MethodologyMethodology
The data explorer is built from the message-level ChatGPT histories of 5,000 US members of a YouGov panel who opted in to share them. This page explains how the data was collected, what it contains, how each metric is defined, and what data limitations exist.
Data collection
YouGov’s eligible pool was its US panelists who had opted in to share their ChatGPT history with YouGov. YouGov restricted the pool to panelists with at least one ChatGPT message dated on or after 1 January 2026. From that pool, YouGov selected 5,000 panelists at random. Each panelist exported their recorded ChatGPT history and shared it with YouGov. Exports took place on a rolling basis between January and August 2026.
Before delivery to Epoch, YouGov replaced the text of every message and chat title with a placeholder giving its word and character count. At no point did we have access to any transcripts. YouGov also attached the self-reported demographics it holds for each panelist. Twenty-two panelists report an age under 18. They are included in all panel-level figures and form the Under 18 subgroup in the age breakdown.
Who is in the panel and what that means
The panel is a fixed group of 5,000 people selected for having used ChatGPT at some point in 2026. Earlier users who had no activity in 2026 are not observed in our data, but the panel can include people who stopped and later returned. Earlier users had more time to stop before 2026, which may affect the composition of the earlier observations. We cannot determine how much this selection shapes the curves.
“First recorded ChatGPT use” shows when activity first appears in these panelists’ shared histories. By December 2025, 86.5% had a recorded prompt. The remaining 13.5% first appear in 2026. The share of panelists active in a given period can rise or fall.
Read these charts as the recorded histories of this selected group. Measures per active user or per conversation also describe this group and may be affected by the selection process.
Contents of the raw data
Each row of the raw data is one message. A row records a hashed panelist identifier, a hashed chat identifier, the message timestamp, the role that produced the message, the model identifier of a reply, the name of any tool the message called, the account plan associated with the message at the time of export (see Subgroups), and the panelist’s demographics. Message text is present only as a word and character count.
Each message carries one of three roles in the delivered data: user, assistant, or tool. We call user messages prompts and assistant messages replies. ChatGPT’s export keeps every version of a message: when a user edits a prompt or sends it again, the earlier version stays in the history next to the new one. About 2% of prompts follow another prompt with no reply in between, which suggests that they fall into this category of edited or resent messages. Tool messages are outputs returned to the model, such as a web search result, and are not counted as prompts or replies.
The raw data holds about 8.26 million messages of all three roles, from 5,000 panelists, across about 660,000 conversations. The messages are dated December 2022 to August 2026.
Each row also contains the account plan recorded for the message at the time the chat history was uploaded to YouGov: free, plus, go, or one of several team, business, education, and other labels. Taking the plan with the largest number of a panelist’s prompts, 4,461 panelists are on Free, 449 on Plus, and 90 on Go or another paid plan, meaning about 11% have a paid or organizational plan. This grouping is the “Plan (most common)” subgroup variable.
Model identifiers on replies are absent in December 2022 and January 2023, present on most replies in February 2023, and complete from March 2023. The metrics that depend on the reply model exclude conversations with no recorded model, so their values in the earliest periods are based on few conversations.
Time coverage
Panelists exported their histories between January and August 2026, so 2026 is only partly observed for the panel as a whole depending on when the history was exported. The number of panelists with observed data falls from 5,000 in January 2026 to about 1,450 in July 2026 as more panelists reach the end of their export period. However, the Active panelists share uses the full panel as its denominator throughout 2026, so the share declines mechanically as fewer panelists remain observable. We cannot tell from the data whether an unobserved panelist stopped using ChatGPT or simply reached the end of their export period.
For this reason, we show only the period before 2026 by default in the data explorer. The “Show incomplete data” setting extends the charts through the last period in the data, with the incomplete periods dashed or hatched. The downloadable files include these periods. In our checks, the panelists who were observed longest into 2026 were lighter users already in 2025, so heavy usage is unlikely to be artificially inflated by excluding the 2026 data. To account for this partial data, we calculate Messages per active panelist in 2026 using only panelists who were active in a period and had not reached their last observed period.
Selection based on 2026 activity affects every period shown, including 2022–2025. Rolling export dates create a separate problem because 2026 is only partly observed. Ending the default charts in December 2025 avoids that coverage problem but does not remove the selection effect.
We aggregate messages into calendar months and weeks starting on Monday. Each message is assigned by its timestamp. Timestamps are treated as UTC. A conversation is active in a period if it has at least one message of any role in that period.
