How we measure what Amazon won't tell us.

The panel

18,042 active Amazon US listings, sampled across six subcategories: vlogging microphones, smart light bulbs, cooking thermometers, hair dryers, electric kettles, and yoga mats. Stratified sampling within each subcategory: 50% top-100 by rank, 30% rank 101–500, 20% rank 501–2000. Smaller panels run on Amazon UK (3,200), DE (2,800), JP (1,900), Shopify (4,100), and TikTok Shop (1,400).

Listings rotate quarterly. We add new listings as old ones de-list or fall outside the sampling window. The panel is always at least 14,000 active listings.

What we log

Daily, per listing: organic rank, organic impressions (inferred from rank distribution), session conversion rate, units sold, new verified reviews, returns rate (when visible), PPC spend share, title length, bullet length, A+ presence, image count, image quality score, and 12 keyword-density features.

What we infer

We run ridge regressions of rank-delta on the signal panel, with listing-level fixed effects and ridge penalty λ=0.1 chosen by 5-fold cross-validation. Standard errors clustered by subcategory.

Inferred weights are notAmazon's actual weights. They are a model of Amazon's model. We can detect that a signal's weight has changed; we cannot prove the exact magnitude. We report magnitudes with this caveat in mind, and we test robustness across at least 10 alternative specifications before publishing.

What we cannot see

Amazon's source code. Their internal CTR models. The personalization layer (we average across users). Buy Box logic. Sponsored ranking (different model, partial overlap).

How to read our dispatches

Trust direction over magnitude. If we report a +11pp shift, the direction is high confidence; the exact size is ±2pp. Trust the qualitative tactic over the quantitative target. Most importantly, run our recommendations on a small subset of your listings first, measure, then scale.

Replication and corrections

We will share the regression code and the de-identified daily aggregates on request from credentialed researchers ([email protected]). If you spot an error, email us. Corrections are dated, transparent, and linked from the affected dispatch.