AI citations in Google AI Mode, ChatGPT, and Perplexity shift from day to day, with a previously cited domain staying cited only 37.8% to 68.5% of the time depending on the platform. Episode 7 of The GEO Show covers that daily volatility alongside six other developments, including Google's position that AI Overviews need no special technical layer beyond SEO fundamentals, new data on how much of the web is now AI-authored, and the emergence of "agent experience" as a discipline downstream of GEO.
AI citations don't behave like search rankings. A domain cited by an LLM today can vanish from the same answer tomorrow, and the newest data on how often that happens should change how any team reports GEO progress. Episode 7 of The GEO Show also covers Google's stance on technical GEO, how much of the web is now AI-written, and why "agent experience" is becoming the next discipline teams need to track.
No. Google's John Mueller has confirmed there is nothing website owners need to do specifically for Google's AI Overviews or AI Mode beyond the same fundamentals that already drive strong SEO: crawlability, indexation, a logical information architecture, and clear entity signals. Google's AI results draw from its existing search index plus additional query fan-outs, which is why the position doesn't extend to ChatGPT, Claude, or Perplexity. Differentiation inside Google's own AI results comes from measurement, proprietary knowledge, and off-site authority, not from any hidden markup.
According to Pew Research Center's analysis of 490,000 English-language pages from Common Crawl's July 2026 sample, 10% of the entire sampled web shows meaningful signs of AI authorship or editing. Among pages published after ChatGPT's November 2022 launch, that figure rises to 35%. AI-generated content is no longer a differentiator; it's ambient infrastructure. The only defensible moat left is information gain: subject matter expert interviews, proprietary datasets, benchmarks, original technical analysis, customer evidence, and first-hand experience that a generic AI writer simply doesn't have access to.
Vendor research firm Chloro compared identical prompts across consecutive daily runs from August 1 through August 16 and found that citation persistence varies sharply by platform. A domain cited on one day remained cited the next day only 37.8% of the time in Google AI Mode, 49.9% of the time in ChatGPT, and 68.5% of the time in Perplexity.
"You actually need to run it, in our experience, probably in the order of thirty to forty times in a row to get a statistically relevant, high-confidence AI visibility observation out of that. And then that needs to be repeated."
That's the standard behind Hop AI's own SignalForge measurement, which runs each tracked prompt forty times every seven days to hold its margin of error to roughly plus or minus two percent. The takeaway for anyone reporting GEO results: never call a single lost or won citation a trend. Use repeated runs, rolling averages, and confidence bands, especially in front of executives.
Not according to Muck Rack's analysis of more than 25 million cited links across ChatGPT, Gemini, and Claude. Citation behavior varies materially by platform and by query type, and domain authority or domain rating doesn't factor into which sources get cited at all. What matters instead is relevance: a niche, highly relevant publication beats a high-authority generalist site. Industry trend queries cite journalism at more than twice the rate of how-to queries, and more than half of all journalism citations came from content published within the previous twelve months, so freshness compounds the advantage for reputable outlets.
RedMonk's new framework draws a clear line between two disciplines that are often conflated. GEO determines whether a product gets discovered or recommended by an AI system. Agent experience, or AX, determines whether an agent can actually use that product once it arrives: filling out forms, running searches, applying filters, or completing a transaction. Agents are already crawling websites at a far higher rate than human visitors, and tools like Grokbot now give every agent its own cloud-based computer that browses and acts like a person would. The practical shift: GEO gets you recommended, but AX decides whether AI can actually operate your site once it shows up.
The common thread across all seven stories: GEO keeps rewarding teams that measure carefully, build genuinely differentiated knowledge, and stay platform-specific rather than chasing a single blended score.
Hi, everybody, and welcome back to The GEO Show, episode seven. The GEO Show is brought to you by GEOforge, full self-driving for AI visibility. Now let's get into our stories for today. The first is that Google says there's no special technical GEO layer for AI Search. Google's John Mueller said there is nothing really special websites need to do specifically for generative responses in Google Search.
