NVIDIA is acquiring Hugging Face for $12.93 billion, a platform hosting 3 million models, 500,000 datasets, and 1 million apps used by over 18 million developers. On episode 21 of The GEO Show, recorded September 6, 2026 (Unification Day in Bulgaria), Paris Childress covers open-source fragmentation for per-platform GEO, ChatGPT-6 Astra model volatility, Pepper Agent Atlas and HubSpot GEO recommendations as execution layers, llms.txt adoption without evidence plus Ars Technica security risk, Brand Ghost's 263-day median citation age, USA Today's volume trap, ChatGPT ads splitting paid vs organic visibility, AI Overview mediation of high-intent B2B queries, and G2's finding that 51% of B2B buyers now start with AI chatbots.
NVIDIA is acquiring Hugging Face for $12.93 billion. Hugging Face hosts 3 million models, 500,000 datasets, and 1 million applications used by more than 18 million developers, and NVIDIA says the platform will remain open without requiring NVIDIA compute.
Paris Childress recorded episode 21 of The GEO Show solo on September 6, 2026, Unification Day in Bulgaria. He treats the deal as a signal that American open source will fragment AI engines beyond the frontier stack, then walks model volatility, execution-layer GEO tools, llms.txt evidence and security risk, citation freshness myths, publisher volume collapse, paid ChatGPT visibility, high-intent AI Overviews, and G2's buyer-behavior shift.
Because it strengthens open-source distribution and pushes brands to measure AI visibility per platform, not only against ChatGPT, Google AI Mode, Google AI Overviews, and Claude. Paris argues open-source models have been underweighted in optimization stacks, and this deal makes fragmentation harder to ignore.
I think that this move is going to really bolster open source models because now, NVIDIA effectively is going to be the biggest advocate now for American open source.
OpenAI's ChatGPT-6 Astra rollout includes browsing, research, and multi-step agentic work, with selected organizations first and broader availability soon. OpenAI also rates Astra at a critical cybersecurity capability threshold. Paris's point is that model releases can alter retrieval and recommendations with no brand-side change, so GEO reporting needs annotated release markers even when early behavior looks continuous.
Yes. Pepper launched Agent Atlas to connect GEO tracking, Search Console, analytics, competitive data, and CMS signals, then recommend and execute workflows such as content optimization. HubSpot's GEO Recommendations beta analyzes citation patterns across tracked prompts and recommends site content, third-party outreach, and social actions, including gap-driven content and mention tracking. Paris notes GEOforge already connects Search Console, Analytics, CMS, and competitor share of voice, and that HubSpot's CRM-plus-marketing stack is a strong moat for measurable revenue impact.
Ars Technica reported researchers scanned 6,214 corporate domains and found 120 llms.txt or llms-full.txt files referencing unregistered packages or domains. After registering abandoned names, coding agents including Claude, Codex, and Hermes executed proof-of-concept code, and at least one misconfigured reference already pointed toward real malware. Separately, Brand Ghost analyzed 6,762 AI citations and found a median cited age of 263 days, with 42.2% older than one year and only 21.5% under 90 days. Paris says protecting durable research, benchmarks, docs, and evergreen assets beats manufacturing superficial date updates.
USA Today told Search Engine Land that producing more content no longer solves audience growth as platforms keep users. OpenAI's ad tooling deepens custom audiences and conversion measurement, so earned GEO and paid ChatGPT visibility need separate scoreboards. SER Interactive data via Search Engine Land shows AI Overviews on 95.4% of X vs Y queries, 86.3% of review queries, 83.4% of price/cost/buy queries, and 81.3% of best-of queries. G2 finds 51% of B2B software buyers now start with AI chatbots versus Google, up from 29% eleven months earlier, with 69% selecting a different vendor than planned after chatbot recommendations.
00:28. NVIDIA acquires Hugging Face for $12.93B; open-source GEO fragmentation.
02:18. ChatGPT-6 Astra and model volatility for retrieval behavior.
05:07. Pepper Agent Atlas and HubSpot GEO Recommendations as execution layers.
