Keyword Prioritization

Keyword Prioritization

Rank only fully evidenced candidates that passed BID. Read references/scoring-policy.md before calculating scores.

Required inputs

Require for every candidate:

• keyword, target country, discovery source, volume, KD, traffic potential, parent topic, and current content-gap status;
• BID verdict, brand fit, product fit, product applicability, business value, free-seeker flag, observed intent, tool-led flag, current brand DR, median top-ten DR, weak-link evidence, and BID provenance;
• AIO mode, presence, verdict, reasoning, and AIO evidence provenance;
• an approved, versioned brand/product snapshot, current Pleasur.ai content inventory, and current evidence-backed max_targetable_kd;
• the Stage 01 formula/policy version and evidence-freshness decisions.

Use needs_data when a required input is missing, stale, contradictory, or country-mismatched. Do not rank incomplete candidates.

Eligibility and routing

Evaluate every candidate, then route before ranking:

1. BID FAIL → rejected with the BID reason.
2. BID needs_data or AIO needs_data → needs_data.
3. aio_verdict=FAIL_CANNIBALIZED → rejected.
4. aio_verdict=RISKY → eligible only when the Stage 01 packet explicitly accepts the named information-gain/click requirement; otherwise needs_data for human decision.
5. serp_intent=tool-led → routed_tool_opportunity, never the blog candidate set.
6. Existing strong Pleasur.ai coverage → track_existing_strength, never selected for a new article.
7. Remaining BID-PASS and AIO-eligible candidates → scoring pool.

Do not describe routed or unselected candidates as rejected unless a gate actually failed.

Scoring process

1. Traffic score

Calculate a 0–10 opportunity score from current traffic potential and KD:

traffic_raw = min(10, 2 × log10(traffic_potential + 1) - kd / 20)

Clamp the result to [0, 10]. If traffic potential is unavailable, do not silently substitute volume; return needs_data unless the approved run policy defines and versions an alternative normalization.

Record raw inputs, calculation, clamping, and formula version.

2. Brand and product fit

Reuse the evidence-backed BID scores; do not rescore them during prioritization.

• brand_fit remains 0–10.
• product_fit remains 0–10 and is the dominant factor when products exist.
• Preserve the named product and fit rationale.

If the approved brand snapshot has no products, use the no-products formula below and record product_fit_applicable=false.

3. Winnability

Score 0–10 relative to current Pleasur.ai authority, never KD alone. Require current brand_dr, current evidence-backed max_targetable_kd, median top-ten DR, and named weak links. Require max_targetable_kd > brand_dr; otherwise return needs_data.

Calculate base_winnability exactly:

• kd <= brand_dr → 10;
• brand_dr < kd <= max_targetable_kd → 5 + 3 × (max_targetable_kd - kd) / (max_targetable_kd - brand_dr);
• max_targetable_kd < kd <= max_targetable_kd + 15 → 3;
• kd > max_targetable_kd + 15 → 1 and above_ceiling=true.

If weak_link_count >= 3, set winnability = min(10, base_winnability + 2.0) and name the three or more URLs supporting the override. Otherwise set winnability = base_winnability. Keep full precision through final ranking.

If current max_targetable_kd is unavailable, return needs_data. Do not derive an alternative from qualitative bands or revive a legacy cache value.

4. Free-seeker penalty

Set free_seeker_penalty=0.4 when BID flagged the query as free-seeking; otherwise use 1.0. Keep the candidate eligible unless another gate fails. Add top-of-funnel / low-business-value to its rationale.

5. Priority score

When products exist, calculate:

priority_score = (0.2 × traffic + 0.2 × brand_fit + 0.4 × product_fit + 0.2 × winnability) × free_seeker_penalty

When the approved snapshot has no products, calculate:

priority_score = (0.3 × traffic + 0.4 × brand_fit + 0.3 × winnability) × free_seeker_penalty

Keep full precision for sorting and display a consistent rounded value separately. Apply no boost for commercial wording, source, subjective enthusiasm, or preferred outcome.

Ranking and selection

Sort eligible candidates by:

1. full-precision priority_score, descending;
2. traffic potential, descending;
3. KD, ascending;
4. normalized keyword, lexical ascending.

Assign deterministic rank after routing. Approve at most the rank-1 eligible blog candidate. Mark other eligible candidates evaluated_not_selected; they remain available for a future run.

If the scoring pool is empty:

• return needs_data when any candidate could become eligible after required evidence or human acceptance;
• otherwise return terminal Stage 01 status rejected and preserve every gate reason.

Never start Stage 02, draft an article, publish, install a skill, trigger another workflow, or write external systems.

Output schema

Persist the following in the immutable Stage 01 packet:

• every candidate and its route/status;
• traffic inputs, calculation, score, brand fit, product fit, product applicability, winnability inputs, named weak links, winnability score, business value, free-seeker flag, penalty, and final priority score;
• the exact formula and policy versions;
• tie-break values, deterministic rank, selection reason, and non-selection reason;
• AIO acceptance requirement for every selected RISKY candidate;
• all source provenance, freshness, nulls, conflicts, uncertainty, and human decisions.

Summarize the ranked eligible set, tool opportunities, existing-strength routes, rejected candidates, and needs_data candidates. Clearly name the single selected candidate or state that none was selected.

Quality checks

• Only BID-PASS, AIO-eligible, fully evidenced blog candidates were ranked.
• Tool-led and existing-strength candidates were routed out before ranking.
• Product fit is the dominant weight when products exist.
• Winnability varies with current Pleasur.ai authority and observed SERP evidence.
• The exact interpolation and +2.0 weak-link adjustment reproduce the same winnability from identical evidence.
• No above-ceiling candidate received high winnability without a named, evidence-backed weak-link override.
• Free-seeker penalty was applied exactly once.
• Full-precision values, tie-breaks, and lexical normalization make the result reproducible.
• At most one candidate is selected; all other survivors are evaluated_not_selected.
• No local cache, CSV, repository, CMS, workflow, or production system was mutated.