Claude vs ChatGPT: What 35,000+ Marketing Teams Do With Each

1 in 9 data queries on our platform now comes from an AI assistant instead of a human clicking around a dashboard. We looked at a week of that traffic to answer a question that so far gets answered with opinion instead of data: when marketers hand their data to Claude versus ChatGPT, what do they do differently?

Data: aggregated, anonymized usage of Windsor.ai’s MCP integration, one full week (Aug 20 to 27, 2026). All figures are percentages, ratios, or per-team rates. No customer-identifiable data.

The short answer: almost everything. The two ecosystems differ in who uses them, how hard they push, which ad platform they trust the AI to touch, and even in which metric they mean when they say “users.”

Key findings

• AI assistants now generate roughly 11% of all marketing data queries on our platform, and rising.
• Claude is the default choice for marketing data analysis. Teams analyzing their data through Claude outnumber teams using OpenAI’s clients roughly 7 to 1, and Claude accounts for about 81% of all AI-driven data queries.
• Inside the OpenAI ecosystem, Codex has overtaken ChatGPT. More teams now query their marketing data from OpenAI’s coding agent than from the ChatGPT consumer app. Marketing data is turning into something developers wire into agents and automated pipelines.
• OpenAI users query about 35% more intensively. ChatGPT and Codex teams average roughly 33 data queries per week, versus about 25 for Claude teams. Claude’s user base is much wider; OpenAI’s is more concentrated and heavier per seat.
• Claude and ChatGPT literally count your users differently. Ask both “how many users did my site get last month” and they will often pull different GA4 metrics and give you different numbers (details below).
• When it comes to write access, Claude leans Google, ChatGPT leans Meta. And the most widespread thing marketers let any AI do to an ad account is delightfully cautious: adding negative keywords.

Where marketers point each assistant

Both ecosystems concentrate on the same core stack: Meta Ads, Google Ads, Instagram, GA4, and Search Console make up the top 5 for both. The differences are at the margins:

Data source	Claude teams	OpenAI-client teams
Meta Ads (Facebook & Instagram Ads)	35%	48%
Instagram	34%	26%
Google Ads	33%	25%
Google Analytics 4	19%	18%
Search Console	11%	12%

Share of each client’s active teams that queried the connector at least once during the week. Teams can appear in multiple rows.

Claude’s audience spreads wider into organic and social sources (Instagram, TikTok Organic, Google Business Profile, YouTube), while OpenAI’s user base is more heavily concentrated on Meta Ads. Both treat GA4 and Search Console almost identically: unified web data is the common denominator no matter which AI you ask.

The metric vocabulary problem: “users” is not “users”

This is our favorite finding, because it changes what you see when you compare AI answers.

When teams pull GA4 data, both ecosystems agree on the #1 metric: sessions (requested by over 90% of GA4-querying teams on both sides). But on the #3 concept, user counts, they diverge:

• Claude overwhelmingly requests totalusers (39% of its GA4 teams). The active_users metric does not even crack Claude’s top 15.
• ChatGPT and Codex overwhelmingly request active_users (61% of their GA4 teams, their #3 field overall).

Both are legitimate GA4 metrics, and they produce different numbers. So a marketer who asks Claude and ChatGPT the same question about “users” can get two different answers, both technically correct. If your team standardizes on one assistant, this never bites you. If you mix them, it will.

(We have also watched this vocabulary shift happen overnight, platform-wide, with no announcement from the AI vendor: read about the day ChatGPT changed how it counts your users.)

Beyond the users split, OpenAI clients request measurably richer queries: engaged sessions, engagement rate, source/medium, and e-commerce purchase fields all show up in a larger share of their GA4 queries than Claude’s. That matches the intensity finding: fewer OpenAI teams, but each one digs deeper per session.

Write access: what marketers let each AI touch

Beyond reading data, both ecosystems can act on ad accounts through Windsor.ai (Google Ads, Meta Ads, TikTok Ads, LinkedIn Ads, Microsoft Ads, plus Instagram and Google Business Profile posting). Here the personalities split sharply:

Claude users lean Google. The most-used write actions among Claude teams are Google Ads operations: creating ad assets, building responsive search ads, and pushing keywords. A dedicated cluster of Claude power users has also automated Google Business Profile review replies at serious scale.

OpenAI users lean Meta. The most widespread write actions among ChatGPT and Codex teams are Meta operations: creating campaigns, ad sets, and ads.

Everyone starts with the safety rail. Measured by how many different teams use it, the single most widespread write action across Claude users is pushing negative keywords. The first change most marketers allow an AI to make in a live ad account blocks wasted spend and touches nothing else. From there, teams expand write access a rung at a time:

1. Negative keywords: blocks wasted impressions and leaves everything else untouched. The most widespread write action.
2. Keyword pushes: additive but low-stakes, easy to audit and undo.
3. Budget changes: spend moves up or down, within campaigns that already exist.
4. Full ad creation: campaigns, ad sets, creatives. The AI builds from scratch.

Success rates favor Claude, modestly. Write actions initiated from Claude succeed 90.8% of the time versus 86.6% from OpenAI clients. On the read side the gap nearly vanishes: ChatGPT errors on about 8% of queries, Claude about 9%, Codex about 10%. Roughly 1 in 11 AI data queries fails regardless of vendor, usually from the AI requesting fields that do not exist or over-broad date ranges. Near-parity here is itself news: neither side has “solved” talking to marketing APIs.

What this means for marketing teams

1. Pick one assistant per workflow and standardize. The users/active_users split means mixed-assistant teams will produce mismatched reports through no one’s fault.
2. Climb the trust ladder in order. Start AI write access with reversible, defensive operations (negative keywords, pausing) before letting it create campaigns.
3. Watch the Codex signal. Marketing data querying is shifting from chat interfaces into coding agents and automated pipelines. The teams doing this run 35%+ more queries. If your competitors have wired their reporting into an agent, they are iterating faster than you.

Methodology

Aggregated MCP query and action logs across 350+ marketing data connectors on Windsor.ai, one full week (Aug 20 to 27, 2026), covering 35,000+ active teams. Clients identified by MCP user agent (Claude; ChatGPT; Codex; Agent Builder, excluded from most cuts due to small base). All figures reported as shares, ratios, or per-team rates. No individual customer data was accessed for this analysis beyond automated aggregation, and none is published.

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