Demographics and use pattern comparison
Each pair shows our panel’s estimate followed by the comparison group’s estimate. For example, “14.1% vs. 28.7%” means that 14.1% of the panelists included in that comparison were aged 18–29, compared with 28.7% of ChatGPT users in the Ipsos survey. We use different panel groups to approximate each source’s user population and reference period. A dash means no comparable estimate is available.
| Measure | Census/CPS 2026 | Epoch/Ipsos March 2026 | Pew February 20261 | OpenAI usage paper |
|---|---|---|---|---|
| Source’s population and reference period | US adults, 2026 Current Population Survey Annual Social and Economic Supplement | Past-week ChatGPT users surveyed March 3–5, 2026 | ChatGPT ever-users surveyed February 17–23, 2026 | Global weekly active ChatGPT users, July 2025 for gender and prompt intensity |
| Panel group and period used | All eligible adult panelists, using recorded 2026 demographics | Panelists active February 25–March 3, 2026, approximating the survey’s preceding week | Panelists with any recorded use by February 20, 2026, the survey midpoint | Weekly active panelists during July 7–August 3, 2025, covering four complete weeks |
| Demographics | ||||
| Aged 18–29 | 14.4% vs. 20.2% | 14.1% vs. 28.7% | 14.5% vs. ≈28% | – |
| Aged 30–49 | 40.9% vs. 33.0% | 41.9% vs. 39.2% | 41.0% vs. ≈42% | – |
| Aged 50–64 | 26.4% vs. 22.3% | 26.0% vs. 22.4% | 26.3% vs. ≈19% | – |
| Aged 65+ | 18.3% vs. 24.5% | 18.0% vs. 9.7% | 18.2% vs. ≈11% | – |
| Female2 | 54.4% vs. 51.7% | 54.8% vs. 55.0% | 54.2% vs. ≈52% | 53.0% vs. 52.4% |
| Bachelor’s degree or higher, aged 25+ | 55.0% vs. 40.2% | 59.4% vs. 51.9% | – | – |
| Hispanic3 | 8.2% vs. 18.9% | 8.4% vs. 18.9% | – | – |
| Plans4 | ||||
| Paid plan or paid access | – | 15.6% vs. 19.8% | – | – |
| Free plan or no paid access | – | 84.4% vs. 76.1% | – | – |
| Usage5 | ||||
| Used on one day in the preceding week | – | 36.4% vs. 34.0% | – | – |
| Used on 2–5 days in the preceding week | – | 51.7% vs. 49.4% | – | – |
| Used on 6–7 days in the preceding week | – | 11.9% vs. 16.6% | – | – |
| Mean prompts per calendar day among weekly active users, July 2025 comparison6 | – | – | – | 4.28 vs. ≈3.67 |
| Growth in prompt volume, June 20–26, 2024 to June 20–26, 20257 | – | – | – | 5.58× vs. 5.82× |
The activity windows are approximate matches, except for prompt-volume growth, which uses the same dates in both datasets. Panel demographics come from recorded 2026 profiles and have not been reconstructed for earlier dates.
Metric definitions
Active panelists
Share of the 5,000 panelists who sent at least one prompt in a period. The denominator is always the full panel, or the full subgroup in the subgroup tables, regardless of when a panelist started using ChatGPT. This measures earlier activity within a panel selected for use in 2026. It does not estimate the share of Americans using ChatGPT or overall ChatGPT adoption. Because every panelist was selected for having used ChatGPT in 2026, and some of them sent their first prompt in 2026, the Active panelists share cannot approach 100% in any published period.
First recorded ChatGPT use
Cumulative share of the 5,000 panelists whose first recorded prompt falls on or before the end of a period. The complementary category contains panelists with no prompt yet recorded. By December 2025, 86.5% had a recorded prompt. The remaining 13.5% first appear in 2026. These dates describe only the observed data and may differ from when panelists first used ChatGPT.
Panelists by days active
Active panelists grouped by the number of distinct days on which they sent a prompt during a period. Monthly bands are 1 day, 2 to 5 days, 6 to 10 days, 11 to 20 days, and 21 or more days. Weekly bands are 1, 2, 3, 4, and 5 to 7 days. Shares represent the percentage of active panelists within each band, and sum to 100%.
Messages per active panelist
Prompts plus replies per active panelist in a period. Tool messages are not counted. We compute the mean and median number of messages.
Prompts per conversation
Prompts per conversation, over conversations active in a period. A conversation with messages but no prompt in a period counts as zero prompts. We compute the mean and median number of prompts.
Most-used model
Each conversation is assigned to the model that produced the largest share of its replies in a period, with ties broken alphabetically. Models are bands of the model identifiers on replies, which group a model with its named variants (for example Thinking, Instant, and Pro). For example, GPT-5 covers gpt-5, gpt-5-thinking, gpt-5-instant, gpt-5-auto-thinking, and gpt-5-pro. The bands are listed below.
