How our keyword estimates work

The observation matrix

One expansion run asks Amazon's search-suggestion service a planned set of questions: your seed, the seed with a trailing space, curated modifier probes, every a–z continuation, and — on deeper runs — a–z prefixes, word-prefix probes, and depth-two expansions of the strongest terms. The answers form a matrix: queries down one side, ranked suggestion slots across the other.

Most tools collapse that matrix to a single count and throw the rest away. We keep it. Where a suggestion appears, how many of its compatible queries it appears in, and how high it ranks each time is the evidence everything below is computed from.

One caveat we state up front rather than in a footnote: estimates for different suggestions in one run come from shared queries, so their errors are correlated. This design is about extracting more from each fan-out — a claim of data efficiency, not one of statistically independent samples per keyword.

From observations to a demand estimate

Per suggestion, in order:

mass     = Σ over appearances of  rank-weight × specificity-weight
censored = mass adjusted upward when its slates came back full
M̄        = censored mass per opportunity, normalized to a reference
            opportunity count so runs of different depth stay comparable
volume   = saturating curve of M̄, adjusted for phrase length
            (and, when a results page has been read, for result count)

The rank weight declines smoothly from the top slot to the last — no lookup tables, no cliffs where one position jump halves a score. The specificity weight discounts appearances on much narrower queries than your seed and caps the reward for appearing on broader ones. The censoring adjustment exists because a full suggestion slate is a truncated observation: an eleventh-place term on a full slate may be far stronger than its rank shows, and we widen its uncertainty rather than pretending the cutoff is data.

Every estimate ships with a band

A keyword estimate is an estimate. Ours always renders as a range with a confidence badge — High, Med, or Low — and the band is computed, not decorative. It widens with thin evidence (few compatible queries), censored evidence (full slates), stale data (served from cache), missing signals (no results-page count yet), and disagreement between the independent views we do have. A number you see with a tight band earned it.

When the evidence is too thin to support a metric at all — a difficulty score without readable review counts, a Book Intent Ratio where either department's evidence is sparse — the suite renders insufficient data. We think that beats a confident wrong answer.

What we calibrate against

Here is the honest part most tools skip: absolute Amazon search volume is unverifiable without Amazon's own data. Nobody outside Amazon has it — including us, and including every tool that shows you a bare integer.

Our calibration source is our own Amazon Ads search-term impression reports — real impression data from real campaigns. Against it we calibrate ordering, not magnitude: if the estimator says keyword A outranks keyword B, the impression data should agree. The curve's absolute scale is anchored to public reference points and a deliberate conservatism haircut, and we treat it as an index, not a measurement.

Calibration is ongoing: today's constants are structured priors — chosen under monotonicity constraints and checked against worked examples — and they will be refitted as our own impression data accumulates. Every estimate is stamped with the algorithm version that produced it, and a version change re-scores stored history rather than silently mixing two methods in one view.

Freshness and history

Suggestion data carries a multi-day cache horizon; results-page signals a shorter one. Anything served past its horizon is labeled as history with its observation date — never passed off as current. Trend deltas are computed only between runs that executed the identical query set, because a delta across two different expansion plans measures the plan, not the market.