2026 AI Search Shift

AI Crawler Control Center

Cloudflare now lets every site decide — per bot — whether AI crawlers get in free, get blocked at the edge, or pay per crawl. Set the switches and watch what happens.

Drag to orbit · pinch or scroll to zoom

Three kinds of AI crawl

They look identical in your logs but do very different things:

  • Training crawls — bulk-collect pages to bake into future model weights. Your content becomes the model; no link back.
  • Search / RAG crawls — index pages so AI answers can cite and link you (this is where visibility lives).
  • User-triggered fetches — an assistant grabs one page live because a user asked. Closest to a human visit.

robots.txt vs edge enforcement

robots.txt is a polite note on the door — compliant bots read it, others walk past. Managed enforcement at a CDN edge like Cloudflare actually verifies bot identity (signatures, IP ranges, behavior) and drops or challenges the request before it touches your origin. A rule you can't enforce is a preference, not a policy.

Pay-per-crawl

The middle switch. Instead of a binary allow/block, the edge returns HTTP 402 Payment Required with a price; crawlers that present payment credentials get the page, and you get paid per fetch. It turns "scrape me for free" into a licensing market — especially relevant for publishers whose traffic is being replaced by AI answers.

The visibility tradeoff

Block every AI bot and your content stops leaking — and you also vanish from ChatGPT, Claude, Perplexity and Google AI answers, where a growing share of discovery now happens. The strategic question isn't "block or not," it's which crawl types create value for you and which just extract it. Most sites land on: allow search/RAG, control or charge training.

Know your crawlers

User-agentOperatorPrimary purposeWhat allowing it gets you
GPTBotOpenAIModel training corpusLittle directly — your text may inform future GPT models
OAI-SearchBotOpenAIChatGPT search indexCitations & links inside ChatGPT answers
ChatGPT-UserOpenAILive user-triggered fetchYour page shown when a user asks about it right now
ClaudeBotAnthropicTraining + indexingPresence in Claude's knowledge and web answers
Google-ExtendedGoogleGemini training opt-out tokenBlocking it does not remove you from Google Search
PerplexityBotPerplexityAI search indexCited sources with outbound links in Perplexity
CCBotCommon CrawlOpen web archiveFeeds many research datasets — and many model trainers

Key nuance: one company can run several bots with different jobs. Blocking GPTBot (training) while allowing OAI-SearchBot (search) is exactly the kind of split policy edge-level controls make possible — robots.txt alone can express it, but only enforcement makes it real.

Rule of thumb

Sell attention? Charge the bots. Sell products? Let AI search in so assistants can recommend you. Ship developer docs? Allow broadly — being the answer an AI gives is your marketing. There is no universally right switch — only the right switch for how your site makes money.

Enjoy this tool? Build your own with Super