The panel is made up of 5,000 US YouGov panelists who used ChatGPT at some point in 2026 and opted to share their chat histories. YouGov selected them randomly from that eligible pool. The sample is unweighted and need not be representative of US adults or ChatGPT users overall.

Not necessarily, no. These curves describe a fixed group selected for having used ChatGPT in 2026. Earlier users with no activity that year are not observed in our data, while people who stopped and returned can be included. “First recorded ChatGPT use” can only rise or stay flat because it is cumulative.

We received one row per message, with a hashed panelist identifier, a hashed chat identifier, the timestamp, the role (prompt or reply), the model that produced a reply, the name of any tool used, the account plan, and the panelists’ self-reported demographics. Message text was replaced with its word and character count before delivery. At no point did we have access to transcripts, and only aggregate statistics are published.

Panelists exported their histories between January and August 2026, so coverage ends at different dates for different people. December 2025 is the last full calendar month before any of these exports. The “Show incomplete data” setting extends the charts into 2026, when declining activity can reflect histories ending rather than people stopping use.

In 2026, Messages per active panelist is computed over only active panelists whose data extends beyond a period. Note: the number of panelists decreases mechanically in this period and the behavior of the remaining panelists may not be representative of the panel as a whole. The downloadable files include these periods.

A panelist who sent at least one recorded prompt during a time period. The share of active panelists uses all 5,000 panelists as its denominator, regardless of when their recorded histories begin. It measures activity within this selected panel, not the proportion of Americans using ChatGPT.

It shows the cumulative share of panelists whose first recorded prompt falls on or before the end of a period. This identifies when activity first appears in their shared histories, which may differ from when they first used ChatGPT.

Not necessarily, no. Participation was opt-in, the sample is unweighted, and everyone in the panel used ChatGPT at some point in 2026. These limitations affect both reported levels and changes over time. Comparisons within the panel should not automatically be generalized to other users.

Subgroups are available for monthly data and for panelist-level metrics only. Weekly series and conversation-level metrics are not broken down.

We group panelists by the account plan associated with a message at export, using the plan attached to the largest number of their prompts. This is a fixed grouping, not a historical record of subscription status. About 10% of panelists have more than one plan across their messages. For this reason, we use the most common value as a fixed attribute of the panelist: the plan recorded on most of their prompts, with ties broken alphabetically. We group these plans into Free, Plus, and Go and other paid (Go, Pro, Pro Lite, Team, Business, Enterprise, Edu).

A value in the data files is blank if it has been suppressed due to a small number of underlying observations. See Blank values in the Methodology.

We classify a reply as coming from a reasoning model when its model identifier names a Thinking, Thinking mini, or Pro variant, or an o-series model. A conversation counts as a reasoning conversation when more than half of its replies with a model identifier meet this rule.

The full dataset is available as an Excel workbook and as three long-format CSV files, listed on the Downloads page. The data is released under a Creative Commons Attribution license. Please credit both Epoch AI and YouGov, for example “Source: Epoch AI / YouGov, 2026”. AI not authorized by YouGov is not authorized to ingest or scrape this data.