Three measurements, one direction
Pangram measures model shares from the texts submitted to it, and reported that Gemini's share fell from 12 per cent to 1.9 per cent in July 2026; Pangram calls that drop a collapse. In the same measurement OpenAI has stayed above 50 per cent every month since tracking began, and Anthropic rose from 4.3 per cent to 14.9 per cent. On the Similarweb side, Google's website share slipped in the preceding month from 27 per cent to 26.8 per cent, and a year earlier the same figure was 9.4 per cent.[1]
The three numbers do not measure the same thing, and The Decoder writes out each one's blind spot. Pangram mostly sees writing tasks; OpenRouter's traffic skews toward open-weight models; Similarweb leaves mobile apps out. They converge on direction and diverge on size. What a reader is left holding stops right there, at the direction.[1]
What does Google's number count?
The 1 billion monthly Gemini App users Google reports is a measure too, yet it counts something else entirely: the number of people who opened the app within a month. None of the three independent measurements looks at that surface, and Similarweb leaves the mobile app out to begin with. This column drew the same distinction for Bluesky's registered-account figure: a count of registered accounts shows neither active use nor how many people actually open the product.[1], [2]
A shrinking share in Pangram's sample has two separate readings. People may genuinely be choosing Gemini less; or Gemini use may have moved to surfaces that never leave as submitted text: inside a search result, inside office apps, inside the assistant on a phone. In the second case Pangram's table would still show a decline while use had not fallen. The three measurements at hand do not carry the data that would separate those two readings.[1]
What is left for the reader?
The question these numbers can answer for a reader is narrow. They do not show which assistant does an everyday job better, and they do not directly say how much a tool is used. What they do say is this: in a writing-heavy sample Gemini's share fell sharply, and three separate methods pointed the same way. For someone deciding whether to open a product, these numbers describe the surfaces on which the product can be measured, and what they say about the product itself is limited.[1]
The one thing that would close this gap is a study that measures mobile assistant use with its method disclosed. If such a measurement is published by 30 November 2026, it becomes testable whether the fall in Gemini's share sits in use itself or only in the surfaces where text is submitted. Google could do the same job: placing alongside the monthly user count a measure of how many people come back more than once within a month would be enough.[1]