Blog index

App Keyword Install Estimates, Explained by Our Data Scientists

by ,  Data Scientist AppTweak Releases

App Keyword Install Estimates, Explained by Our Data Scientists

AppTweak estimates the number of organic installs that keywords bring to your app; this is a fundamental KPI that can help you measure a keyword’s potential or the impact of your latest metadata changes. Our Data Scientists have recently been working on an improved version of these estimates to further increase the accuracy of the information we provide. In this blog, our Data Scientists detail their methodology and share specific examples that illustrate the changes.

Improved app keyword install estimates using the new Google Play Console

When Google released its new Developer Console, it made the installs per keyword available per country and on a daily basis. AppTweak started fetching this new data as soon as possible: Having a day-to-day breakdown of an app’s downloads for its ranked keywords would allow the Data Science team to improve the precision of our organic download estimates per keyword.

The new Google Console now shows the store acquisitions from search terms on a daily basis

The new Google Console now shows the store acquisitions from search terms on a daily basis.

Now is a good time to recall Google’s changes made to the values provided on its Developer Console. Whereas before, organic installs were only given through worldwide monthly reports, when the new version came out, Google had merged the organic and paid installs that an app receives from a keyword and split the data per day and per country.

Discover all the changes made to the Google Console at the end of 2020.

As AppTweak wants to provide separate estimates for organic and paid traffic, we had to find a way to split the total keyword downloads into two sources. Shortly after the Developer Console update, Nadir Garouche published an article explaining how to isolate the paid installs for a keyword. The data mentioned in the article was unavailable to us; as a result, we had to find a proxy.

We noticed that for many keywords, the number of downloads generated by one specific app may vary by multiple orders of magnitude. We believe this is due to apps varying their advertisement pressure on some keywords from day to day. Using statistical tools, we were able to automatically estimate on which date an app did not perform any advertisements on a given keyword, therefore extracting the organic part from the total number of downloads. However, industrial secrets won’t allow us to disclose exactly how we did this!

From here, we became happy Data Scientists thanks to some neat data:

  • For the thousands of apps that are connected to AppTweak, we estimate how many organic downloads they make on any keyword, on any day, and in any country.
  • We can therefore extrapolate this information to build an estimator for any app-keyword combination, on any day, and in any country.

We built a model to estimate the keyword installs for any app in the stores

We fed our deep learning model with the newly collected data from the thousands of apps connected to AppTweak. For each app-keyword combination, we considered features related to:

  • The app’s metadata
  • The app’s rank on that keyword
  • The keyword’s search popularity
  • Whether the keyword was branded and (if so) the brand owner
  • The country

Using our expertise in deep learning (check out this article on our download estimates), we knew we could train a model to learn the relationship between these attributes and the number of daily organic downloads.

Once trained, our model could generalize the learned relationship to any new app-keyword pair to estimate its number of organic downloads. One of the key takeaways of using deep learning is that we can produce a model that works across countries. The model can learn the relationships between the different countries and thus benefit from the data of one app in multiple countries around the world. This is especially beneficial to smaller countries for which less data is available.

Once the model was deployed, we were able to identify some interesting behaviors. For example:

Branded keywords usually convert better than generic keywords.

When comparing keywords with the same popularity scores in the US, we found that branded keywords generated more installs than generic keywords. It is particularly interesting to note that the model learned this behavior solely from the data itself and not from a signal provided by AppTweak’s experts. Below, we see three examples of such behaviors. We also notice that as the popularity increases, the gap between the total installs for a branded keyword and a generic keyword grows exponentially.

Comparing the number of installs for brand vs. generic keywords with similar volumes.

For branded keywords, the brand (usually in first position) receives over 90% of downloads.

Installs from branded keywords mainly go to the corresponding app.

AppTweak provides more accurate estimates thanks to extra data

One of the major improvements is that our new model uses a day-to-day breakdown from worldwide apps. This means that this new model has learned from approximately 3000 times more data than before, allowing it to make even better predictions. Thanks to all the metrics we tracked during the learning process, we are confident that this new model will be able to estimate an app’s installs per keyword with even higher accuracy.

Examples of AppTweak’s improved app installs per keyword estimates

The similarity between the app and the keyword is better depicted

Let’s analyze the estimated keyword installs for the apps “PUBG MOBILE 1.5: IGNITION” and “Call of Duty®: Mobile”. We notice that “PUBG MOBILE” is the brand for the keyword “pubg” and generates 240,941 installs (or 91.96% of the keyword’s total installs). Ranked second on the same keyword, the brand “Call of Duty®: Mobile” receives 15,588 installs (or 5.95%) for the keyword “pubg”.

If we now compare the estimated keyword installs for the apps “Call of Duty®: Mobile” and “Call of Duty Companion App” for the keyword “call of duty”, we see that “Call of Duty®: Mobile” generates 236,847 installs (or 91.96% of the total installs) and the app “Call of Duty Companion App” generates 17,671 installs (or 6.86%). Both apps now have the keyword “call of duty” in their title and we see that “Call of Duty Companion App” gets 0.91% more installs on the keyword “call of duty” than “Call of Duty®: Mobile” gets on the keyword “pubg” (despite both apps ranking second on these keywords with equivalent popularity). This shows that when an app has some/all of a keyword in its title, this will result in higher conversion.

The closer the app title is to the keyword, the higher its conversion.


To summarize, AppTweak’s new keyword install estimations are now even more accurate thanks to:

  • More precise data from the Google Console (split by country and by day)
  • The model, now able to learn the relationships between different countries
  • Better predictions for branded keywords, accounting for the similarity between keywords and app titles

Don’t miss out on AppTweak’s unique KPI to start measuring the impact of your ASO today!

Start a 7-day free trial

by ,  Data Scientist

Lionel is a Data Scientist at AppTweak. He is passionate about machine learning and deep learning. He’s always on the look on how the newest state-of-the-art techniques can be applied to AppTweak’s data. He might try to lure you into cybersecurity and heavy metal.