How the figures on the home page are counted

The home page shows four numbers with a date. This page shows where each one comes from and how it is computed, so it can be recomputed by anyone with the same public pages. The installs and ratings headings are the home page figures, both counted on 13 October 2026, and the table under each one is that count listing by listing. The average is computed from the ratings table. The latency figure is Google DeepMind’s and is dated separately in its own section.

Installs: 1.41M+

Google Play publishes installs as a range, not a count. The figure is the sum of the lower bound of each range across the six listings published under our developer account, NEXT AI Lab. The App Store publishes no install count, so PDF AI's iOS installs are not included. That is why the figure is a floor and carries a plus sign. Why a floor rather than a midpoint or a third-party estimate is set out in Counting installs.

Ratings: 17,162

The number of ratings on the seven listings, six on Google Play and one on the App Store, added up on 13 October 2026. A rating count can fall as well as rise: a review removed by its author or by the store lowers it. Between 16 and 27 September two listings lost one rating each while the other five gained, so the figure is the count on the date beside it, not a floor, and it carries no plus sign.

The home page carried 15,936 dated 13 September 2026 until 16 September. That figure has been withdrawn rather than updated: 15,936 on 13 September and 16,539 on 16 September differ by 603, while the seven listings together gained 8 ratings between the evening of 15 September and 16 September. The two cannot both be right at that accrual rate, and nothing published lets a reader check the 13 September count, so it is gone. The corrections page records the change.

The scores and counts above come from the same captured listing's structured data. Google Play was requested in English with hl=en and gl=US; the App Store source is the US listing. These request settings do not establish the country or device population behind Google's scores. Apple's reviewCount field is counted as ratings, corroborated by the visible Ratings label and the US lookup count; its listing score of 4.8 is used, not the lookup score of 4.80128. Rounded display labels elsewhere on a listing are not calculation inputs. US App Store source: https://apps.apple.com/us/app/id6581489325

Average rating: 4.55

The listing-reported scores are weighted by their paired rating counts: the sum of (score × ratings) across the seven listings, divided by the total number of ratings. This is a listing-score metric, not a pooled lifetime average of individual users. The calculation retains the source precision shown above and rounds only the final result. With the 13 October 2026 values, 78127.972258377076384 ÷ 17162 = 4.552381555668…, which rounds to the 4.55 in the heading and on the home page. An average moves in both directions, so it carries no plus sign.

Dictation polish latency: 488 ms, down from 876 ms

Not our measurement. Google DeepMind reports the median latency of Cue's dictation polish step falling from 876 ms to 488 ms after that step moved to Gemma 4 E4B running locally through Ollama, measured on Apple Silicon over a 227-sample real-voice benchmark covering English and mixed-language input, figures current as of May 2026. The figure applies to the polish step only. Source: the case study.

Refreshing

The date beside the figures on the home page is the date those figures were counted; each table above carries the date it was last checked. Every change to a published figure is listed on the corrections page.