LSHP Visibility Dashboard

Two independent signals, gathered on the same scheduled job:

• AI visibility — how often Layman Smart Home People surfaces in answers from ChatGPT, Gemini and Claude when a Singapore user asks genuine smart-home buying questions (no brand named in the question).
• Google search demand (GSC) — real impressions/clicks for laymansmarthome.com from Google Search Console, used to size which questions the site should prioritise winning.

Non-brand SOV = the share of answers to questions that do not name the owner brand, in which the owner brand is mentioned anyway.

Example: "who can install Home Assistant in Singapore" does not name Layman. If 20 of 48 such answers mention Layman, SOV = 20/48 ≈ 41.7%. This measures discovery by people who don't already know the brand — the number that matters for content strategy.

• Excluded from SOV, trends, and gap detection: all brand-intent prompts (e.g. "is Layman Smart Home People any good"). Those are leading questions — they almost always surface the brand and would inflate the metric. They're stored for sanity checks only.
• Metric = 100 × owner_runs / runs, per engine, per cycle, over non-brand prompts only.
• Single weeks are noisy (~48 samples per engine). Read the trend across weeks, and use the monthly-average toggle to smooth weekly churn and the 20→24 prompt-set change.

• Fixed prompt set (config/prompts.yaml, 24 questions) of real Singapore buying-intent questions: HDB/BTO/condo/landed installs, Home Assistant, device troubleshooting, and comparisons (DIY vs installer, Aqara dealer vs HA specialist, Matter vs Zigbee).
• Each prompt runs twice per engine per cycle → 48 raw answers per engine, ~144 per week, to damp single-answer churn.
• Every answer uses live web search (current web, not training data) with Singapore-facing instructions:
  • ChatGPT: gpt-4o-mini + web_search_preview, user location SG
  • Gemini: gemini-2.5-flash + Google Search grounding
  • Claude: claude-sonnet-5 + web_search tool

Each raw answer goes through a second, cheap LLM call (gpt-4o-mini) that extracts structured facts — deliberately not regex, because models write brand names loosely.

• Per mention: the brand name, matched canonical identity, its position in the answer, sentiment (positive/neutral/negative), and whether the answer recommended it for the user's situation.
• Canonicalization via config/brands.yaml (owner + known competitors + auto-discovered): variants like "Roen" / "Roen Pte Ltd" consolidate to one identity. Matching strips company suffixes, punctuation and spacing.
• Device/protocol brands (Aqara, Control4, KNX, Matter, Zigbee, Tuya…) are explicitly excluded — they are products, not service-provider competitors.
• Cited URLs (urls_cited) are extracted per answer.
• A parser failure on a substantial answer re-parses once; if still empty, the run is marked errored (visible in Run Status) rather than silently counted as "not mentioned".

• Where Layman shows up (per question): ● marks which engine(s) surfaced the owner for each question. Brand prompts included (full diagnostic view) — but this is a view, not the SOV number.
• Competitor mentions: how many times each competing provider was mentioned, and how many of those were recommendations, per engine. Device brands excluded.
• Who owns the content (most-cited sources): source domains AIs actually cite. ChatGPT and Claude only — Gemini returns Google redirect URLs, not real domains, so its citations can't be attributed. These are the pages to out-rank with content.
• Google search demand (GSC): monthly impressions/clicks from Search Console for the site, from the "final" (settled) data state. This is the real demand behind the questions tested.
• Action Engine briefs: prioritised gaps — questions where Layman did not surface on at least one content engine, demand is real (≥10 impressions), and a competitor is being cited instead. Ranked by a gap score of (missed engines × 10) + (100 − SOV) × 0.1 + demand. Each brief ships a ready-to-write title, out-ranking angle, H2 sections, cite-proof points and an index path.

• AI collection is weekly (Monday 03:00 SGT), fully automated: an external cron fires a workflow_dispatch, GitHub Actions collects, builds and deploys to Cloudflare Pages.
• GSC refreshes on the same job (last 3 settled days). Its "through" date therefore lags the AI cycle date — the dates are intentionally different and cadences differ.
• Run Status (top) is the health check: green = all 3 engines completed cleanly for the latest cycle; red = any errored run, missing engine, or incomplete cycle. Trust the numbers only when green.

• 48 samples per engine per week gives directional trends, not statistical precision — treat point changes between single weeks as noise.
• AI answers vary run to run with the same question (model churn, live search ranking shifts). Two runs per prompt + monthly averaging exist to damp this.
• "Recommended" reflects the AI's phrasing, not any real endorsement.
• Brand matching is alias + parser based; parser drift can surface odd canonicalisations — tracked under auto_discovered in brands.yaml for human review.