New research found that measuring ChatGPT through the API recovered only 26.2% of the exact pages the ChatGPT app actually cited, and 48.7% of the source sites. Paris Childress covers the week's GEO news: parallel agents in GPT-6 Pro, a six-engine citation index, reported work at Anthropic on AI agent negotiations, a directory of 134 GEO tools, 15 AI visibility indicators for hotels, a Cloudflare and OpenAI discovery pilot, and a trusted agent transactions demo from GoDaddy, Infoblox, Google and Cloudflare.
New research suggests the ChatGPT that most AI visibility tools track is not the one your customers use. Paris Childress, founder of Hop AI and co-founder of GEOforge, covers this week's GEO news and what it means for brands that want to be found in AI answers.
Researchers studied ChatGPT's actual web search behavior with 80 questions, 40 in English and 40 in Chinese, across several model configurations. For GPT-5.5 Instant, combining Google Search results with OpenAI's Responses API recovered only 26.2% of the exact pages and 48.7% of the source websites that ChatGPT cited. The study did not test commercial trackers. Paris's takeaway: most tools measure ChatGPT visibility through the API, so their numbers may differ materially from the consumer app.
OpenAI updated GPT-6 Pro so multiple agents can work on different parts of a task at the same time, with better tool and plugin use. As research and buying tasks get delegated to several agents, prompt-based measurement may not fully capture how brands are discovered and compared.
The Machine Relations Index reports 31,721 cited domains across 27,227 answer runs and 200,158 source events from six AI engines, collected between May 10 and October 18. Reddit, YouTube, LinkedIn and Forbes lead. The data comes from the index operator and depends on its sampling method. Paris expects tool providers to offer source intelligence by industry, such as B2B SaaS or cybersecurity.
The Financial Times reports that Anthropic is consulting legal specialists as it explores how AI agents might negotiate agreements and commercial terms. This is reported, exploratory work, not confirmation of a launched negotiation service. Paris sees agents moving from recommending vendors to negotiating with them, which would bring clearer attribution.
AiSightKit lists 134 tools and platforms, including 45 AI visibility tools, 39 content improvement products and 29 impact measurement products. G2's AI search software category already has more than 650 listings. With vibe coding there isn't much of a barrier to entry, so leaders will need to stand out on marketing.
The AI Hospitality Alliance released an implementation and measurement guide with 15 indicators that separate website accessibility, AI visibility, recommendation accuracy and booking capability. Paris expects vertical-specific standards to appear in other industries.
A Cloudflare and OpenAI discovery pilot using about 1% of Cloudflare's network helped OpenAI discover and index hundreds of millions of previously unknown pages. It does not demonstrate that those pages were later cited or recommended. GoDaddy, Infoblox, Google and Cloudflare also demonstrated a prototype that connects agent discovery, identity verification, authorization and transaction execution. Paris's view: that full path will be attributable, so GEO teams will soon be able to tie their work to real revenue.
Paris: Hi, everybody. Welcome to episode 65 of The GEO Show, brought to you by GEOforge: Full Self-Driving for AI Visibility. Today, we'll start with ChatGPT. Research finds major differences between ChatGPT citations and API results. Researchers examined ChatGPT's actual web search behavior using 80 questions, 40 in English and 40 in Chinese, with repeated observations across several model configurations.
Paris: For GPT-5.5 Instant, combining Google Search results with OpenAI's Responses API recovered 26.2% of the exact pages ChatGPT cited and 48.7% of the source websites ChatGPT cited. So, the takeaway here is that an API-based measurement system may produce materially different citation results from the consumer-facing ChatGPT application, and most tools that are out there are in fact measuring ChatGPT visibility on an API-based measurement system.
Paris: So, make sure that you're aware of that if you're using one of those tools. Next up, OpenAI introduces parallel agents in GPT-6 Pro. OpenAI updated GPT-6 Pro to support multiple agents working on different portions of a task simultaneously. The release also improves tool and plugin use. Users can follow individual agents' progress and inspect completed work.
Paris: Availability is rolling out to eligible Pro subscribers. So now, complex research and commercial tasks may increasingly involve several AI agents rather than a single assistant answering one prompt. So we need to think now about how parallel research agents discover, evaluate, and compare competing brands, and existing prompt-based measurement may not fully capture these delegated workflows. Next up, six-engine citation index expands to 31,721 domains.
Paris: The latest Machine Relations Index reports 31,721 cited domains across 27,227 observed answer runs and 200,158 source events. The dataset covers six AI engines, with observations collected between May 10 and October 18.
