Competition Analysis

Competition Analysis

This skill computes nothing itself. Every number and label below is already computed by deterministic, tested Python in the wrapped engine — hunt deepen-start runs the competitor autopsy and lineage checks and prints the result. Your job is only to read that result correctly and interpret it using the exact thresholds in references/competition.md — never recompute a concentration figure, never round differently, never relabel a level the engine already assigned.

When to use this

After a candidate has reached ECONOMICS_CHECK (via amazon-validation and economics being set). This is the entry point to the "deepen" pipeline: COMPETITOR_ANALYSIS → VOC_ANALYSIS → DIFFERENTIATION_ANALYSIS → RED_TEAM → HUMAN_REVIEW.

Steps

1. Run the autopsy.

   hunt deepen-start <run_id>
   

   This is deterministic, no LLM involved. It advances every eligible ECONOMICS_CHECK opportunity in the run to COMPETITOR_ANALYSIS and prints a JSON competitor_result per opportunity. Note every printed opportunity_id — each needs its own pass through the rest of this pipeline.

2. Read the concentration figures, not just the labels. For each of review_concentration, brand_concentration, oos_frequency, check:

   • The reported value (or None/UNKNOWN).
   • competitor_set_size (post parent-ASIN-dedup — this is the number that matters) vs competitor_set_size_before_parent_dedup (raw rows). If these differ a lot, sibling variants were correctly collapsed — do not describe them as separate competitors in any summary you write.
   • The assigned level: HIGH / MODERATE / LOW / UNKNOWN / INSUFFICIENT_SAMPLE. See references/competition.md §3 for the exact thresholds (0.6 review/brand concentration, 0.15 OOS frequency, min_sample_size=3).

3. Interpret levels correctly — do not upgrade or downgrade them yourself:

   • HIGH review or brand concentration → a small number of listings/brands dominate the visible competitor set. Worth flagging as a headwind, but this is advisory, not itself a hard reject.
   • MODERATE → between half and full threshold; note it, don't treat it as equivalent to HIGH.
   • LOW → no obvious dominance signal in what was observed.
   • UNKNOWN → the underlying value was None (empty/degenerate input). Never describe this as "low concentration" — it means the metric could not be computed, not that it was computed and came out low.
   • INSUFFICIENT_SAMPLE → fewer than 3 competitor observations. Never report the raw ratio behind this (e.g. "100% concentrated" from one competitor) as if it were a reliable figure — the engine deliberately withholds the precise decimal below the sample gate (competition.md §4); do not reconstruct it from other output.

4. Check the lineage result for each opportunity. The lineage_status will be one of CLEAN, SUSPECTED_CONTAMINATION, or UNKNOWN — it can never be CONFIRMED_CONTAMINATION at this stage (competition.md §5.1).

   • SUSPECTED_CONTAMINATION → flag for human attention explicitly in your summary. Say what triggered it (parent ASIN + high review count + short history, per competition.md §5.2; or an abrupt ≥25% review-count jump, per §5.3). Do not treat this as disqualifying and do not skip the candidate — only a recorded human decision can promote this to CONFIRMED_CONTAMINATION, and only that literal value fires the hard gate (competition.md §5.5). Continue the candidate through the rest of the pipeline unless a human tells you otherwise.
   • UNKNOWN → note that lineage could not be assessed (no or thin price history) — this is not evidence of either a clean or contaminated listing.
   • CLEAN → at least 3 stable observed snapshots with no red flags.

5. Move on. Once you've recorded the interpretation (concentration levels + lineage flag) in whatever running summary you're keeping for this candidate, proceed to reddit-voc / the automated VOC path for each opportunity_id still in COMPETITOR_ANALYSIS.

What NOT to do

• Do not compute max(counts)/sum(counts) or any other formula yourself — read the engine's output.
• Do not report a precise decimal for a metric the engine marked INSUFFICIENT_SAMPLE or UNKNOWN.
• Do not reject or silently deprioritize a candidate because of SUSPECTED_CONTAMINATION — surface it, don't decide on it.
• Do not call hunt deepen-start again on the same run expecting new competitor data beyond the budget cap — it is scoped and budget-limited by design (competition.md §1.2); repeated calls will not fetch more than the configured window allows.

See references/competition.md for the full rule set with exact source line references and tests.