09:24. Brand Ghost: median citation age 263 days.
17:03. G2: 51% of B2B buyers start with AI chatbots, up from 29%.
The throughline is platform fragmentation plus buyer behavior: open-source and paid surfaces multiply while B2B shortlists form inside AI chatbots first.
Paris: Hi, everybody. Welcome to episode 21, and for my Bulgarian audience that's listening, „Честит празник!" Today is Unification Day, September 6th, here in Bulgaria. So thank you for listening in. I'm your host, Paris Childress from GEOforge, full self-driving for AI visibility. All right, let's get right into today's top stories. Starting with, what's making headlines everywhere, and most of you have probably heard, NVIDIA is now officially acquiring Hugging Face for $12.93 billion.
Paris: NVIDIA agreed to acquire Hugging Face, whose platform hosts 3 million models, 500,000 datasets, and 1 million applications used by over 18 million developers. NVIDIA says Hugging Face will remain open and won't require NVIDIA compute. So what is the connection here to GEO?
Paris: I think that this move is going to really bolster open source models because now, NVIDIA effectively is going to be the biggest advocate now for American open source, and this is going to further, I believe, fragment the AI engines, beyond just the leading frontier models. So today, obviously, ChatGPT has the largest market share. That's probably the engine that most people are optimizing towards.
Paris: You've got, of course, Google's Gemini-powered AI Mode and Google AI Overviews and then Anthropic's Claude, and those are the main ones. But the open source models, no one is really paying that much attention. A lot of tools, do include optimization for some of the open source leading models like DeepSeek and, Kimi K3.
Paris: But I think this move is going to really, fragment this picture a lot more, and it's going to make it even more important to analyze and report on AI visibility per platform because it differs so much per platform. On to the next story here. Another big one, ChatGPT-6, codenamed Astra, raises the model volatility problem for GEO. OpenAI launched ChatGPT-6 Astra. It hasn't rolled out fully to everyone, but probably will in the next few days.
Paris: It includes major improvements in browsing, research, and multi-step agentic work. Rollout has begun with selected organizations. Broader availability is planned soon. And OpenAI also rates Astra as a critical cybersecurity capability threshold. All right. So far, what we're seeing is that major model changes can alter retrieval and recommendation behavior even when a brand doesn't change anything.
Paris: So I think it's important, even though we haven't really seen any major changes yet or nothing has been reported in the way that ChatGPT-6 Astra is, basically going through this retrieval process. For all we know, it's really the same general process and it hasn't really changed the rules fundamentally for GEO. but this is possible with every model release across every major lab. So I do think that it's very important to track and annotate these changes in your reporting.
Paris: Next up, a company called Pepper now has a new GEO agent that actually executes the work. Pepper launched Agent Atlas which connects GEO tracking, search console analytics, competitive data, and CMS information then recommends and executes workflows such as content optimization rather than simply reporting visibility. pretty interesting stuff here. the game here with Pepper is going quickly from just monitoring to diagnosing and then recommending and then executing.
Paris: And it's interesting here the way that they are connecting now Google Search Console, Google Analytics, competitive data, and CMS information. effectively, our platform GEOforge does all those things as well. one of the things that you do in setup is to connect a bunch of integration tools. That does include Google Search Console, Google Analytics, and your CMS.
Paris: And of course, we are benchmarking share of voice against up to 10 competitors, so the system is, constantly monitoring the activity of those competitors and taking that into account when it suggests topic recommendations. While Pepper here is doing something similar, I haven't dug in too deep, but it definitely looks like the execution layer is coming now into a lot of these platforms. Which brings us to another major platform, that's moving into the execution layer, HubSpot.
Paris: HubSpot's new GEO Recommendations beta analyzes citation patterns across tracked prompts, and it recommends actions spanning site content, third-party outreach, and social channels. It can generate content from identified gaps and subsequently track changes in mentions and citations. All right. No big surprise here because HubSpot is sitting on so much data that its customers have, starting with CRM data, but also its marketing automation platform has all social data connected.
