Fern's marketer keeps the brief, copy, images and weekly results as files inside the marketing site repo and lets Claude agents work the loop, which more than doubled qualified demos, took roughly 90% of brand and competitor search impression share, and lifted site traffic 202%. The agents never touch bidding, a media partner still owns the spend, so the leverage came from creative-strategy throughput rather than automated buying.
Why it matters: The transferable part is the file layout, not the tooling. Campaign state an operator owns, agents on the repetitive layer, a human owning the angle.
Four months after sponsored prompts became billable, Amazon says shoppers who click one convert 48% more often and spend 21% more on average, the first hard conversion data on any paid placement living inside a conversational shopping surface.
Why it matters: First real benchmark for what an agentic ad surface is worth, which is the number every budget conversation has been missing.
Productrise ran 100,000+ shopping queries through both surfaces: products appeared on 88% of standard searches versus 23% in AI Mode, 22.5 listings per page versus 4.3, and only 0.8% of products visible in standard search also appeared in AI Mode for the identical query. Winning the Shopping carousel tells you nothing about winning the conversation.
Why it matters: That 0.8% overlap is the cleanest proof yet that feed and product-data work is a separate discipline from AI-surface visibility.
IAB Tech Lab's AAMP 2.3 adds no new agent capability. It adds enterprise deployment paths (Bedrock AgentCore, Meta buying, Google Ad Manager reporting), privacy controls, and pricing guardrails on automated transactions. The bottleneck was never whether an agent could buy media, it was whether anyone would let it touch real money.
Why it matters: Standardized guardrails move the question from can an agent do this to who governs it, and that is where the operator's job now lives.
geoSurge ran 66 buyer prompts 60 times each across 3,960 responses and 13,281 fan-out searches: familiar brands got searched 55.7% of the time versus 17.4% for anything outside a model's top ten, and 63% of brand-specific searches went to one of five names. Only 31% of fan-out searches named a company at all, which is the gap a challenger can still walk through.
Why it matters: Puts a number on the incumbency tax in AI search, and points the challenger play at unbranded fan-out queries instead of brand terms.
Zellify let an optimizer build a unique journey per visitor over six weeks with a team of five, holding cost near $0.01 per visit via 97% token caching, moving conversion from 2.0% to 2.4% and compressing payback from 60 days to 50. The transferable lesson: a human rolled back a winning variant 10 hours before the automated guardrail would have, because it was gaining conversions while a retention proxy slid.
Why it matters: The strongest argument yet for a deterministic human checkpoint any time an agent's objective is narrower than the business it is optimizing.
Graph API v26.0 permanently blocks order management with no replacement API, drops Instagram Explore Feed and Messenger Stories placements, and removes dailyoutcomescurve, budgetguardrail and estimatedau. Most restrictions apply across every API version by October 27, so anyone running homegrown budget alerts is rebuilding them this quarter.
Why it matters: Anyone with custom Meta pacing automation has a dated break coming. Audit the integrations now, not in October.