Google Lighthouse 13.5.0 adds Agent Resource Discovery, putting agent crawl readiness next to llms.txt in mainstream web audits. Paris Childress covers Profound Dynamic Bot Rendering, citation fidelity near 50%, crawler spoofing, YouTube's 25.66% share of GEO citations, and why passing an audit is not the same as winning recommendations.
Paris Childress hosts episode 39 of The GEO Show on Google Lighthouse Agent Resource Discovery, Profound Dynamic Bot Rendering, citation fidelity near 50%, crawler spoofing, and YouTube's outsized role in GEO citations. The through-line: agent readiness is entering mainstream web audits, but passing a checklist is not the same as being cited or recommended.
The episode also covers weak country-code domain citation volume, llms.txt placeholders, Funky Media's AI Recommendation Gap, Swap Discovery inside Shopify, and Thailand's Recharge Land multi-run tracker.
I think this is the biggest story of the day and the most important one.
Lighthouse 13.5.0 (September 18, 2026) adds an Agent Resource Discovery audit and groups ARD with llms.txt under a new agent discovery category. Google expected the release to reach Chrome 156 DevTools and PageSpeed Insights within roughly two weeks. Paris ran getgeoforge.com through PageSpeed Insights live on air and treated the new section as a strong signal that agent crawl and discovery deserve the same attention as classic web vitals.
A citation-fidelity instrument covering 15,525 events found that only 54.8% of resolvable ChatGPT citations and 50.1% of resolvable Perplexity citations supported the claims they anchored. ChatGPT failures often looked like wrong-page or decorative attribution; Perplexity more often overstated what a plausible source actually said.
Yes. Machine Relations audited traffic claiming AI crawler identities. After retaining client IPs for a corrected one-day run, 73% matched vendor-published ranges and about 27% did not. User-agent matching alone can contaminate bot analytics with spoofed traffic, something Paris has also seen on getgeoforge.com in Cloudflare.
In the AI visibility / GEO category, editorial publications supplied only 5.9% of citation weight versus 11.3% across 16 categories. YouTube led at 25.66%. For marketers in this category, YouTube is currently one of the highest-leverage citation-building surfaces.
Watch the full episode on YouTube: https://youtu.be/nKsrvaP4oJk.
Paris: Hi, everybody. Welcome back to episode 39 of "The GEO Show," brought to you by GEOforge, full self-driving for AI visibility. Today is September 20th. It's a beautiful Sunday here in Sofia, Bulgaria, and as always, I am showing up to bring you everything that you need to know today about what's happening in the world of GEO.
Let's get started. We have a lot of news today about agents, agentic readiness, measurement, and recommendation versus visibility. So the first story of the day is about Profound. Profound moves from observing agents to serving them. Profound's September 18th release introduced dynamic bot rendering, which detects answer engine agents and can serve them a fully rendered version of a page without requiring changes to the customer's content management system.
Policies can vary by bot, page, page pattern, or domain, and the feature currently supports CloudFront, Cloudflare, and Vercel. Profound also added Google Search Console data alongside AI citations and new page-level AI visibility keyword metrics. All right. Profound is pushing into new territory. They're pushing into the retrieval and delivery layer, not just reporting.
And what they're actually doing here is something similar to what Cloudflare and a couple of other tools are doing, which is that they're giving websites the ability to serve AI bots a clean markdown file as opposed to those bots crawling whatever they find on that website, whether that be HTML or even more difficult would be JavaScript.
And I think more and more sites are actually gonna start to consider what used to be called, I don't know, 20 years ago in black hat SEO, double serving, which means that you serve one version of your website to humans, and that might be your JavaScript-heavy website that has really great and slick UI, and you serve a different version of the site, a really stripped down, .md file markdown language to AI bots so that they can really quickly parse it and understand it.
So Profound is jumping into this game as well, in a way now a little bit overlapping with tools like Cloudflare and others. Next story. Google Lighthouse makes agent discovery a first-class web audit. I think this is the biggest story of the day and the most important one. Google Lighthouse, which is also known as Google PageSpeed Insights, released version 13.5.0 on September 18th, which adds an agent resource discovery, ARD, audit, and they grouped this ARD with LLMS.txt under a new agent discovery category.
Google expects to release to to reach Chrome 156 DevTools and PageSpeed Insights within roughly two weeks. Well, I just tested it today, and of course, Google Lighthouse routes you directly to Google PageSpeed Insights. And what's on the screen here is what came up for GEOforge's website under the agent accessibility section, the new agentic browsing section, or in this case, it's agent resource discovery.
