agentleFS
XPOZpublicXPOZpublic/geo-seo-agent/.claude/skills/visibility-measure/SKILL.md

visibility-measure

Skill0 starsChanged 3 months agoMITView on GitHub

Measure the company's visibility for every active target prompt in AI engine answers (GEO), plus the Google surface where Search Console and GA4 are connected, and record dated, comparable snapshots. Judgment-free sensor; runs weekly.

  • Reads credentials

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 0 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈0

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Every active prompt in strategy/target-prompts/target-prompts.md.

    line 14
  2. Very specific:

    Run .venv/bin/python scripts/target-prompts-visibility.py.

    line 18

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Visibility Measure

The sensor of the GEO loop: the only skill that produces measurement data rather than judgment. It answers one question on a fixed cadence: for each target prompt, how visible is the company right now, on each surface, and who is visible instead? Every downstream decision (gap diagnosis, brief prioritization, outcome verdicts) is an interpretation of what this skill records.

What a run does

1. Read the panel

Every active prompt in strategy/target-prompts/target-prompts.md. The panel is the only input deciding what gets measured: never ad-hoc queries, never "top queries we rank for."

2. Measure the AI surface (GEO)

Run .venv/bin/python scripts/target-prompts-visibility.py. It queries each engine configured in config.json (Claude, ChatGPT, Gemini via GEMINI_API_KEY or Vertex AI) with web search enabled, samples_per_engine times per prompt, in parallel, and writes a per-run folder recordings/target-prompt-visibility-runs/<timestamp>/ with two files:

  • visibility.json: the raw per-prompt, per-engine, per-sample record (full answer text, search queries, search results, cited URLs), each sample carrying a computed stats block: brand mentioned/cited/in-search-results, brand URLs, mention snippets, cited-domain counts.
  • visibility-summary.md: the agent-facing digest: per prompt×engine rate tables (OK / Mentioned / Cited / In search results), merged cited-domain counts, deduped search queries, brand URLs, and mention snippets.

The script does all deterministic extraction and aggregation; this skill reads the summary (the raw JSON only for deep dives; structure, per-engine differences, and slice recipes in references/visibility-json.md) and adds the judgment:

  • How prominently is the brand mentioned: the recommendation, or one option among many? Judge from the mention snippets.
  • Which cited domains are competitors. From the cited-domain counts, compute share of voice versus competitors, never a binary yes/no. Competitor data is what gap analysis later uses for root-causing (e.g. "the answer is assembled from listicles we're not in").
  • Where the answers come from. From the per-engine domain counts (.venv/bin/python scripts/cited-domains.py <run-folder>), classify the run's cited sources by surface type: vendor's own site, small-tool dedicated page, review site (G2/Capterra/Gartner), official platform docs, community thread, marketplace/directory, publisher platform (Medium/Substack/YouTube). Per engine, as shares of cited-source instances. This is the panel-level reading of which surfaces engines actually trust; recorded as observation only. Week-over-week shifts here are the evidence behind gap analysis's content-plan proposals.
  • What the search queries reveal: the fan-out queries expose the phrasings that decide which pages get pulled into answers; discovery and gap analysis both use them.

3. Measure the Google surface (SEO), when Search Console is connected

Skip this section when integrations.search_console is false in config.json, and note the gap in the snapshot's method section.

For each prompt, record its Google reading from Search Console: position, impressions, clicks, CTR for the matching queries. Same intent, second surface. The two readings together enable diagnoses neither channel could express alone: "ranks well but never cited" vs. "cited but invisible in Google." Alongside it, the GSC topline (clicks, impressions, average position, period-over-period) as site-wide context.

4. Measure AI-referral traffic, when GA4 is connected

Skip when integrations.analytics is false; note the gap.