Shares represent the percentage of conversations with at least one reply carrying a model identifier and sum to 100%. Conversations with no recorded model are excluded. On the data explorer, the legend groups the bands by model family. The charts show a band only if it has a non-zero share in at least one of the displayed periods. The models in the GPT-5.3 to GPT-5.6 bands were released in 2026 and appear in the data from March 2026, so these bands are shown only when “Show incomplete data” is selected in the data explorer. The downloadable files include their rows for every period.
| Band | Model identifiers |
|---|---|
| GPT-3.5 | text-davinci-002-render and variants, gpt-3.5-turbo |
| GPT-4 | gpt-4, gpt-4-gizmo, gpt-4-plugins, gpt-4-code-interpreter, gpt-4-dalle, gpt-4-browsing, gpt-4-mobile, gpt-4l |
| GPT-4.1 | gpt-4-1 |
| GPT-4.1 mini | gpt-4-1-mini |
| GPT-4.5 | gpt-4-5 |
| GPT-4o | gpt-4o, gpt-4o-canmore, gpt-4o-jawbone |
| GPT-4o mini | gpt-4o-mini |
| o1 | o1-preview, o1, o1-pro |
| o1-mini | o1-mini |
| o3 | o3, o3-pro |
| o3-mini | o3-mini, o3-mini-high |
| o4-mini | o4-mini, o4-mini-high |
| GPT-5 | gpt-5, gpt-5-thinking, gpt-5-instant, gpt-5-auto-thinking, gpt-5-pro |
| GPT-5.1 | gpt-5-1 and its thinking, instant, auto-thinking, and pro variants |
| GPT-5.2 | gpt-5-2 and its thinking, instant, and pro variants |
| GPT-5 mini | gpt-5-mini, gpt-5-t-mini, gpt-5-a-t-mini |
| GPT-5.3 | gpt-5-3, gpt-5-3-instant |
| GPT-5.3 mini | gpt-5-3-mini |
| GPT-5.4 (incl. mini) | gpt-5-4-thinking, gpt-5-4-pro, gpt-5-4-auto-thinking, gpt-5-4-t-mini |
| GPT-5.5 | gpt-5-5, gpt-5-5-thinking, gpt-5-5-instant, gpt-5-5-pro |
| GPT-5.5 mini | gpt-5-5-mini |
| GPT-5.6 (incl. mini) | gpt-5-6, gpt-5-6-thinking, gpt-5-6-pro, gpt-5.6-sol-wm, gpt-5-6-mini, gpt-5-6-t-mini, gpt-5-6-t-mini-mini |
| Other | research (Deep research), alpha and gated test identifiers, and any identifier not listed above |
Tool use
Conversations grouped by the set of tool families used in a period. Families are image generation, web search, file search, memory, canvas, and other tools. A conversation that used exactly one family falls in that family’s band, for example “Web search only”. A conversation that used more than one family is “Multiple tools” and one that did not use any tools is “No tools”. Shares represent the percentage of conversations active in a period and sum to 100%.
Reasoning model use
Share of conversations in which more than half of the replies with a model identifier came from a reasoning model. Reasoning models are the Thinking, Thinking mini, and Pro variants and the o-series models, identified from the model identifier on each reply. Conversations with no recorded model are excluded. Conversations in which only an occasional reply came from a reasoning model are not counted.
Subgroups
Breakdowns by subgroups are available for monthly data and for panelist-level metrics only. The monthly subgroup file breaks down Active panelists and Messages per active panelist by gender, age, education, household income, employment status, and plan. On the data explorer, the “Active panelists by plan (most common)” chart draws the Active panelists series of the plan subgroups; the other subgroups are available in the file. The weekly table and the conversation-level metrics are not broken down. The subgroups are listed on the Records page. Panelists with no recorded value for a subgroup variable are left out of that breakdown, so subgroup sizes do not always sum to 5,000.
Plan (most common) groups panelists by ChatGPT plan. Each message carries the account plan associated with the message at the time of the export that included it. This is the plan recorded for a message at export, so a history exported once in 2026 by a Go panelist may show Go on messages from 2023, before the Go plan existed. About 10% of panelists have more than one plan across their messages. We assign each panelist the plan attached to the largest number of their prompts across all dates, breaking ties alphabetically, and map it to Free, Plus, and Go and other paid (Go, Pro, Pro Lite, Team, Business, Enterprise, Edu). Go and other paid never reaches 100 active panelists in a month, so Messages per active panelist is not reported under the rule below. Its Active panelists share is still reported.