But what it's showing is that I passed one out of two criteria. The first is agent accessibility, and it looks like I have a problem with my accessibility tree, and that has to do with, the logo itself, so that's something I need to look into. And then down, at the bottom, I've passed an audit for cumulative layout shift, which is something that has been, a common part of Google Search Console, particularly for mobile UI, for quite some time.
But what's interesting here is that a really solid product from Google itself, which has been around for ages, Lighthouse and Google PageSpeed Insights, is now including an agent resource discovery section. So that is really a strong signal that we should start paying a lot of attention to how agents are able to crawl and discover our content.
Next story. Roughly half of resolvable AI citations actually supported the claim. A citation fidelity instrument covering 15,525 citation events found that 54.8% of the 1,255 resolvable GPT citations and 50.1% of the 1,298 resolvable Perplexity citations supported the claims they anchored.
Coverage differed materially. 77% of sampled ChatGPT citations were resolvable versus 47.6% for Perplexity. The researchers also report different failure patterns. ChatGPT failures skewed towards wrong page or decorative attribution, while Perplexity more often cited a plausible source but overstated what it supported. What's happening here? This is quite interesting. Some of the citations that are appearing are simply not supporting the claim that is made in the answer itself, and that is pretty disturbing. citation presence is an incomplete quality metric. the, URL can appear beside an answer without supplying meaningful evidence for the associated claim, so it's something we should not take for granted.
This is one of the first institutions' citation fidelity analysis. this is one of the first reports that actually is clicking into each of the citations and verifying whether or not the content of that page is supporting the claim properly, the claim that's made in the answer. And it turns out that a large percentage of the time, it is not, and it seems to be worse for Perplexity than it is for ChatGPT.
But still, it's about half. So that's something that we wanna look into, and I'm gonna want to follow this story for sure Next story. AI crawler dashboards can be materially wrong without identity verification Machine Relations audited 59,166 requests Across six web properties that claimed AI crawler identities Its original logging pipeline lacked client IPs, making genuine identity verification impossible. After retaining the IP for a corrected one-day run, 2,584 out of 3,540 requests, which is 73%, matched vendor published ranges. But 932, which is about 27%, did not verify. That's a problem. And 24 used an impossible Google extended HTTP identity.
The study explicitly says that the 73% figure applies only to that day and to those six properties. All right, what is happening here? User agent matching alone can contaminate AI bot analytics with spoofed or unrelated traffic. We have seen this firsthand with, AI bots crawling our own website, getgeoforge.com. We went into Cloudflare, and we saw clear evidence of spoofing, which means it's an AI agent that is pretending to be OpenAI's training agent, and it's named exactly the same way, but it does not have a traceable IP address, so it's fake, and it is there to do who-knows-what. Maybe to try to steal data or to do something it's not supposed to do.
So not only this is yet another reason why we need to be vigilant about tracking bot activity and even looking at what bots are there crawling our websites and try to identify potential fraudsters and spoofed bots. Next story. GEO buying answers rely unusually little on editorial sources. Machine Relations reports that editorial publications account for 5.9% of citation weight in the AI visibility/GEO category versus 11.3% across the 16 measured categories in its September 18th index.
So in its how to choose an AI visibility tool segment, no editorial publication appeared in more than 9.73% of observed runs. And most interestingly, YouTube led the overall source set at 25.66%. Very interesting. So this is our category, AI visibility and GEO. And because it is such a new category, the citations are coming: one out of four citations is coming from YouTube, and there's only a small percentage of citations coming from editorial sources And that's a sign of this ecosystem being still very immature. the AI systems are still assembling, buying answers from platforms, from small vendors, operator sites, communities, and long-tail sites rather than established trade media.
That, to me, is great news. In fact, that presents a great opportunity because it means that smaller sites where we might be trying to get brand visibility, brand mentions, these have a higher than usual chance of becoming citations. And most importantly, YouTube is twenty-five percent. Within this category, if you're marketing for AI visibility or GEO, YouTube by far is the highest leverage citation-building tactic that you can be pursuing.
I'm convinced of that. All right, next up. Country code domains barely register in AI citation volume. Reported by Machine Relations, a country domain citation study across 15,883 answer runs and 22,213 cited domains. National country code top-level domains represented just 6.58% of cited domains, but only 2.73% of citation volume. By contrast, .ai and .io together captured 8.6%, which is three times the citation volume of all national country code domains combined.
Only three of the 1,462 national ccTLDs cleared the study's publication evidence floor And all of them were .US. All right. So the country code domains are really not performing so far. a local domain suffix should not be treated as evidence that an AI engine is using genuinely local market sources. So if you're doing GEO in another country, like a European country or somewhere outside of the US right now, it still doesn't make a lot of sense to go after citations on those local country's, TLD domain websites thinking that that's gonna be, really, a local strategy, or that the AI is pulling from local market sources.
