ASO Patents

Document type	Academic study, not a patent
Title	App search ranking prediction towards enhanced app store optimization using ML and NLP
Authors	M. Abrar, Nitsa J. Herzog
Journal	Franklin Open, Volume 12, September 2025
Platform	Google Play Store, United Kingdom
Data	Top 60 apps across 32 categories, 1,964 apps, 50 features reduced to 13
Source	sciencedirect.com

What This Study Covers and Why It Matters for ASO

Scope note first. This is Google Play, not the App Store, and the two stores index metadata differently. On Play the long description is part of the searchable text. On iOS it is not. Do not carry the conclusions across without that in mind.

With that said, this is the better designed of the two ASO prediction studies on this site. Where the earlier study predicted app rating while describing it as ranking, this one states its target precisely: it classifies an app’s title keyword ranking as High, meaning positions 1 to 3, Medium, meaning 7 to 10, or Low, meaning beyond 10, in Play Store search results.

That is an actual search position, measured for the app’s own title keyword, which makes the feature importances meaningful in a way that rating prediction never was.

What predicted rank

The two strongest features, at equal coefficient:

• Sentiment of the description, 0.48. More positively worded descriptions accompanied higher ranks.
• Density of title keywords in the long description, 0.48. Repeating the title’s key term through the description, without stuffing, accompanied higher ranks.

Then app rank in category, followed by last updated date, content rating and genre at moderate importance.

The exclusions are as informative as the inclusions. Feature selection dropped short description sentiment and title keyword presence in the short description, meaning those carried less signal than their long description equivalents.

The description sentiment finding is the interesting one

Keyword density in the description is unsurprising on a store that indexes descriptions. Sentiment tying with it is not.

Two readings are available and the study cannot separate them. Either positive language genuinely helps, or well resourced teams write more confident marketing copy and also do everything else better. The second is at least as likely, and neither the sample nor the method distinguishes correlation from effect.

Still, it puts a number on something the ASO industry asserts without evidence, and it is worth testing directly rather than assuming.

Model performance

Model	Accuracy	Precision	Recall, High class	F1, High class
SVM	75%	77%	96%	85%
Random Forest	71%	77%	88%	82%
Gradient Boosting	68%	77%	82%	80%

Seven algorithms were compared with grid search hyperparameter tuning and SMOTE for class balancing. Ten fold cross validation gave SVM a variance of 0.000383 across folds, which indicates a stable rather than a lucky result.

Three quarters accuracy on a three class problem is real but modest. A quarter of apps are placed in the wrong band by features that include category rank, which is itself close to the outcome.

How to use it

• On Play, treat the long description as a ranking surface and carry your title keyword through it naturally. The study puts this at the top of the list.
• Do not port that to iOS. The equivalent iOS question is whether description vocabulary feeds relevance modeling rather than keyword ranking, which is what US9280595B2 and US20120191694A1 address.
• Rerun the method inside your own category. A cross category model averages away exactly the differences that matter.

Study Summary

The authors build a supervised classification model to predict an app’s title keyword ranking on the UK Google Play Store, using features grouped as user related, developer controlled and platform controlled. The dataset covers the top 60 apps across 32 categories, 1,964 apps in total, with 50 initial features reduced to 13 through feature elimination.

NLP processing includes YAKE keyword extraction from titles using three word n-grams, sentiment analysis of descriptions and reviews on a five point scale, and keyword density calculation within long descriptions. Seven machine learning models are compared, with grid search tuning and SMOTE class balancing.

The support vector machine performed best at 75% accuracy classifying apps into High, Medium and Low ranking bands. The authors report that developer controlled ASO features were the most influential group for model performance, with description sentiment and the density of title keywords in the long description as the two leading predictors.

A detailed breakdown of the feature set, the NLP pipeline and the model comparison will follow in a later update to this article.