Agent Beat for September 16, 2026

Welcome to Agent Beat, your daily briefing on how AI finds brands and how agents choose what to buy. Today is Wednesday, September 16, 2026. Here is what happened today.

In funding news, Profound, a startup that sells software tracking how brands appear and get cited inside AI answers, raised a $180 million Series D at a $1.8 billion valuation. That round came less than seven months after a $96 million Series C. The new round is much larger than the last one. Marketers buying AI-visibility tooling now have a much better-funded vendor to consider. They should also expect the category's products and sales coverage to keep changing at this pace. For buyers, that means new features and new sales conversations may arrive quickly.

For marketers preparing a site for AI agents, Vercel's Is Agentic reports can now be viewed through one of four site types: docs and content, business, app, or commerce. The commerce view highlights payment and checkout standards such as x402, UCP, and ACP. Scores do not change between views, so the switch is about what the report emphasizes, not a different grade. A site can set its default view with one meta tag. Marketers preparing a storefront for AI agents should run the audit in the commerce view to see which agent payment and checkout standards their pages are missing. That makes the commerce view the one to choose when the goal is agent-ready checkout.

Cloudflare added a Disallow AI Training option that blocks crawlers used for AI training while leaving mixed-use crawlers such as Googlebot free to index a site for search. Googlebot is an example of a mixed-use crawler. Google and Apple honor the training opt-outs, and Bing has not yet added robots.txt support, according to Search Engine Journal's report. The distinction matters for sites that want search visibility but not training use. Marketers who want their pages cited in AI answers should re-check what their robots.txt and Cloudflare rules actually block. Opting out of AI training no longer has to cost them search indexing.

At Dreamforce, Salesforce introduced Koa, its first CRM reasoning model, built with Nvidia to work through multistep sales, service, and customer workflows. The company says Koa decides which tools and actions a job needs. It does more than answer questions or write text. Marc Benioff, Salesforce's chair and CEO, said in a statement that the company's knowledge of how enterprise business works is put inside the model itself. Marketers should expect CRM vendors to pitch their own models for agent work. They should ask how a vendor model will use their account and product data.

Search Engine Journal published a guide on auditing a brand's presence in AI search, finding false claims and impersonators, and matching each problem to a correction, a report, or a preventive step. The guide treats AI answers as a place where a brand's identity can be altered or copied, as well as cited. Marketers should add branded-query checks in AI assistants to their regular monitoring, next to the reports they already run for ads and search results. Those checks can show where an assistant repeats a false claim or points to an impersonator.

Meta now lets AI agents handle the boring parts of WhatsApp Business setup. A new WhatsApp Business MCP server lets developers point AI coding agents such as Claude, Cursor, Codex, and ChatGPT at business setup, message templates, testing, and troubleshooting. Meta's business messaging tools are the target of the integration. Marketers who run WhatsApp Business accounts should expect agencies and platform vendors to configure templates and fix errors through agents instead of manual console work. The setup work that used to sit in a console can move into an agent workflow.

Marie Haynes turned her 15-year archive of Google algorithm updates into an MCP server at algo.mariehaynes.com. The post explains the architecture and how to connect it to Claude Desktop. MCP is the standard that lets AI assistants call outside tools and data. Marketers who hold a useful content archive can publish it the same way, so assistants query that data directly instead of waiting for a page to be crawled and cited. The archive becomes something an assistant can call directly, rather than only a page it might find.

Google is testing new forms of query expansion in AI Mode beyond standard follow-up questions. The tests include questions with multi-selection tools, a query refinement option, and paginated responses. Search Engine Roundtable reports these are tests, not a released feature. Marketers optimizing for AI answers should assume one prompt now fans out into several related questions. They should cover the sub-questions buyers actually ask instead of a single head term. That means writing for the follow-ups an AI assistant may generate from the first question.

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