---

name: open-geo description: Use this skill when measuring or rechecking a brand's visibility in AI answers, citations, source lists, or competitor rankings; when resuming an interrupted GEO run; or when another workflow needs captured answers and a portable measurement artifact. Use it even if the user says only “am I recommended?”, “who gets cited instead?”, or “run these questions”. Capture rendered answers through an available engine playbook, persist validated data, and export JSON; add PDF/dashboard only when requested. Accept a query CSV or a semantic-core handoff. Question research requires the separate opt-in harvest flow; this skill does not audit or edit the target website.

Open GEO

One run = one question set, target, and engine. Return a portable open-geo.run-artifact.v1 JSON file, with optional PDF/dashboard. pipeline/INTERFACES.md defines the contracts; pipeline/schema.py validates captures.

Invocation and defaults

/open-geo <questions.csv> <engine> <domain-or-URL-prefix> --brand "<name>"
          [--n-worker N] [--output data|dashboard|pdf|both]
          [--artifact-out <path.json>] [--period today|all]
          [--lang en|ru|zh|ar] [--repeat R]

These are skill arguments, not a shell command. Defaults: one worker, data, reports/run-<id>.json, all, en, one capture. --lang controls the exported presentation and user summary, not query language or captured sentiment. --repeat R costs R independent captures; never infer it from “check again”.

Read only what this run needs

| When | Read | |---|---| | Runtime is not resolved, dependencies are missing, or caller paths differ | runtime.md | | Before creating/resuming a run or processing captures | run.md — exact CLI sequence, counts, failure policy | | Before capture/finalization, or when a result looks suspicious | gotchas.md — observed failures and corrections | | User requests question research instead of supplying a CSV/core | harvest.md — grounded research and review gate | | PDF, dashboard, combined-engine report, or repeats requested | deliverables.md | | Explaining a metric or its denominator | metrics.md | | Editing/testing this skill | validation.md — replay evidence and acceptance checks |

STEP R — resolve the runtime and paths

Prefer the current repo containing pipeline/INTERFACES.md, then caller-supplied OPEN_GEO_ROOT, then a writable installed package, then the managed runtime. Bootstrap instructions are in runtime.md; perform setup rather than asking the user to launch services. Use the runtime's .venv/bin/python from its root.

Before changing directories, resolve the input CSV/core and explicit artifact path against the caller's original directory. Resolve one absolute database path (normally data/aeo.db under that root, or the caller's explicit isolated DB) and pass it to every CLI.

STEP A — resolve the measurement

Reuse arguments and free-text context. Ask only for missing brand, target, engine, and question input; concurrency defaults to 1. List engines from actual playbooks under engines/, excluding README.md and FAST_PATH.md. A missing playbook is a blocker, not permission to invent capture steps.

Accept a supplied query,lens CSV directly. For a core.json handoff, read questions_csv, brand, domain, and totals.coverage (INTERFACES §8); keep the coverage for the final summary. Missing or empty input blocks run creation.

Host primitives

Ask with the host's question tool (AskUserQuestion on Claude), or in conversation. Spawn with its native agent tool (Claude Agent, Grok spawn_subagent, equivalent on other hosts). Read the host-native worker contract at the pointer below. Without delegation, perform the same capture contract sequentially; do not substitute an API or headless scrape for the rendered browser surface.

STEP A.5 — questions, only if requested

Use harvest.md only when the user opts into generating/researching questions. It produces a reviewed CSV and provenance; a supplied CSV/core bypasses this step. Do not replace the user's questions with a newly invented set.

STEP 1 — validate, then create or resume

Follow run.md: validate CSV before writing a run, retain file order, count unique (query,lens) pairs. Reject malformed rows rather than silently dropping them. Check for a compatible unfinished run; an explicit resume request or fully specified invocation resumes it automatically. If the set differs, create a separate run and identify the unfinished one. Capture only the --pending rows.

For --repeat R > 1, read deliverables.md before creating runs; repeats share a group, but never share captured rows.

STEPS 2–3 — capture pending rows

Read the whole engines/<engine>.md and the host-native ../../agents/capture-worker.md. Probe the host's visible-browser connection and playbook-required clean context before creating a new run. A missing connection, login wall, or unsupported clean context blocks live capture.

Default to sequential capture. If concurrency is requested and the host/browser supports isolation, split pending rows into min(N, pending count) ordered chunks. One worker owns each chunk and its tabs; account limits still apply across workers. Give each worker:

• Full engine playbook, exact (query,lens) rows, target including its path, brand, engine, and pointers to INTERFACES §1 and pipeline/schema.py.
• A unique absolute output path inside this invocation's temporary directory (mktemp -d), plus chunk index. Never reuse a global /tmp/open_geo_cap_1.json.

Workers persist a validated array after each query, then return its path, JSON, and captured/missing keys. They never receive the DB/run ID or ingest commands. The orchestrator owns writes and exports. On hosts without workers, keep that same separation of capture and ingest in the sequential execution.

Hard capture constraints:

• Read source/citation URLs in place; never visit the cited sites. Close only tabs opened for this capture, including on failure.
• Preserve overview_present, source order, and citation/source inclusion. Compute ranks with target_ranks; a shared domain alone does not match a URL-prefix target.
• Preserve prior_context and its evidence; a contaminated answer is flagged data, not a clean measurement. Screenshots are transient (screenshot_path=null).
• On CAPTCHA or rate limit, stop the affected capture and preserve completed work; do not solve challenges or keep increasing concurrency.

STEP 4 — persist and reconcile

Follow the commands in run.md. Validate each returned file's engine and keys against its assignment, then ingest incrementally. Read errors even on exit 0. skipped is a duplicate, not a corrected overwrite. Verify stored keys/counts after ingest; do not equate worker success with durable storage.

Finalize done only when every unique CSV pair is stored with no extra keys. Missing captures are not overview_present=false answers. Preserve a paused run as running with its ID and pending count for resume; finalize an abandoned or unrecoverable run failed. Never silently label a partial run as complete.

STEPS 5–5b — metrics and sentiment

Aggregate with pipeline.aggregate and save its JSON stdout. It computes metrics and domain statistics, but does not synthesize sentiment. Read per-query sentiments via pipeline.run --sentiments; write one neutral sentence for each represented lens plus all, using only those notes and their language. All-null notes mean a null summary. Persist the {lens: summary} object with pipeline.lens_sentiment; verify the returned written lenses.

STEP 6 — export

Export JSON with pipeline.artifact using the same DB; open the file and check schema version, run ID/status, target, result counts, metrics, and lens summaries. The default data output ends here. Start presentation services or generate a PDF only for the requested output, following deliverables.md.

STEP 7 — return the result

Use the all row from aggregate stdout: answer coverage, target visibility in sources/citations, average best source/citation ranks, source→citation conversion, and brand mention rate. null means no data, never zero. Explain denominators when needed via metrics.md; brand mention is independent of linked presence.

Report captured/expected counts, status, unresolved blockers and prior-context flags alongside metrics. For partial exports, metrics describe captured rows only. Include core demand coverage if supplied, then the absolute artifact path and any verified PDF path/dashboard URL. Do not claim a fixture replay is a live GEO result.