Google's Nov 17 allow-list tells sites to skip CAPTCHAs and per-IP limits for Google-Agent, including on-site search, GraphQL, add-to-cart and booking, because robots.txt generally does not apply and many users share one egress pool. Blend already saw bot protection block its shopping agent at 29 of 200 UK/US retailers.
Why it matters: Pull CDN and WAF rules this week and whitelist Google-Agent through checkout APIs, or agent shoppers never reach the cart.
Task mode dropped the $19.99 paywall in the US, Japan, Korea, Canada, Australia and India, running on Hydrogen. In the US it can call businesses (it discloses it is an AI; businesses opt out in Google Business Profile), sign into sites with Password Manager, apply a discount code, and pay with Visa/Mastercard AP2 tokens; on some sites it still hands the last click back.
Why it matters: Local and retail teams need a GBP automated-call setting and a promo-code policy for agent checkout before this hits volume.
Google's new Search Journey Awareness help page: tCPA, tROAS and Max Conversions on Search (including AI Max) and PMax Search inventory automatically learn from secondary actions if they sit in a real goal category, not Default. Changing secondaries does not swing live bidding; changing the primary goal does.
Why it matters: Audit every secondary conversion category before Black Friday week; Default-bucketed add-to-carts are invisible to the model that is already bidding.
Ribar (ex-Optus EGM, JPMorgan): keep north of 80% of customers on the champion, because every extra live test pulls people off it and puts revenue at risk. Agents, simulation and live tests are one loop; "do nothing" is a valid next-best action, and a 12-month universal control is how you stop experiment spaghetti from contaminating the P&L.
Why it matters: Before the next lifecycle test calendar, put a universal holdout in place and run the long list of ideas in simulation, not on the paying base.