The traffic that visibility converts into, and the only surface where AI-mediated humans are directly countable: GA4 sessions referred from AI engines. From the GA4 Data API, same 28-day window every run:

  • Per referrer: sessions and active users by sessionSource, broad-matched against AI engine names (chatgpt, openai, perplexity, claude, gemini, copilot, anthropic, deepseek, mistral, grok; new referrers appear regularly, so matching stays broad). Verify each unfamiliar source before counting it; broad matching produces false positives (e.g. ngrok-free.dev matches "grok").
  • Per referrer × landing page: top landing pages by sessions. Which pages AI engines actually route humans through is revealed demand: a discovery signal for /target-prompts-discovery and an attribution signal for retro, recorded here, interpreted there.

Sites that additionally tag AI-agent page fetches by user-agent (a custom GA4 event) can report engine fetch counts as an extra middle-of-funnel reading; that instrumentation is site-specific and outside this repo's scope.

5. Track leading pages, when Search Console is connected

The top ~20 pages by clicks: per-page clicks, impressions, and position, with week-over-week and 4-week trends. This watches the site's existing equity; a leading page losing traction is a rework candidate even if it targets no panel prompt. As everywhere in this skill: the snapshot records the trend and the deltas summary may flag it ("page X: -40% clicks over 3 weeks"), but deciding whether a decline is real and warrants rework belongs to /content-gap-analysis.

6. Write the snapshot

Two artifacts per run:

  • A dated snapshot in memory/visibility-snapshots/ (conventions and section order in that folder's index): the full per-prompt, per-engine record, linking to the raw run folder it interprets. Append-only: snapshots are never edited after the fact; they are the system's ground truth for trends and retro attribution. Mention and citation rates are both structured per-prompt columns, never collapsed into one: cited-but-unmentioned means your page sourced the answer while the brand went unnamed; mentioned-but-uncited means the model knows the brand without retrieval. On prompts where engines run no search, mention rate is the only reading that can move.
  • A short deltas summary: what moved since the previous snapshot: prompts that gained or lost visibility, published items whose target prompts moved, new AI referrers or notable referral shifts, leading pages losing traction, anomalies worth a human glance. Observations only, no conclusions.

What it never does

  • No judgment. It may note "prompt X went 0/4 → 3/4 and content targeting it published two weeks ago", but concluding causation, judging what worked, or proposing actions belongs to gap analysis and retro. A dumb sensor is a trustworthy sensor.
  • No panel edits. Dead prompts or emerging patterns are noted in the summary for /target-prompts-discovery to consider. The panel has one writer.
  • No content reading, no briefs, no recommendations.

Principles

  • Consistency over cleverness. Same prompts, same engines, same sample count, same recording format, same date-window math every run; trends only mean something if the instrument does not change between readings. When the method must change (new engine, different sampling), the snapshot notes the discontinuity explicitly.
  • Multi-sampling is non-negotiable. At least 2 samples per prompt per engine (samples_per_engine in config.json); single runs are noise, and volatility findings show roughly half of cited pages change monthly.
  • Missing credentials abort before anything runs. A configured engine with no credentials is a configuration error, not a partial failure: the script refuses to start and names the fix (add the key, or remove that engine from config.json engines). The human decides which.
  • Partial failure never aborts the run, except out-of-credits. An engine unreachable or a call failing transiently (timeouts, 503s) is recorded as a gap in the snapshot, and the run continues; every such run produces its snapshot, even a degraded one. Billing is the exception: the script preflights each configured engine with a tiny call before sampling, and a credits/billing error on any of them aborts the whole run before a single sample is taken (no run folder is written). When that happens, stop the measurement, tell the human which platform needs credits, and wait; rerun the sensor after the top-up. A snapshot missing an entire engine breaks week-over-week comparability, which is worse than a late snapshot.
  • Expect lag. New content typically takes days to weeks to appear in answers, and 2-4 weeks to be cited stably. Absence of movement in week one means nothing.

Cadence

Weekly. Judgment-free by design, so it is the one skill safe to schedule unattended later. Gap analysis invokes it automatically when no fresh snapshot exists for the current week.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.