Butler/Till and iHeart ran a four-week streaming-audio and podcast buy as a matched MCP pair: a planner wrote the brief, the buyer agent talked to the publisher agent, and about $10k cleared at 42% lower CPMs than the client's own direct benchmark. 48% of podcast impressions landed in non-skippable mid-roll, versus 33% on the traditional plan. Earlier tests went through SSP middlemen. This one did not.
Why it matters: The first clean receipt that agent-to-agent can skip the exchange on a channel that is still mostly not programmatic.
Query expansion, generated copy, and final-URL routing went live globally on August 19 as an opt-in Search toggle, with brand lists capped at 20 and per-feature experiments from day one. Exact match still beats AI Max in the Microsoft auction, unlike Google, and there is no forced conversion date.
Why it matters: Baseline Microsoft Search this week, before Copilot-query inventory starts mixing into the same campaigns.
VTA (1d) is a fixed one-day window, click wins ties, and the number is excluded from Conversions, CPA, bidding, billing, and conversion optimization. Docs updated August 18-19. No email, no blog post.
Why it matters: Treat VTA as a diagnostic column, not a KPI. Anyone using it to defend ChatGPT spend is ahead of what OpenAI will stand behind.
A 41-day test on 1,062 pages found GPTBot, ClaudeBot, Bingbot, PerplexityBot, and Meta and Amazon crawlers followed zero JS-injected links. Googlebot reached only 2% of those pages. The crawlers that feed AI answers never executed the JavaScript.
Why it matters: If a site's information architecture is client-rendered, AEO work on those URLs is theater.
Promptwatch put Reddit at 3.83% of ChatGPT Search citations from July 18 to August 7, then 0.52% from August 14 to 17. No matching one-day drop in Google AI Overviews. The cause is unconfirmed, and Promptwatch has not ruled out a collection issue.
Why it matters: Reddit-seeding as an AEO tactic just lost its ChatGPT premise overnight. Measure per model, not a blended AI visibility number.