Keyword score methodology: popularity, difficulty, relevance and opportunity

What these scores are, and are not

On the App Store, popularity is a modeled popularity index, not store-reported Apple search volume. On Google Play, it is a modeled popularity index, not store-reported Google Play search volume. Difficulty, relevance and opportunity are ASOClimb models too.

The inputs are evidence the stores publish for anyone to see: the autocomplete suggestions they show as you type, the apps they rank for a search, those apps’ rating counts and star ratings, and your own listing’s text. Nothing is bought from a third party and nothing is estimated from install counts.

Scores are for comparing keywords with each other in the same store and country. They are not a forecast of rank, downloads or revenue.

Current score version: asoclimb-wow-keywords-v1. It is returned with every score ASOClimb serves, so a saved result says which formulas produced it.

Popularity (P)

P = 0.6 × A + 0.4 × R

A

Autocomplete evidence. For each prefix a shopper types on the way to the term — the first 2 characters (weight 1), the first 4 characters (weight 0.85), the first word (weight 0.7), the full term (weight 0.5) — ASOClimb reads the store’s suggestions. Where the term appears at position p of n suggestions it scores max(10, round(100 − (p − 1) × 90 ÷ (n − 1))); where only a longer suggestion starting with the term appears, half of that; where it does not appear, 0. That score is multiplied by the prefix weight, and A is the best of them: being suggested after a few letters is stronger evidence than being suggested only for the whole term.

R

Ratings. The median rating count of the top 10 apps for the term, log-scaled against a cap: R = 100 × min(1, log10(1 + median) ÷ log10(1 + 50,000)).

A component that was never observed (no suggestion list answered, no rating count read) is left out and its weight moves to the other, rather than being scored as an observed zero.

Difficulty (D)

Computed over the top 10 apps the store returns for the term.

D = 0.5 × Rt + 0.2 × T + 0.15 × Q + 0.15 × B

Rt

Rating strength: the mean of 100 × min(1, log10(1 + rating count) ÷ log10(1 + 100,000)) across those apps.

T

Title match: the share of those apps whose title contains every word of the term (whole words, plurals folded), × 100.

Q

Quality: the mean star rating above 3, scaled so 3★ or lower is 0 and 5★ is 100.

B

Big publishers: the share of those apps from a developer with at least 2 apps in the top 10, or with at least 100,000 ratings, × 100.

As with popularity, a part that could not be observed hands its weight to the parts that were.

Relevance (Rel)

Relevance is computed only when there is a listing to compare with (in Wow Keywords runs, not in the free keyword checker).

Rel = 0.5 × L + 0.3 × C + 0.2 × E

L

Lexical: for each word of the term (connector words left out), the strongest place your listing uses it — title 1, category 0.8, description 0.6, absent 0 — averaged across the words, × 100.

C

Competitors: for each competitor of your app, 1 if it ranks in the top 10 for the term, 0.6 in the top 30, 0.3 deeper, 0 if not ranked; averaged, × 100.

E

Embedding: the cosine similarity between the term and your listing, computed on ASOClimb’s servers, clamped to 0–1, × 100. Optional.

When E is unavailable, or there are no competitors for C, that weight moves to L.

A candidate with Rel below 40 is listed as dropped and never written into a variant.

A phrase is only written into a title or short description when Rel is at least 50: a popular store suggestion can still be someone else’s product.

Opportunity (O)

O = P × (100 − D) ÷ 100 × Rel ÷ 100

Opportunity weighs modeled demand against how hard the current top results look, then by how well the term fits your app. Without a listing (the free keyword checker) the relevance factor is left out: O = P × (100 − D) ÷ 100.

All four scores are rounded to whole numbers from 0 to 100.

How much is read

A Wow Keywords run reads at most 60 store pages per country and scores at most 50 candidates per country. Store search results are cached for 24 hours and suggestions for 7 days, so a score can be up to that old; each result carries the time it was observed.

A term the budget never reached is not scored and not shown. No score is filled in or guessed.