TLDR;
Cloudflare launched Redirects for AI Training to send eligible training crawlers from older pages to current canonical content. Brands can use it to supply better information for future training. It cannot remove outdated information that a model has already learned.
What happened
Cloudflare launched Redirects for AI Training, turning eligible canonical signals into redirects for verified training crawlers. The feature addresses outdated or duplicate material that remains accessible for historical reasons, applying to training crawlers rather than promising to rewrite existing model knowledge.
Why it matters
Enterprises often retain legacy pages for historical, support or regulatory reasons. Those pages may remain accessible even when a newer source should represent the current product or policy. Redirecting eligible verified training crawlers can influence future retrieval, but existing model knowledge and separate search indexes follow their own update processes.
Historical content can be both legitimately useful and misleading when treated as current. An old specification may help a customer support an older product, while a revised specification should guide someone considering today's version. Canonical signals and training redirects need to respect that distinction. Sending every historical page to the newest page could remove context rather than improve it. The enterprise should establish which materials are duplicates, which are superseded and which describe genuinely different versions. The feature's training-crawler scope is also essential. It can influence what an eligible crawler receives in a future request; it cannot promise that a model forgets an earlier claim. Answer accuracy therefore needs its own review process, separate from checking that a redirect was technically delivered.
How your brand can benefit / be affected
Audit canonical signals and the relationship between current and historical content before enabling redirects. Check that retained pages still serve legitimate human and support needs.
Test the behaviour for verified training crawlers and avoid assuming search or user-triggered fetches behave identically. Record which versions are supplied and monitor current answer accuracy separately. Keep clear dates and status labels on legacy material that must remain publicly available.
Inventory the legacy content that creates the greatest confusion and confirm the intended relationship with current material. Add clear dates, version labels and status explanations where a page must remain available to customers. Review canonical targets before relying on them for crawler redirects, especially when older products still have distinct support needs. A current product page is not automatically the right destination for every document bearing the same family name. Ask product and documentation owners to approve the mapping. This reduces the risk that a broad technical rule substitutes a different product's facts for the historical information being requested.
Test the supported behaviour for eligible verified training crawlers and record what content is returned. Keep separate checks for ordinary visitors and other retrieval routes instead of assuming they share the same response. Monitor observed answers for outdated claims, but do not interpret a correct redirect as proof that every assistant's knowledge has been updated. Where the source remains confusing, improve the visible explanation as well. Revisit the mapping when products or policies change so the current destination stays appropriate. The practical benefit is a more deliberate supply of authoritative information for future training requests, with honest limits around what the feature can change in existing models.
News date: 17 April 2026. Editorial review: 16 September 2026. Analysis includes subsequent developments where stated.