Agent Beat for September 28, 2026

Welcome to Agent Beat, your daily briefing on how AI finds brands and how agents choose what to buy. It's Monday, September 28, 2026. Here is what is happening today.

A survey from Initial, a Japanese research firm, found that 85% of B2B purchasing respondents use generative AI. Web担当者Forum reports that more than half choose the vendor they recalled first. The article also asks which business information source buyers used ahead of Nikkei and WBS. For a marketer selling to companies, first recall and the AI answers buyers read are where vendor shortlists start.

If AI answers shape recall, then citation patterns matter. Peec AI researcher David Konitzny posted data on LinkedIn showing that YouTube lost 91% of its share of ChatGPT citations between August and September 25, 2026. Facebook.com was down 88%, wikipedia.org was down 78%, and forbes.com was down 72%. Konitzny says the drop is not simple dilution, because cited volume also fell in absolute terms. ChatGPT can still retrieve a domain without citing it. Marketers should track retrieved sources separately from cited sources, since a domain can lose citation share while still feeding answers.

Citations point to pages, but how those pages are built for machines matters. A column argues that stripping a site down to markdown for machines removes the layer that tells a machine how the page is built. The machine can then only read about the content instead of using it. The visual layer can go while the structure should stay. Marketers serving text or markdown versions of their pages to crawlers and agents should check that headings, lists and other structure survive the conversion.

The surface where those pages appear is also changing. Rankshift compares tools for tracking brand visibility in Google's AI Mode. It describes AI Mode as a conversational tab that splits a question into sub-queries and pulls from a wider set of sources than AI Overviews. The post argues that AI Overviews tracking alone will not cover that surface. Marketers measuring AI visibility should check whether their tooling tracks AI Mode separately from AI Overviews.

This connects to a shift described in a field report. The article describes a move from a search model based on blue links to one driven by answer engines. It says that shift has already redistributed traffic and changed SEO and SEA practice. The field report covers 2025 and 2026. Marketers should treat answer engine visibility as an operational part of search work, not a future scenario.

The models behind those answer engines are changing too. OpenAI introduced GPT-6 Sol and Luna, two models it says bring frontier intelligence to everyday work at different balances of capability and cost. The post gives no prices and no benchmark results. Marketers who track which models sit behind AI assistants and agent products should wait for pricing and evaluation details before assuming anything about how their content gets cited.

Model changes also raise questions about spam detection. Abondance read the three-page SAFE paper, presented at IEEE TPS 2026 and signed by seven Google authors. It reports that the paper's diagrams use YouTube channel data and that the text never mentions websites, search ranking, or the August and September 2026 spam updates. The paper describes four AI agents that investigate coordinated networks publishing mass AI-generated content. It mentions "initial deployment results" without publishing figures. Marketers should not treat the paper as proof of a site-level AI-content filter: nothing in it says SAFE affects how web pages rank.

With all this movement, search traffic itself is under pressure. An article says organic traffic has been falling for months without a change in strategy. It frames a set of new Google rules as a redistribution of search. It is written for businesses that depend on search traffic. Marketers should audit how much of their organic traffic depends on Google and test how their pages perform in AI-driven results.

The response to that pressure may not require a separate program. Ahrefs published a guide to optimizing for AI search. The guide says the work is mostly the same as SEO: content still has to be crawlable, indexable and useful. Google says its AI search features rely on its core search systems. The post cites Microsoft making a similar point. Marketers should extend existing technical and content work into AI search rather than build a separate program for it.

For teams that want to hand off some of that work, Semrush published eight SEO workflows that hand defined tasks to an AI which pulls its own data. That includes a free Claude setup the company says it ran on a live site. The post is a how-to built on Semrush's own process, not a product launch. Marketers with repetitive keyword, competitor or reporting work can copy the setup, but should test the prompts on their own data before trusting the output.

That is today's Agent Beat. Thank you for listening. For the full written brief, go to agent-beat.com.