Paris: So it's no surprise here that they want to leverage all that data for their GEO recommendations. And GEO for HubSpot is really now becoming embedded inside their main marketing platform, rather than just being a bolt-on specialist software. again, another recurring theme here is that the GEO competitive moat is shifting towards proprietary data, off-site authority acquisition, and measurable revenue impact. So I think the big advantage that HubSpot GEO has is that HubSpot is the CRM.
Paris: So not only can you track your visibility, and of course you can take actions through your, publishing to your CMS. A lot of people also have HubSpot CMS as well. So HubSpot can publish content for you. It can push content through your social channels, through third-party citation acquisitions, but most importantly, it can measure the downstream revenue impact, which is reported inside your HubSpot CRM.
Paris: So they have a pretty strong competitive edge here, primarily because of the aggregation of all the marketing stack and the CRM. Next up, 41% of B2B SaaS companies have llms.txt without evidence that it works. This is one of the most hotly debated topics in GEO right now. A company called Treyci, spelled T-R-E-Y-C-I, scanned 100 B2B SaaS companies and found that 41 of them had implemented llms.txt. However, no controlled comparison showed that the adopters received any more recommendations or citations.
Paris: This is being reported by NetContent SEO. So with all the talk about llms.txt, there still isn't evidence that it moves the needle. I think we should generally treat this as a cheap experiment rather than a GEO best practice, and always measure the before and after citation and recommendation rates rather than just selling implementation itself as an outcome. And the next story also has to do with llms.txt, which is becoming a cybersecurity issue reported by Ars Technica.
Paris: Researchers scanned 6,214 major corporate domains and found 120 llms.txt or llms-full.txt files referencing unregistered packages or domains. After researchers registered several abandoned names, AI coding agents including Claude, Codex, and Hermes executed their proof-of-concept code inside corporate environments. Ars Technica reports at least one misconfigured reference was already pointing toward real malware. Okay, so here, actually an attempt, a perhaps benign attempt to optimize for GEO with an llms.txt file has actually triggered, malware.
Paris: So keep in mind that machine-readable content can also be machine-executable trust infrastructure. All right, next story debunks a very popular myth that that AI cites only fresh data. So the story is that fresh data pushes back on the AI-only-cites-fresh-content narrative. Brand Ghost's latest run analyzed 6,762 AI citations. The median cited content age was 263 days, and 42.2% of citations were more than one year old. Only 21.5% were under 90 days old.
Paris: So most of the narrative we've been hearing in GEO is that there is a strong preference, a strong citation preference for freshness or fresh content that is less than 90 days old. this is one study that debunks that popular narrative and, basically saying that the freshness, refresh everything after 30 days is just too simplistic as a strategy.
Paris: And, what we need to do is really protect the durable research, benchmarks, product documentation, and authoritative evergreen assets rather than constantly just manufacturing superficial date updates. This is not a strategy. The next story comes to us from none other than USA Today. In fact, it is a story reported by Search Engine Land about USA Today. USA Today says producing more content no longer solves audience growth, and this was from September fourth. USA Today has reorganized its audience team amid layoffs.
Paris: An internal memo cited search pressure and said producing more content is no longer as effective, while platforms increasingly keep users rather than sending traffic outward. This is a harsh but a true reality, that ChatGPT, of course, is doing it, Google is doing it. The zero-click world is real, and USA Today is feeling it. The volume content model is structurally weakening, and it's not just merely experiencing an SEO downturn, but of course, it's really hitting major publishers like USA Today.
Paris: So content programs for brands that want to win in this space, they really need to optimize for information gain, which we talk about here a lot, citations, off-site citations, recommendations, branded demand, and pipeline, not sheer volume and, the articles per month KPI, I think is a bad idea. All right, next up. ChatGPT is becoming a real performance media stack. OpenAI's latest ad tooling adds much deeper custom audience controls, including customer and prospect uploads, inclusion, exclusion, and bid adjustments.
