Fluency's Eric Picard makes the operator case that probabilistic models belong at the intent layer and nowhere near execution: ask an LLM the same budget question five times and you get five answers. The execution layer has to be deterministic infrastructure with guardrails the model cannot route around.
Why it matters: The durable build is the guardrail layer, not the agent. Teams shipping agentic media buying should be able to name exactly which decisions the model is allowed to make.
DataBeat's read across 35 billion monthly impressions and 200+ bidders has agentic buyers in 86% fewer auctions than conventional DSPs, clearing at $6.13 CPM against $6.95 for everyone else, while challengers price guaranteed agentic products near 1% versus the roughly 20% incumbents take. Fixed-price pre-negotiated volume: television's model, rebuilt.
Why it matters: If agents move spend into negotiated deals, the auditability story changes shape. Signed containers replace an inspectable supply path, and buyers lose a lever they have leaned on for a decade.
NIQ takes Optiq Bridge out of beta in early September, distributing panel and retail measurement data over MCP, the same interface Google, Meta and Amazon already expose, with Publicis-Carrefour JV Unlimitail as first Activate retail media client. Pricing is still undecided, and pricing decides which agencies get access.
Why it matters: Share of prompt is arriving as a vendor-measured line item later this year. Worth settling your own AI-visibility reporting vocabulary before a measurement vendor defines it for you.
Agencies are selling Reddit comments as a lever on what ChatGPT and Gemini recommend, while mod rules vary wildly: some ban brand steering outside paid ads entirely, others allow brands only in designated threads. Reddit's own position is that nothing guarantees a comment reaches an LLM at all.
Why it matters: Real citation upside sitting next to real ban risk. Decide your stance in writing before a stakeholder asks you to run it.