Blank values
A value in the data files is blank if it has been suppressed due to a small number of underlying observations. These values are excluded from the graphs in the data explorer. Messages per active panelist is suppressed when it is calculated with a base of less than 100 active panelists. Active panelists has the full panel or the full group as its base and therefore shows the Go and other paid group, though it has only 90 panelists. Conversation-level metrics are reported regardless of the number of active panelists. Most-used model and Reasoning model use are also blank in December 2022 and January 2023, when no reply carries a model identifier.
Model and tool release dates
The tables below record the start of publicly-documented rollouts in ChatGPT. These dates may precede availability to all types of accounts or subscription plans. Model dates refer to availability in ChatGPT, which can differ from announcement dates. Sources are linked in the tables.
For tools, we show one initial rollout marker each for DALL·E 3 image generation, saved memory, ChatGPT search, and Canvas. The web-search marker refers specifically to ChatGPT search, rather than to the earlier Browse with Bing feature. Later upgrades and expansions to additional plans are not marked. We omit rollout dates for File search because we could not establish a launch date for the specific tool family recorded in the exports, and for Other tools because it combines features with different launch dates.
When “Show model/tool releases” is on, the Most-used model and Tool use charts mark these dates with dashed vertical lines labeled with the band or tool name. GPT-3.5’s rollout precedes the first period with model identifiers, so its line is not drawn, and the 2026 bands and Other have no line.
| Model | ChatGPT rollout start | Note |
|---|---|---|
| GPT-3.5 | 2022-11-30 | ChatGPT launch |
| GPT-4 | 2023-03-14 | Plus |
| GPT-4.1 | 2025-05-14 | Plus, Pro and Team |
| GPT-4.1 mini | 2025-05-14 | Includes Free fallback |
| GPT-4.5 | 2025-02-27 | Pro research preview; Plus followed |
| GPT-4o | 2024-05-13 | Free and paid rollout |
| GPT-4o mini | 2024-07-18 | Free, Plus, and Team |
| o1, including o1-preview | 2024-09-12 | Full o1 followed on 2024-12-05 |
| o1-mini | 2024-09-12 | Plus and Team |
| o3 | 2025-04-16 | Plus, Pro, and Team |
| o3-mini | 2025-01-31 | Includes Free via “Reason” |
| o4-mini | 2025-04-16 | Includes Free via “Think” |
| GPT-5 | 2025-08-07 | Free and paid rollout |
| GPT-5.1 | 2025-11-12 | Paid users first |
| GPT-5.2 | 2025-12-11 | Paid users first |
| GPT-5 mini | 2025-08-07 | Fallback after usage limits |
| Tool | Initial rollout date | Note |
|---|---|---|
| Image generation | 2023-10-16 | DALL·E 3 beta rollout |
| Memory | 2024-02-13 | Saved-memory test begins |
| Web search | 2024-10-31 | ChatGPT search launches for Plus and Team |
| Canvas | 2024-10-03 | Canvas beta for Plus and Team |
Limitations
- Participation was opt-in and the sample is unweighted. Results should not be generalized to US adults or to all ChatGPT users.
- Every panelist used ChatGPT at some point in 2026. Earlier periods describe those people’s recorded histories, so earlier users who had no activity in 2026 are not observed in our data. This selection can shape both activity levels and trends.
- Model identifiers are missing before February 2023.
- Demographics are self-reported to YouGov and are not weighted to any population.
- Usage is highly concentrated. The 10% most active panelists sent 63% of all prompts, and the single busiest panelist-day holds 698 prompts. The panel cannot distinguish one heavy user from several people sharing an account.
-
Pew user-composition estimates are derived from its published adoption rates and Census population shares.
-
The panel records self-reported gender, CPS records sex, and OpenAI classifies first names, excluding unknown classifications.
-
Ethnicity comparisons use a provisional mapping between panel and survey categories.
-
Panel plans are the most common labels across recorded prompts, not verified plans held at the survey date. Ipsos measures paid access at interview, including employer- or school-funded subscriptions.
-
The panel measures ChatGPT use. Ipsos measures any AI use among ChatGPT users.
-
OpenAI’s estimate is calculated from rounded global weekly totals. The panel estimate pools four complete weeks from July 7 to August 3, 2025. Both count user prompts per calendar day among weekly active users.
-
Both estimates compare June 20–26, 2024 with June 20–26, 2025. OpenAI covers global consumer users. Our panel is selected for activity in 2026, so similar growth does not establish that it represents overall ChatGPT usage trends.