Confirmed
Cloudflare splits AI-training blocks from search crawling so sites can stop training without losing Googlebot
Cloudflare announced a Disallow AI Training setting that publishes no-training preferences in robots.txt through the Google-Extended and Applebot-Extended tokens, letting a site block AI model training while still allowing Googlebot, Applebot and Bingbot to crawl for search. That differs from Cloudflare's existing Block option, which stops those same mixed-purpose crawlers outright because it blocks every behavior, training or search, at once. New defaults take effect September 15, 2026; Bing's equivalent token support is targeted for early 2027.
In plain termsRobots.txt is the file a site uses to tell automated visitors what they may fetch. Google, Apple and Bing's crawlers do double duty, feeding both search results and AI training, so a rule meant to stop the training half was also stopping search indexing by accident. This setting adds a line to that same file that targets the training half only.
Why it mattersSites that toggled a blanket AI-training block for unrelated reasons have a documented fix as of September 15, rather than the support-ticket workaround the same collision required a month earlier.
Our takeI reran a 15-origin probe across 14 large US retailers plus my own site in August, reaching 14 of those origins, and found 12 of them named zero AI crawlers in robots.txt with none carrying Cloudflare's Content Signals directive either, meaning most sites weren't making any training-versus-search distinction to begin with. This setting only helps sites that already manage that distinction inside Cloudflare; the bigger problem in my sample was sites with no signal set at all, which a toggle does nothing to fix unless someone turns it on.
Deep dive: Cloudflare's AI agent checklist, tested on 15 retailers
Confirmed
Google puts Gemini 3.8 Live inside Search Live for real-time voice troubleshooting
Google said on September 15, 2026 that Gemini 3.8 Live powers Search Live, the voice-and-video mode inside the Google Search app, giving it step-by-step, real-time troubleshooting help. The same rollout reaches the Gemini API, Google AI Studio, and Google Workspace's Docs, Gmail and Keep Live tools, plus a private preview in Gemini Enterprise. A companion model, Gemini 3.8 Live Extended Thinking, adds live progress narration for multi-step background tasks.
In plain termsSearch Live is the version of Google Search a person talks to out loud, using their phone's camera or microphone, and as of September 15, 2026 it runs on the same real-time voice model Google built for its other products, instead of an older one.
Why it mattersSearch Live becoming a live-model voice interface rather than a scripted assistant changes what counts as visibility for how-to and troubleshooting queries that used to land on a text results page.
Our takeFolding a general-purpose live model into Search Live blurs the line between Google's assistant and its search product, which matters for anyone tracking where how-to and troubleshooting queries go once a spoken answer replaces a results page. Google's own post does not say whether a Search Live session ever surfaces a clickable source link, and that omission is worth watching, since a voice answer with no visible URL removes the one signal a publisher uses to check whether it was the source.
Observed
Google tests sending shoppers straight to the merchant page from product grid results
Search Engine Roundtable reported on September 16, 2026 that Google is testing a change to free product grid listings in core Search: clicking a product opens the merchant's own product page directly instead of the multi-retailer overlay panel. The report is based on a discovery SEO consultant Brodie Clark shared on X, and notes Google piloted a similar direct-link approach inside the separate Google Shopping tab back in June 2026.
In plain termsClicking a product in Google's free shopping results often opens a pop-up panel that shows the same product at several stores. This test skips the pop-up and sends the shopper straight to that one store's own product page instead.
Why it mattersA direct-to-merchant click skips the overlay where a shopper previously compared several retailers carrying the same product, which changes both the traffic an eligible listing can expect and the visibility smaller retailers lost inside that overlay.
Our takeThis follows a pattern of Google testing changes to the layer that sits between a Shopping result and the merchant's own page, from AI-rewritten ad copy on the paid side to a routing change here on the free side. It is worth tracking once this test is visible in server logs, since Google Search Console's existing Shopping filters won't separate direct-from-Search sessions from Shopping-tab sessions without a merchant checking referrer data directly.
Confirmed
Snowflake ships per-user spending quotas for AI Functions and Cortex Agents
Snowflake announced general availability on September 15, 2026 of per-user quotas that cap individual spend on governed AI services, covering Cortex AI Functions such as AI_COMPLETE and AI_SUMMARIZE, Cortex Agents, CoWork, and warehouse compute. Block enforcement automatically pauses a single user's access to a specific service once they hit a daily or monthly limit, resuming on the next cycle, replacing blanket access controls with per-person guardrails.
In plain termsInstead of an administrator deciding upfront who may use the company's AI tools at all, this lets everyone have access while automatically cutting off any single person once their AI usage for the month gets too expensive.
Why it mattersA team running LLM classification or extraction jobs inside a client's Snowflake instance can be capped without losing access outright, which changes how that access gets negotiated with the client's data team.
Our takePer-user, per-service caps are a more surgical version of a control that usually ships as an all-or-nothing access grant. The detail worth watching is whether block enforcement fails open or closed for a client-facing agent mid-conversation, since a hard pause on Cortex Agents mid-task is a different failure mode than a pause on a batch AI_SUMMARIZE job that picks up again next cycle.