keyword-vet-aio

Evaluate Google AI Overview presence and click-cannibalization risk for BID-passing Pleasur.ai Stage 01 keyword candidates using current Ahrefs evidence. Use after BID validation and before prioritiza

SKILL.md

GitHub

Keyword Vet — AI Overview

Evaluate only candidates with bid_verdict=PASS. Read references/aio-rubric.md before deep scoring.

Required inputs

Require:

  • BID evaluation, keyword, target country, observed intent, and tool-led routing status;
  • current Ahrefs Keywords Explorer evidence containing serp_features;
  • a human-approved run mode: presence_only or deep;
  • for deep, a fresh Google AI Overview response body from current Ahrefs Brand Radar coverage.

If run mode is absent, use the safer presence_only mode and record that decision. Apply the run packet's approved evidence-freshness and bounded-retry policies.

Evidence contract

  • Use the configured Ahrefs integration as the only external SEO data source.
  • Inspect live schemas before requesting fields. Do not assume legacy tool names.
  • Detect AIO presence only from Keywords Explorer serp_features; the literal feature is ai_overview.
  • Do not request serp_features from SERP Overview; it does not provide that field.
  • Use a Google AI Overview response captured by Ahrefs Brand Radar as the only acceptable deep-scoring body.
  • Never use general web fetch, another provider, a search snippet, or model memory as a body substitute.
  • Record exact query, country, report/tool, fields, filters, retrieval time, provider freshness/reference, status, and error.

Process

1. Detect presence

Batch compatible keyword requests when the live integration supports it, while preserving candidate-level provenance.

  • If ai_overview is absent, set has_aio=false and aio_verdict=PASS with reason aio_absent.
  • If present, set has_aio=true and continue.
  • If current presence evidence is unavailable after bounded retries, return needs_data with reason aio_presence_unavailable.

Do not infer absence from a null or failed response.

2. Evaluate routing exceptions

For an AIO-present candidate:

  • tool-led intent is outside the blog queue; retain PASS only as an AIO routing note with reason tool_led_immune.
  • comparison-led commercial investigation is not automatically exempt. Record the specific reason readers may still need options, criteria, evidence, or firsthand comparison, then apply presence-only mode or the deep verdict table normally.

No blog candidate is exempt merely because an intent label is commercial.

3. Apply presence-only mode

For each non-tool-led AIO-present candidate, set:

  • aio_completeness_score=null;
  • aio_click_intent=null;
  • aio_verdict=RISKY;
  • aio_reasoning=aio_present plus a candidate-specific information-gain requirement for any later article.

Presence-only mode must never emit FAIL_CANNIBALIZED; that verdict requires observed-body evidence.

4. Apply deep mode

Obtain a fresh Google AIO response body from Ahrefs Brand Radar. Record the body source, engine, capture/freshness timestamp, query and country match, and cited-page evidence when available.

If the body is missing, stale under the approved policy, for the wrong engine, query, or country, or cannot be retrieved after bounded retries, return needs_data for deep evaluation. Do not silently downgrade the same candidate to presence-only unless the run packet explicitly authorizes that fallback.

Score the observed body using the reference rubric. Judge how completely the AIO satisfies the search need, not how polished it sounds. Record:

  • aio_completeness_score from 0–10;
  • aio_click_intent as yes-deep, yes-shallow, or no;
  • one specific rationale naming what the AIO covers and what a click would still provide;
  • uncertainty and any missing evidence.

Use an independent review pass only when the active Buzz runtime exposes a general review capability. Do not select or name a provider or model. When no independent pass exists, perform a second documented self-review against the same rubric and set review_mode=self_review.

5. Set the deep verdict

Apply this table to every non-tool-led candidate:

Completeness score Click intent Verdict
0–4 yes-deep or yes-shallow PASS
5–7 yes-deep or yes-shallow RISKY
8–10 yes-shallow RISKY
8–10 no FAIL_CANNIBALIZED

Treat any unlisted combination as internally inconsistent. Perform one documented rescore against the observed body. If it remains inconsistent, return needs_data with reason aio_score_click_mismatch; do not choose a verdict.

Never force a distribution of PASS, RISKY, and FAIL results.

Output schema

Persist for every BID-PASS candidate:

  • keyword, country, aio_mode, has_aio, and the raw observed serp_features;
  • aio_completeness_score, aio_click_intent, aio_verdict, aio_reasoning, and aio_reason_code;
  • aio_body_source, engine, body freshness/capture reference, cited pages when used, and nulls when no body was requested;
  • review_mode, information-gain requirement, uncertainty, retry outcomes, evidence provenance, and rubric version.

Keep this evidence in the immutable Stage 01 validated-research packet. Do not write caches, CSVs, repositories, CMS records, or production systems. Do not trigger prioritization or Stage 02.

Summary

Report candidate counts for:

  • no AIO / PASS;
  • AIO present / exempt or routed;
  • AIO present / RISKY in presence-only mode;
  • deep scores 0–4, 5–7, and 8–10;
  • needs_data with reason;
  • deep-body source coverage.

List the highest-risk candidates with their exact evidence and reasoning.

Quality checks

  • Every BID-PASS candidate has a terminal AIO result or needs_data.
  • Presence came from the literal ai_overview feature in Keywords Explorer.
  • No failed/null response was interpreted as AIO absence.
  • Presence-only mode emitted no FAIL_CANNIBALIZED verdict.
  • Every deep score used an observed, fresh, query-matched Google AIO body from Ahrefs Brand Radar.
  • Every deep verdict follows the completeness-score/click-intent table; retained click value is never marked FAIL_CANNIBALIZED.
  • Every score and exemption has candidate-specific click reasoning.
  • No provider, model, sub-agent primitive, environment flag, cache path, cron behavior, or exit code is assumed.