EZY Research · Study 2 · July 27, 2026
For every training crawl OpenAI makes, real ChatGPT users trigger four live visits
147,830
ChatGPT-User fetches
Live requests for a real person
39,042
GPTBot training crawls
Same 125 sites, 12 weeks
3.8x
Usage over training
Live fetches per training crawl
There are two kinds of OpenAI traffic in a server log. GPTBot crawls to collect training data. ChatGPT-User fires when a real person, mid-conversation, asks something that makes ChatGPT fetch your page right now.
Across 125 connected business sites over 12 weeks: GPTBot made 39,042 visits. ChatGPT-User made 147,830.
AI traffic to business websites is now roughly 4:1 usage over training. A live human is on the other end of most OpenAI hits. Add OAI-SearchBot's 26,181 search-index fetches and the picture sharpens: the majority of OpenAI's presence on these sites happens at answer time, when a citation can turn into a customer.
EZY Research · 125 connected sites · 12 weeks
OpenAI's traffic is usage, not training
3.8 live user-triggered fetches for every training crawl
3.8x
EZY Research · server-side logs · 27 Apr - 19 Jul 2026
OpenAI traffic composition
| OpenAI agent | Role | 12-week fetches |
|---|---|---|
| ChatGPT-User | Live fetch for a real user | 147,830 |
| GPTBot | Training crawl | 39,042 |
| OAI-SearchBot | Search index | 26,181 |
User-triggered AI fetch share
| Engine | Share of user-triggered AI fetches |
|---|---|
| ChatGPT | 96.0% |
| Claude | 3.2% |
| Perplexity | 0.8% |
The same server-side lens gives the cleanest market-share reading we know of, because it works even when browsers send no referrer. Of user-triggered AI fetches: ChatGPT 96.0%, Claude 3.2%, Perplexity 0.8%. Referral clicks tell the same story more loudly - 17,007 identifiable clicks from chatgpt.com in four weeks against 83 from Claude, Perplexity, Copilot and Gemini combined - though referral counts undercount every engine whose apps strip the referrer, so treat the fetch-based split as the honest one. Public data lands in the same place: the ChatGPT-vs-Google traffic tracker shows ChatGPT at roughly three-quarters of all AI-assistant referrals web-wide, and industry analyses published this month put its share of standalone AI referrals above 90%. Three methods, one verdict on concentration.
The practical shift: content strategy for AI is no longer about being in a training set someday. It's about being the page an assistant fetches in the four seconds after a customer asks. That page needs to load fast, answer in its first lines, and state facts an engine can lift.
This is the moment EZY optimizes for: we track the questions your customers ask (three runs per engine, because single answers vary), and build the pages those live fetches land on. The visit is already happening - the only question is whose page it lands on.
Methodology
Data: EZY.ai's 556-domain research panel - 125 connected business websites (WordPress and Cloudflare) with server-side logs, and 387 never-connected audited sites, 27 Apr-19 Jul 2026, UTC. Bots identified by user agent with signature verification where available. Panel skews toward small-business sites whose owners sought an AI-visibility tool. Aggregates only. Full tables available on request.
Aggregate CSVs for every table are available on request. The fixed-cohort check (42 domains present all quarter) is applied to all growth claims; raw-panel growth is reported only alongside it. Dataset license: CC BY 4.0.