Gemini's Share Fell From 12% to 1.9%, on One Dataset That Counts Text

Pangram, an AI text-recognition platform, recorded Gemini's share of the texts submitted to it falling from 12% to 1.9% during July 2026. In the same dataset OpenAI has held over 50% every month since tracking began, and Anthropic climbed from 4.3% to 14.9%, concentrated in technical and scientific writing.

What each dataset actually measures

• Pangram measures model shares across texts submitted for AI detection. It is weighted toward writing tasks, which is a category where Gemini has historically underperformed.

• OpenRouter shows a similar shape on very different usage patterns, and its mix skews toward open-weight models.

• Similarweb measures website traffic and excludes mobile apps. On its numbers Google's share moved from 27% to 26.8% over the last month, against 9.4% a year ago.

The Decoder names the blind spot in each source, which is the right call. They are not three measurements of one thing.

What this means for you

A 12% to 1.9% drop in a text-detection sample is not a collapse in Gemini usage, and it is not a reason to move optimisation effort off Gemini. The gap between the Similarweb trend and the Pangram trend is the whole lesson: the same month reads as a collapse or as flat depending on what got counted.

For deciding where to put effort, the only number that pays is your own. Which engines cite you, on the prompts your buyers actually type. We run that per client precisely because published share figures cannot answer it, and because a model with a small share of writing tasks can still be the one recommending you in your category.

Treat all three of these as weather reports on the model market. Useful for context, and not an input to your own allocation.