GPTBot
GPTBot is a training crawler operated by OpenAI. Its job is to read public web pages and feed that content into a machine-learning pipeline that trains future versions of the model.
Unlike a search-engine crawler, a training crawler does not send users back to your site. The content is consumed once, baked into the model, and shows up later in the model's responses. There is usually no citation and no referral traffic.
If OpenAI ships a new model version, you will likely see GPTBot traffic spike for a few weeks while it gathers fresh data, then quiet down again.
See GPTBot on your own site
Match the User-Agent header on incoming requests against the pattern below.
regex
Known UA examples
example 1
For higher confidence, also verify the source IP against the operator's published ranges. UA strings can be spoofed; IP ownership is harder to fake.
Renders JavaScript
No
IP verification
Published IP ranges
Crawl frequency
Periodic, broad
Honors robots.txt
Yes
Honors Crawl-delay
No
OpenAI runs 17 bots in total. Each one is a separate user-agent so you can allow or block them independently.
Agentic Browser
8Live-Fetch AI
2Training Crawler
2- GPTBotYou are here
- OpenAI
Agentic Commerce
2AI Search Index
1AI Coding Tool
1Link Unfurler
1Share of AI bot traffic
11%
↓ 63% vs 12 mo ago
% of top sites blocking it
62%
↑ 11% vs 12 mo ago
Top page categories crawled
Geographic origin of requests
Training Crawlers as a category, share of all bot traffic
68%
↓ 21% vs 12 mo ago
Sum of every training crawler we have public stats on. Compare to GPTBot's individual trend above to see if it's pulling ahead of its category or falling behind.
Crawl-to-click ratio
1,091 : 1
For every 1,000 fetches, this agent sends ~0.92 human visits back. Cloudflare crawl-to-click data
Snapshot updated
2026-04
Connect Agent Analytics for daily, per-site numbers.
Source: Cloudflare Radar AI Insights, Vercel 'Rise of the AI crawler' (Dec 2024), BuzzStream 2026 news-site blocking study
Should I let GPTBot through?
There's a real trade-off here. Training crawlers consume content without sending users back. Decide whether the trade for being in the model is worth your bandwidth. If OpenAI actually drives traffic or citations back to you, letting it through usually pays for itself. If it just consumes bandwidth, block it.
Does blocking GPTBot affect my Google rankings?
No. GPTBot collects training data, not search-index pages. Your classical search rankings stay intact. The actual trade is whether you want your content folded into the next model release.
How do I confirm a request is really from GPTBot?
Two checks. The User-Agent header should match a known GPTBot string, and the request's source IP should fall inside OpenAI's published ranges. The User-Agent alone is trivially spoofable, so the IP check is what gives you confidence. OpenAI publishes the ranges so you can validate at the CDN or edge.
What happens to my content if I let GPTBot fetch it?
It gets pulled into OpenAI's training pipeline and stored. Whether and how it influences a future model release is rarely disclosed. The only real lever you have on the outcome is what you allow at fetch time.
How is GPTBot different from OpenAI's other bots?
OpenAI splits work across multiple user-agents so site owners can decide on each one independently. Training crawlers, live-fetch agents, search indexers, and agentic browsers each get their own name. Worth scanning the rest of the OpenAI family above to see which ones actually matter for your site.
What's the cleanest way to control GPTBot?
Two layers. Robots.txt for the polite crawlers that read it, and rules at your CDN or edge for the ones that don't. Rankly's Agent Experience handles both from a single config, so you can allow, block, rate-limit, or serve a stripped-down version per bot. Agent Analytics handles the observation half so you know which bots are actually worth a rule.
Verify everything above against the operator's own documentation.