Paris: So Search Engine Land also reports here expanded conversion measurement, campaign creation, and international rollout. And this is from OpenAI's Help Center data. Conversational discovery is rapidly developing both an earned GEO layer and a paid acquisition layer. So just like in the SEO and PPC, SEO versus PPC days, it's now becoming very important to split organic recommendations and paid ChatGPT visibility separately. And right now, I haven't seen any platforms that are breaking out paid ChatGPT visibility specifically.
Paris: So it's very likely that share of voice data, AI visibility data that you're seeing now, particularly for ChatGPT, could strongly be influenced by paid ads. And I think the risk there is that you're comparing apples and oranges because we should be comparing organic non-paid visibility versus paid visibility for all the platforms.
Paris: And it's interesting to think how that might also affect Google's AI Mode when AI Mode becomes the default as well, because it will have ads, and will those ads, be treated separately within the major GEO platforms as paid visibility versus the organic visibility that is currently being reported now? Next story, also from Search Engine Land. Commercial B2B searches are far more AI-heavy than the intent labels suggest.
Paris: SER Interactive data cited by Search Engine Land shows Google AI Overviews appearing on 95.4% of X versus Y queries, 86.3% of review queries, 83.4% of price/cost/buy queries, and 81.3% of best of queries. Even though aggregate commercial query AI Overviews data pre-prevalence looks much lower.
Paris: Okay, so lots of statistics here, but the main takeaway is that B2B searches with high intent at the bottom of the funnel where people are saying, "Compare X versus Y," or, "I want to see reviews for these companies. I want pricing costs. I want the best of category," whatever queries. Here, you have a massive prevalence of AI Overviews.
Paris: So it was previously believed, and I think it's true, that AI Overviews started by taking away a lot of the clicks from the top of the funnel from educational, informational queries and prompts, and that is definitely true, and there's a lot of data to support that.
Paris: But now what we're seeing is that's moving down the funnel, and now we see also a massive overtaking of AI Overviews for high-intent queries, particularly for B2B searches. the intent categories are hiding the real threat here. the exact formats that B2B buyers use to create vendor shortlists are now heavily exposed to AI mediation. All right. What does this mean really for content? let's think about this.
Paris: A GEO program, we normally advise that the first types of pages that need to be created for a GEO program are not more blog posts at the top of the funnel, but rather bottom of the funnel pages like battle card pages, head-to-head comparison pages, alternatives pages. of course, pricing pages, making sure that those are consistent and accurate. best vendor types of pages, and even listicles, because they do still work.
Paris: So really focusing on refining and generating pages at the bottom of the funnel where users are close to making a decision or they're really in shortlisting mode, this can typically help you to get out of the gate much faster with GEO. All right. We do have one more bonus story here which builds on the last story, and that is coming from G2.
Paris: The latest G2 research has stated 51% of B2B software buyers now start their research with AI chatbots more than Google. So they've confirmed, G2, through a published study that headlines 51% of B2B software buyers now start research with AI chatbots more often than Google, up from 29% just 11 months earlier.
Paris: So think about that. it went from less than a year ago, 29% of B2B software buyers started with an AI chatbot before starting with Google, and today, less than a year later, it is more than half. It's 51% of B2B software buyers starting that research with AI chatbots. 69% of buyers selected a different vendor than originally planned based on an AI chatbot's recommendation.
Paris: Roughly a third bought from a previously unknown vendor after chatbot guidance. 85% view vendors more favorably when mentioned in AI recommendations. Wow. So just think of the massive impact on decisions now that AI chatbots have for B2B software buyers. It's massive, and this is very clearly telling me that if you're a SaaS company, B2B SaaS company, you definitely need to be GEO first. SEO is no longer the game because buyers are starting with AI chatbots.
Paris: They're much more susceptible to putting the brands that are recommended in those chats into their shortlists, even if they've never heard of those brands or if they've never, had any kind of previous favorability of those brands. Because AI chatbots are recommending them, they're making it right into the shortlists. They're having a major influence on B2B purchase decisions, not just early funnel awareness. Okay. So I think that will do it.
Paris: We've had a lot of stories today on this lovely Sunday, September 6th. Thank you all for tuning in, and we'll see you all in the next episode.
