The Day ChatGPT Changed How It Counts Your Users

On August 13, ChatGPT queries from 68 different marketing teams asked GA4 for the users metric through our platform. On August 14, that number was 22. By the next week it was in the single digits. Nothing changed on our side, and no announcement appeared on OpenAI’s. The teams themselves did nothing: this happened to dozens of independent companies on the same day.

Follow-up to Claude vs ChatGPT: What 35,000+ Marketing Teams Do With Each. Daily distinct-team counts, Aug 1 to 26, 2026, via Windsor.ai’s MCP integration. No customer-identifiable data.

This is a story about what it means to build reporting on top of an AI assistant: the definitions underneath your numbers can change overnight, and there is no changelog to warn you.

What we saw

We track which GA4 fields each AI client requests. GA4 has several metrics that all describe “users”: users, totalusers, and active_users. On August 14, the mix of what ChatGPT and Codex requested flipped in a single day:

ChatGPT + Codex	Aug 13	Aug 14	Aug 26
Teams requesting users	68	22	17
Teams requesting active_users	69	99	163
Teams requesting totalusers	75	35	44

The active_users metric went from 1 option among 3 to the default. users collapsed and never recovered. This was a step change; in the 2 weeks since, the mix has not reverted.

The flip is client-specific: the ChatGPT/Codex mix inverts on Aug 14 and stays that way; Claude’s mix keeps its shape through the same weeks (its dips are weekends).

The same day, ChatGPT’s calls to our GA4 field-metadata endpoint (the “what fields exist?” question an AI asks before querying) roughly doubled overnight and stayed at the higher level. Whatever changed, it made the client re-read the schema, en masse, and then choose differently.

The Claude control group

Claude queries the same GA4 connector, through the same integration, against the same field list. Claude’s mix did not move: totalusers first, active_users second, users third, in stable proportions before, during, and after August 14.

A change on our side would have moved both clients. It moved one, which points the investigation outward.

What we ruled out

We went looking for a cause on our side first, in this order:

1. Our GA4 field list. It has been unchanged since March 2026; the only additions all year were property_currency (July 16) and property_timezone (July 28), neither of them related to user counts.
2. Our ChatGPT app listing. The listing was live well before the flip, and the listing-improvement work we did happened on August 22, a week after.
3. A platform-wide schema change. The Claude control group above rules this out: Claude reads the same connector and its mix did not change through the same weeks.
4. A client bug or error spike. ChatGPT’s GA4 error rate is flat across the whole month (10 to 15%), including the field-guessing errors you would expect if the client had started inventing field names. The queries were valid on both sides of the change; only the chosen metric differs.

OpenAI’s changelogs and release notes mention no change to tool or connector behavior on that date.

What probably happened

Everything we can rule out sits on our side; what remains points to a change in how the OpenAI clients decide which fields to request: a tool-schema cache refresh, a system-prompt adjustment, or a model update, none of which are announced at this level of detail. From the outside we cannot confirm the cause; we can only show that every explanation on our side is eliminated.

A plausible trigger is the GPT-5.6 rollout, which OpenAI began on August 6 and continued through that week (ChatGPT release notes).

The new choice matches GA4 itself: active_users is GA4’s primary user metric, the one its interface reports. If this was a deliberate fix on OpenAI’s side, it was a good one. The problem is that it arrived unannounced.

Did anyone’s numbers break?

We tested the numeric impact on our own GA4 property through the same connector, and this time the answer is no: in our schema users is an alias for active_users, so both fields return identical values and the switch was numerically invisible to the affected teams.

The luck is in which metric the client landed on. totalusers, the user metric Claude prefers, runs about 15% higher than active_users on our property. A silent flip involving that field, the same class of change, would have moved every user count by double digits mid-report, with nothing to announce it.

Why this matters for your reporting

The lesson applies to every AI assistant, not one vendor:

1. Pin the metric, not the concept. Prompts that name the exact field (“use active_users”) are immune to this class of change. Prompts that say “users” delegate the choice to a model whose defaults can move.
2. Recurring reports need recurring definitions. If a number feeds a dashboard or a weekly email, the field list belongs in the prompt or the saved workflow, stated explicitly.
3. Watch for step changes in your own metrics. A sudden level shift in a tracked number with no campaign change behind it is a reason to check what the assistant queried, before anyone reruns media plans over it.

Methodology

Daily counts of distinct teams requesting each GA4 user metric via Windsor.ai’s MCP integration, August 1 to 27, 2026, split by client user agent (ChatGPT/Codex vs Claude as control). Team counts per day, not query counts, so a handful of heavy users cannot move the line. All figures aggregated; no customer-identifiable data.

Want your marketing data behind whichever assistant your team uses? Windsor.ai is natively listed in Claude’s connector directory and as a native ChatGPT app: one-click Add, sign in, pick a connector. See how to bring your marketing data into ChatGPT or connect your marketing data to Claude. Free forever, no credit card. Paid plans from $19/mo.