison queries.
What the product does not claim to measure
Be explicit about limits:
- No guaranteed view into proprietary retrieval indexes — you observe outputs (answers, citations, mentions), not internal candidate pools
- Prompt sets are samples — conclusions strengthen with cadence and trend length, not single runs
- Content scores evaluate page fitness — they do not replace live prompt monitoring after you publish