Agent Beat for September 26, 2026

Welcome to Agent Beat, the daily briefing on how AI finds brands and how agents choose what to buy. Today is Saturday, September 26, 2026. Here is what the day brought.

A study of AI and SEO in German e-commerce has a finding that changes where marketers should look. LEAP and SISTRIX analysed 2,496 of Germany's highest-revenue online shops, with data collected in June 2026. Organic visibility in that panel fell for the first time in four years. The study says generative search takes traffic from earlier stages of the buying journey rather than from product pages. So marketers should find out which of their pages AI answers now stand in front of. Product pages may keep their traffic while comparison and research pages do not.

The same shift is showing up on the agent side. Meta has opened its personal agent Muse to outside services. The company introduced Connectors, a mechanism that lets a third-party service be called mid-conversation when a user asks Muse for something. A service then no longer depends on the user searching for it. The article works through an analysis by researcher and podcast host Greg Isenberg, who compares the shift to Apple opening the App Store in 2008. He uses equipment rental as an example of a service pulled in partway through a planning conversation. Marketers should look for the points in a customer's request chain where their service could be the one the agent calls, instead of assuming the customer will search for them.

Advertising is changing alongside discovery. In Germany, the first ChatGPT Ads campaigns are producing mixed early results. OMR asked German marketers about their first campaigns and reports good click rates alongside disappointing tests. OMR also put questions to OpenAI. ChatGPT Ads have been sold in Germany since late August 2026. The piece covers what the format delivers for advertisers there. Marketers planning their own ChatGPT Ads tests can read these results as early benchmarks for click rates and conversion before they commit budget.

Google is also changing what it lets into the index. Google has deployed a new AI spam detector called SAFE. Search Engine Journal says the rollout may be part of Google's September 2026 spam update, and reports that the system is already live rather than planned. Marketers who publish machine-written pages at scale should expect those pages to be re-evaluated in Google's index. Pages that drop out of the index cannot be drawn into AI Overviews or AI Mode answers.

Google is not the only platform pushing its agent forward. Meta is putting its muscle behind Muse. TechCrunch reports that the personal agent is topping app store charts and adding users quickly, and that Meta is promoting it across its own apps and beyond. Amazon blocked Muse from buying on its marketplace this month, because the agent does not see Amazon's ads. Marketers should treat Meta's agent as a growing product discovery surface and check that their product data can be read and recommended inside it.

Tools for watching all of this are moving as well. CINC has updated the prompt generator in AIsearchmap, its GEO and LLMO tool for tracking how AI answer engines mention brands. The tool now builds monitoring prompts from a site's own information, suggesting axes such as main topic, comparison, target and area for users to adjust. Marketers using AIsearchmap should expect to set up answer-engine visibility tracking with less manual prompt design. Teams evaluating GEO tools get another feature to compare.

The way an answer is assembled is changing too. Google announced a new query fan-out framework called R4T-Diffusion. Search Engine Journal describes it as ready for production-scale search. Query fan-out is the step where one submitted question is split into several sub-queries before an answer is assembled. Marketers should watch how Google describes the rollout, because that split decides which pages are eligible to be cited in Google's AI answers.

Video markup is another place where the details matter. Google has updated its VideoObject structured data documentation to support creator and author properties and to clarify which interaction types interactionStatistic covers. The change affects the video markup Google reads from a page. Marketers who publish video can add the creator property and review their interactionStatistic values so their pages stay aligned with what Google documents.

Beyond the mechanics, the question of who gets chosen is getting its own data. In Japan, the category strategy firm suswork published category research data comparing how brands are chosen through AI recommendation with how they are chosen through human recall, using cafes and toothpaste as the examples. The firm says AI recommendation now matters alongside recall in everyday buying scenes. Marketers should track AI recommendation as a goal separate from brand recall, and check which products an assistant names in their own category.

And that question can be gamed. Research from MIT and Stanford points to a risk that AI agents will find the cheapest route to whatever target an SEO team sets, according to Search Engine Journal. The column says teams need metrics that agents cannot fake. The advice is to set AI visibility targets that require real demand, such as sales, rather than counts an agent can produce.

That is the briefing for Saturday, September 26, 2026. Thank you for listening to Agent Beat. The full written brief, with every story and its source, is at agent-beat.com. We will be back tomorrow.