Paris: Its leading cited domains include Reddit, YouTube, LinkedIn, and Forbes. Those are the usual suspects when it comes to citations. The data is produced by the organization operating the index. Its citation rates depend on its sampling methodology and should not be interpreted as universal platform behavior.
Paris: So this is an interesting example of multi-engine source intelligence for a particular category, and that practice is now becoming more recognizable as a research and product category. So I think that what we will see from some of the tools providers are these types of comparable source intelligence specifically for industries, let's say like B2B SaaS or cybersecurity, because generic cross-category citation patterns are really not that useful when it varies so much industry by industry.
Paris: This next story is very interesting from Anthropic. Anthropic explores legal frameworks for AI agent negotiations. The Financial Times reports that Anthropic is consulting legal specialists as it explores how AI agents might negotiate agreements and commercial terms. This is reported exploratory work, not confirmation that Anthropic has launched a commercial negotiation service.
Paris: But I think this is an inevitable next step. AI agents can and probably will evolve from just recommending vendors to actually negotiating with them, and this would introduce a new competitive environment that involves pricing, terms, customer eligibility, and automated decision-making.
Paris: So where we are right now is really looking at just simply visibility and being recommended by an agent, but it will move probably more even towards following that agent-mediated negotiation, perhaps all the way through to the conversion or a sale. And once that happens, I think that we will move more towards a much clearer attribution measurement in the GEO space.
Paris: Next up, updated GEO directory tracks 134 tools and platforms. AiSightKit currently archives 134 tools and platforms across AI visibility and related marketing workflows. Its classifications include 45 AI visibility tools, 39 content improvement products, and 29 impact measurement products.
Paris: So I think this is still just a small share of the entire space of these tools. We reported recently that there are over 650 listings in G2's AI search software category, so this just appears to be a fraction of that. But any way you look at it, this category is absolutely exploding.
Paris: There isn't really much of a barrier to entry for producing an AI visibility tool through vibe coding, and I think it's going to get bigger and bigger, and the leaders are going to have to rely more and more on marketing to stand out. Next story. Hospitality industry introduces 15 AI visibility indicators. The AI Hospitality Alliance released an implementation and measurement guide for hotel AI visibility.
Paris: It introduces 15 indicators and recommends establishing an approved collection of verified property facts before testing AI answers. The methodology distinguishes website accessibility, AI visibility, recommendation accuracy, and booking capability. Importantly, it does not assume that technical improvements automatically generate recommendations.
Paris: So AI visibility here is beginning to develop vertical-specific operating standards, and we may start to see these vertical operating standards start to pop up in other industries as well. And what is interesting here in the hospitality industry is that this methodology includes not only visibility and recommendation accuracy, but also booking capability.
Paris: All right, this next story comes from Cloudflare. Cloudflare and OpenAI discover hundreds of millions of previously unknown pages. Cloudflare and OpenAI describe a discovery pilot using approximately 1% of Cloudflare's network. According to the companies, the pilot enabled OpenAI to discover and index hundreds of millions of previously unknown web pages.
Paris: The initiative uses information from participating network infrastructure to improve discovery. It does not demonstrate that the discovered pages subsequently received citations or recommendations. So AI discovery here is evolving beyond conventional crawling of previously known URLs.
Paris: I think that more and more proprietary datasets are going to be popping up that are going to allow direct access to OpenAI's crawlers.
Paris: Okay, on to our last story of the day. GoDaddy, Google, and Cloudflare demonstrate trusted agent transactions. GoDaddy, Infoblox, Google, and Cloudflare demonstrated an architecture allowing AI agents to complete transactions across independent online systems. The demonstration connects agent discovery, identity verification, authorization, and transaction execution.
Paris: The work represents a technical prototype, not evidence of widespread commercial adoption, at least not yet. So I think the writing is on the wall here. As autonomous commerce develops, identity and authorization are going to become just as important as product discoverability. And of course, the final step is transaction execution.
Paris: And I think this full path will be highly attributable, and we're going to be moving, I think, soon well beyond just visibility of a brand inside of an LLM response. But especially with e-commerce, when there's a transaction at the end of it, we're going to have verified identity, authorization of that agent to transact on the user's behalf all the way through to a transaction.
Paris: So I do think we will be able to actually attribute real revenue from our GEO efforts probably in the near future. Okay. That does it for episode 65. Thanks for tuning in, and see you all on the next one.
