Cloudflare 與 OpenAI 聯手啟動研究計畫：用即時網路訊號讓 AI 搜尋更聰明

On July 8, 2026, Cloudflare and OpenAI announced a first-of-its-kind research pilot to explore how real-time signals from Cloudflare's global network can improve AI search indexing and answer quality . The pilot is the latest piece of a broader Cloudflare strategy, announced on July 1, 2026, that reclassifies AI crawlers into three behavior-based categories, sets new default blocks on mixed-use crawlers starting September 15, 2026, and experiments with a Pay Per Use compensation model for publishers ^TBB.

The Cloudflare–OpenAI Research Pilot

What it is

Cloudflare and OpenAI are launching a joint research pilot that combines Cloudflare's edge-network data with OpenAI's frontier search capabilities . Cloudflare CEO Matthew Prince said the goal is to use "fine-grained network signals to find better ways to improve AI search efficiency, helping users get high-quality answers faster" .

How it works (real-time network signals)

The pilot uses anonymized, aggregate signals from Cloudflare's network, which carries more than 20% of global web traffic . The specific signals include:

• Content freshness – how recently a page was updated or created.
• Traffic quality – patterns that indicate whether a page is actively used and trusted.
• Actual page-change data – whether and how often content actually changes on the server, rather than relying on sitemap timestamps.

These signals help an AI search engine decide what to crawl, when to recrawl, and which content is likely stale or low-quality, without having to fetch every page speculatively . The result is more efficient indexing and more timely, accurate answers.

Broader strategy: the July 1, 2026 policy overhaul

The OpenAI pilot fits inside a much larger set of policy and product changes that Cloudflare announced on July 1, 2026 — which it called "Content Independence Day" ^TBB.

New three-category AI crawler taxonomy

Cloudflare moved away from a simple "AI bot or not" classification and now requires that all crawlers be classified into one or more of three use cases ^TB:

Category	Behavior	Example
Search	Indexing content to answer queries later. Site owners should get referral traffic or compensation.	Standard search-engine crawlers
Agent	Real-time automated actions on behalf of a user (chat fetch, browser-use agents).	ChatGPT-User, Claude, Gemini agents
Training	Crawling to train or fine-tune a model. Data is absorbed into the model permanently.	Bulk training crawlers

Cloudflare "strongly encourages" companies that do all three to run separate, identifiable crawlers for each purpose ^T.

Upcoming default settings (September 15, 2026)

Starting September 15, 2026, Cloudflare will enforce new default settings for all new customers, new sites from existing customers, and free-tier accounts ^TBB:

• Training crawlers → Blocked by default on ad-monetized pages.
• Agent crawlers → Blocked by default on ad-monetized pages.
• Mixed-use crawlers (bots that combine search, training, and/or agent functions in one pass) → Blocked by default on any page that hosts ads ^TB.
• Search-only crawlers → Allowed by default, so sites remain discoverable in traditional search.

The sharpest change targets mixed-use crawlers: these bots collect content for search and AI training in a single pass, making it impossible for site owners to tell which purpose the data serves. Cloudflare's default block forces AI companies to either split their crawlers or lose access to ad-supported pages ^TB.

Pay Per Use: experiments with publisher compensation

On July 1, 2026, Cloudflare also announced it is shifting from its earlier Pay Per Crawl model (announced in 2025) toward a Pay Per Use model ^BB:

• Pay Per Crawl (private beta in 2025): charged AI companies each time a crawler requested content, using HTTP status codes and authentication ^B.
• Pay Per Use (announced July 2026): publishers get paid when their content actually powers an AI answer, not just when it is crawled ^B. This is described as a "metered utility" for AI search — the AI company pays for the value it delivers, not the raw access ^B.

The Pay Per Use model is still described as an "early experiment" and a "research program" ^B. Its exact pricing, revenue splits, and rollout timeline for general availability have not been published.

How the pieces fit together

The OpenAI research pilot, the three-category taxonomy, the default blocks on mixed-use crawlers, and the Pay Per Use experiments are all part of one coherent strategy ^TBB:

1. Force transparency – by requiring separate crawlers for search vs. training vs. agents, site owners can finally see what each bot is doing.
2. Give site owners control – defaults block the most extractive behaviors (training, agents, mixed-use) while preserving discoverability (search allowed).
3. Create a compensation layer – Pay Per Use replaces the broken "crawl for free, get referrals" bargain with payment tied to actual AI answer generation.
4. Make AI search smarter – the OpenAI pilot uses Cloudflare's real-time signals to reduce wasteful crawling and improve the timeliness and accuracy of AI answers, which benefits both publishers (better attribution) and AI companies (lower infrastructure cost).