Agent Beat for October 20, 2026

Welcome to Agent Beat, the daily briefing on how AI finds brands and how agents choose what to buy. It's Tuesday, October 20, 2026.

OpenAI pushed out a batch of ChatGPT Ads product updates late Friday. Search Engine Roundtable reports the batch includes new visual ad formats, a redesigned conversions page, performance emails, new measurement partners, and a bug-reporting and product feedback system. The conversions page and the measurement partners are the parts advertisers can use to see what a ChatGPT ad campaign actually produced. The other updates cover how the ads look, how performance is reported by email, and how advertisers can report bugs and send product feedback.

The question of what these systems can read has a new update from Glenn Gabe. He revisits his August 2025 tests, in which ChatGPT, Claude, and Perplexity struggled to read content rendered by JavaScript. Gemini and Copilot could read that content because they draw on Google's and Bing's crawling systems. The new post, 434 days later, checks whether the assistants with their own crawlers have caught up. Its title asks whether AI search platforms like ChatGPT and Claude are still unable to render JavaScript in October 2026. The comparison is between assistants that render pages themselves and those that draw on the two search crawlers.

The traffic side of that question gets a fresh look from Kevin Indig's Growth Memo. The memo takes up the industry belief that a meaningful share of AI traffic arrives as Direct in analytics tools. He cites his own AI Mode user behavior study, where 18% of exits took users directly to another site. He also cites Similarweb research finding that people who got an AI brand recommendation were more likely to visit those brands within seven days. Most of those visits came through Search. The memo looks at the Direct label in analytics and at what the user behavior study and the Similarweb research show about visits after AI recommendations.

On Google's own ad surface, Search Engine Roundtable reports a confirmation about AI Mode and Personal Intelligence. Google confirmed it will show ads in AI Mode results when Personal Intelligence is turned on. In March, Google said AI Mode with that setting enabled would be ad-free. The confirmation covers ad results in AI Mode when the setting is active. The earlier statement covered the same setting, and the new confirmation says ads will appear with it on.

Developer access is the subject of a separate report from Search Engine Journal. DeepSeek opened its in-house search engine to developers. Each result carries a page-level authority tier. DeepSeek says the API filters attempts to manipulate answers with GEO tactics. The search engine is DeepSeek's own, and the API is the developer access point for it. The authority tier is attached at the page level, and the filter is meant for attempts to manipulate answers with GEO tactics.

In Google's AI Mode, citations are getting a test. Google is testing a "Source preferences" button inside the AI Mode citations sidebar, according to Search Engine Roundtable. The button opens the user's preferred sources list, where sites can be added or removed. The test places the button in the citations sidebar for AI Mode results. The preferred sources list is the place where a user can add or remove sites.

A new paper looks at how a made-up concept can spread through AI search tools. Researchers published the concept on sites that AI search tools cite, and four English-language tools repeated it within a week, Search Engine Journal reports. Most of the other topics the team planted did not land. The paper tracked how the tools handled the planted concept and the other topics. Four English-language tools repeated the made-up concept within a week, while most of the other planted topics did not land.

Measurement is the weak spot in a new survey of search professionals. The survey of 194 search professionals, published by Abondance in partnership with Getfluence, found that 93% have AI visibility work on their agenda for their organizations. 49% plan to increase GEO budgets over the next 12 months. 60% perceive an impact on their business, but only 31% can put a figure on it. The survey covered AI visibility work and budgets, and the responses show a gap between perceived impact and measured impact.

That gap between being in the data and being recommended is the subject of Pedro Dias's new piece. He argues that a model can recite a brand's description and still fail the buyer's question, because sitting in the training data is not the same as being recommended. His piece, "It Knew Every Word," draws on research into training data to explain the gap. The argument separates a model's ability to repeat a description from a buyer's question and a recommendation. The research into training data is the basis for that explanation.

Search Console data is the subject of a study by Jan-Willem Bobbink, published on GSC Wizard and covered by Abondance. The study found raw replies typed to an AI showing up as queries on 94 of 155 sites analyzed. The replies include "yes," "ja," and "yes please." Those queries rose from 5.6 to 116 occurrences per million impressions between June 2025 and September 2026. They carry a 2.9% click rate against 36% for all queries. The study covers 155 sites, and the raw replies appeared as queries on 94 of them.

That's the briefing for Tuesday, October 20, 2026. Thank you for listening to Agent Beat. For the full written brief, go to agent-beat.com.