Cloudflare's AI block now stops everyone but Google, Apple and Microsoft

Cloudflare's customer email of 16 September 2026↗ said legacy Block AI Bots settings would migrate automatically over the following week, and that the switch would vanish from the dashboard once the migration was complete ¹. That week has run. What publishers clicked as a block is now, for the three companies that matter most, a polite note in a text file.

What a block became

The mapping is fixed. A legacy toggle lands on Search Allow, Training Disallow AI Training, and Agent blocked on pages with ads ¹². Disallow AI Training writes a no-training directive into robots.txt while Googlebot, Applebot and Bingbot keep fetching, because all three companies run one crawler for both search indexing and model training ¹². The gap it addresses is real: 17% of sites on Cloudflare's network block training in some form, fewer than 1% block search bots, and the network sits in front of more than a fifth of the web ¹³.

Enforcement now has two tiers

The setting still stops someone. Training crawlers run by OpenAI, Anthropic, Meta and Amazon are blocked outright at Cloudflare's edge ⁴. Google, Apple and Microsoft keep fetching under Cloudflare's Accountable label, and what happens to the page afterwards rests on the operator honoring the file ¹⁴. Microsoft cannot even read the new rule until its robots.txt support arrives, targeted for early 2027 ¹².

Accountable is Cloudflare's own designation, not an outside certification ⁴. The label was granted on capabilities each operator already has, paired with time-bound commitments for the rest ¹². And in August the fourth condition↗ required an operator to "show publicly" that disallowing training does not hurt search results; the September version asks for assurance ⁵¹.

The manual still promises the opposite

Cloudflare's own documentation, last updated 1 July 2026↗, still says mixed-purpose crawlers "will also be blocked by all configurations to block AI training, including the legacy Block AI bots option" ⁶. That page lists three options for each control: Allow, Block on pages with ads, and Block ⁶. The shipped product carries a fourth, Disallow AI Training, which the manual never mentions ¹. The behavior that shipped is the inverse of the page Cloudflare still publishes.

The only way back to a hard block is Block, which now cuts a site off from Google, Apple and Microsoft search altogether ¹². Blocking Bingbot takes a site out of Bing, Yahoo and Copilot at once ¹.

The receipt arrives in weeks, not now

URL-level reporting, the mechanism meant to reveal which disallowed pages were used anyway, is weeks away at Google and due next year at Apple ¹²⁴. Until it lands, the ledger reads one way: publishers keep their search traffic, the giants keep the corpus, Cloudflare keeps both sides as customers, and the only crawlers still stopped by force belong to the giants' AI rivals ⁴. Summaries are next, with central controls targeted for early 2027 ¹³.