There just isn't evidence of that yet. All right, next story. More than half of LLMS.TXT adopters are shipping placeholders GEO Ready's audit of 282 self-selected domains found that 62.8% had an LLMS.txt file, but only 30.1% had a complete structured file. Of the 177 adopters, 92, or 52%, were classified as placeholders rather than useful site-specific manifests.
GEO Ready explicitly notes there is no confirmed causal link between LLMS.txt completeness and citation performance. Very interesting. A lot of people are claiming that LLMS.txt is kind of the robots.txt file for GEO, although I haven't yet seen clear evidence or clear correlation of domains that have LLMS.txt and their citation win rates, so to speak.
So I am not convinced that creating and publishing an LLMS.txt file will put you in a better position for crawling and ingestion by AI bots and therefore help you to win more citations. That evidence to me, I haven't seen it yet. I would love if someone could show that to me. But regardless, it's so easy to do, you might as well just do it.
I even did this for our website today. I went into Claude, I had it create an LLMS.txt file, and then I went into Webflow, and I uploaded it in the site settings in Webflow. It was super simple, and I see it kind of as an insurance policy. So technical GEO checklists are rapidly creating checkbox compliance, checkbox compliance.
So, just the fact that a file exists, like your LLMS.txt file existing, doesn't necessarily mean that it's done correctly and that it's structured properly or that it's useful at all. Next story. AI recommendation gap formalizes the distinction between visibility and preference. The Polish agency Funky Media introduced the AI recommendation gap, describing a five-stage progression: known, mentioned, cited, considered, and recommended.
That's a nice funnel. Its central claim is conceptual rather than independently validated research. A brand can be recognized and cited but disappear when users ask which product or provider they should choose. So an, interesting study coming out of Eastern Europe, one of our neighbors, Poland. they're really showing that the overall category language here is shifting more towards the actual buying decision.
Visibility and citations are increasingly being treated as intermediate metrics rather than the actual outcome, which is good. We're looking past being known, getting mentioned, being cited, and then towards being really considered and recommended by AI because that is the real goal, actually, is to have an AI agent say that your brand is the one that they should choose Next up, Swap Discovery bundles AI visibility and execution directly into Shopify.
Swap's current discovery product measures e-commerce brand visibility separately by AI assistant and country. It identifies competitors winning important buying prompts. It prioritizes challenges and writes prompts. I'm sorry, it writes product, collection, FAQ, and structured data improvements into Shopify. Swap positions Discovery alongside its agentic storefront and global commerce products rather than as a standalone GEO software. A really interesting move by this company, Swap Discovery. They have a very-- It looks like they have a very vertical integration with Shopify. They're going after the e-commerce segment, and they are allowing e-commerce brands to turn visibility intelligence directly into action.
So they're identifying which competitors are winning which prompts. They're prioritizing changes. They're able to write product descriptions accordingly. They can rewrite collection pages, FAQs, and structured data to close those AI visibility gaps against competitors directly inside of Shopify. this is why I think that there is already a very tight link or relationship between the GEO tool space and CMSs.
It's no surprise that the first major acquisition in this category was Scrunch AI getting bought for $225 million by, a major CMS provider. And here, Swap Discovery is making it easier for e-commerce sites to optimize for GEO without having to go into Shopify and directly make these changes inside of the CMS. This is something that we're trying to do with WordPress and Webflow and other major CMS providers as well. And I do think that more and more, the AI agentic layer is wedging itself between a user and, between a CMS user and the CMS itself. And more of our work... In fact, I'm already experiencing this in my work.
A lot of the changes that I make, to the website, creating new pages, updating pages, I do that directly with Claude, and Claude just goes right into Webflow and does it all for me. It's like magic. So great move here by Swap Discovery, vertically integrating as an agentic layer, on top of Shopify.
All right, the last story comes out of Thailand. Thailand gets a locally built multi-run GEO tracker. Recharge Land launched GEO Services, AI visibility services, and an internationally developed AI visibility tool on September 19th. The product tracks Google AI Overviews, ChatGPT, Gemini, and Microsoft's Copilot, and describes related multi-run testing plus cross-platform citation and competitor analysis.
All right. So GEO is getting regional. This is the first news, and brand that I have seen come out of Thailand, and it makes me think that because there is such differentiation through local language, local source ecosystems, market-specific prompt behavior, that there are real advantages to being first movers in a specific country market like Thailand or, gosh, any other country practically other than the US market and the English-speaking market.
It's natural that the opportunities start in the US and English, and they're gonna spread all throughout the world. So these guys are probably one of the first in Thailand. They're going to adapt and customize that product for the local, the local ecosystem in Thailand and probably will do very well. All right. That's all we've got for today. Thank you all for listening and see you on the next one.
