ASO Workbench ranks App Store terms with a model trained on Jev's answers

Mustafa Kendigüzel posted ASO Workbench on September 22. The post says the keyword model was trained on about 55,800 terms from five countries, and it links mustafakendiguzel/aso-workbench. The app is a native Mac program, macOS 15 or later, Swift 6.2. The README calls the license MIT with the Commons Clause: free to use, including for your own commercial apps, and not licensed for sale as a product. GitHub’s license field is Other.

docs/POPULARITY.md is the measurement note. Apple publishes a 1 to 100 popularity for the top of each genre. Since 29 September 2025, everything below that list is reported as 5. The app shows Apple’s number when it has one. The estimate is a pair of CreateML models, about 2 MB, trained on terms labelled with Apple’s popularity. The clues are public autocomplete, public search results, and Jev through OpenRouter. Jev is asked narrow questions: whether the term is a brand, whether it sounds like real demand, which genre it belongs to, and how well it fits the app. The note says those language-fit, demand, and brand answers are the most useful features. If there is no OpenRouter key, the app shows ”< N” and does not fill in a weaker guess.

The held-out table is for a model dated 19 September 2026, trained on 44,658 rows. On 578 terms that have autocomplete clues, the kind the app sees in use, Spearman rank agreement with Apple is 0.74, the mean error is 10.4 points, and 81% of estimates fall within one band. A random 20% of all terms is Spearman 0.45, mean error 15.6, 75% within one band. Holding out Turkey, a country the model never saw, is 0.37 and 17.2 points. Holding out a whole topic (Games, Productivity, or Health) is Spearman 0.27 to 0.35 and a mean error of 20 to 30 points. A separate check, whether a term sits in Apple’s published range, is described at precision 0.78 and recall 0.88.

A golden keyword, in that note, has Apple popularity of at least 30, difficulty of at most 45, is not someone else’s brand according to both Jev and the search results, and has a Jev fit of at least 50%. Difficulty itself is computed from the top ten public results, not from Jev. The README’s shorter line, a correlation of 0.74 and about 10 points on a typical term, matches the autocomplete row and not the random split.

jevsearch asks Jev to re-rank documents and prints Hit@1. This app asks Jev for features, then a local model predicts Apple’s number. The 0.74 is agreement with Apple, on one held-out slice. We did not run the app.