Google's AI results draw from its search index and additional query fan-outs. This undercuts a lot of vendors that are selling mysterious Google GEO hacks. It does not mean that AI visibility equals rankings, though. And of course, it does not apply to ChatGPT, Claude, or Perplexity, but only to Google's AI results, which is Google AI Overviews and Google AI Mode.
I'm not sure I fully agree with this, but I think fundamentally, Google's position so far has been that they don't want to give a lot of legitimacy to GEO as a discipline, and they've been saying that effectively, GEO is just good SEO. What we're talking about here specifically is the technical on-page factors.
I do think this is the area where SEO and GEO have the most overlap, which is the on-page technical foundation. You want your site to be well-structured. You want all the content to be very accessible with a very logical information architecture for all crawlers to come.
You do think about things like crawl efficiency and indexation, because if you still want to be retrieved in real-time search by LLMs, you do need to be indexed by the search engines. Crawlability and indexation are two fundamental things that are still there. But broadly speaking, what Google's John Mueller is doing here is trying to keep everyone focused on SEO fundamentals and not chasing some GEO hacks.
So really, the advice here would be positioning technical GEO for Google as an extension of strong SEO. Again, that means crawlability, information architecture, entity clarity, useful content, and evidence. Differentiation comes through measurement, proprietary knowledge, and off-site authority, but not any sort of secret markup.
All right, next story, coming from the Pew Research Center. Pew finds AI fingerprints on 10% of the web, and 35% of the newer dated pages since the launch of ChatGPT. Let's dive into these numbers. The Pew Research Center analyzed 490,000 English-language websites from Common Crawl. In its July 2026 sample, 10% of all of those pages showed meaningful signs of AI authorship or editing.
And among pages with a detectable publication date after ChatGPT's November 2022 launch, that figure was 35%. So what does that mean? More than a third of pages published since ChatGPT's launch are now showing up with these so-called AI fingerprints. That, to me, is not surprising, and I'd bet that if we looked specifically at the last six to twelve months, that figure is far north of fifty percent.
This is just the new reality we live in now. AI content itself is no longer differentiated in any way. It's actually the norm, and it's becoming ambient infrastructure, let's call it. So again, what is the moat for content creators when you're competing with everyone else who can produce an unlimited amount of AI-generated content?
The moat is information gain, which means things like subject matter expert interviews, proprietary datasets that only you own, benchmarks, original technical analysis, customer evidence, and first-hand experience. Information gain for the win.
Okay, next story. NVIDIA is partnering with Perplexity. Initially, there were rumors that NVIDIA was eyeing an investment in Perplexity, which would value Perplexity at thirty billion dollars. But now we have an update to the story, published today on Perplexity's own website: the partnership between NVIDIA and Perplexity has introduced something called Portable Computer for local-first AI.
Portable Computer runs Perplexity entirely on-device with NVIDIA, keeping private data local and escalating to the cloud only when a task needs it. This is a fascinating play. I don't know who NVIDIA is not partnering with in this ecosystem, but this is the first significant partnership between NVIDIA and Perplexity.
As you may know, NVIDIA has now produced its own AI computer, or laptop, with its own special chips to run AI locally on-device, and sometimes this is referred to as AI on the edge. I believe this is where Apple is going to be a major force in the very near future.
So what this is doing is that the Portable Computer is an NVIDIA-powered computer running a local model, so nothing ever goes to the cloud unless it's a task that needs it, and I don't know how they determine that. But this is going to be very big for companies that really value privacy and that aren't yet willing to send their proprietary data and their brand's alpha straight up into cloud-based systems hosted by the likes of OpenAI and Anthropic.
I think it's a fascinating play, and I would personally love to get my hands on Portable Computer. I do think it probably needs a better brand name, though, from a marketing perspective, but in any case, it's a significant move.
All right, let's go to the next story. The first public, searchable ChatGPT Ads library has launched. This is coming to us from LLMPulse. LLMPulse launched an independent ChatGPT Ads library on August 21st and updated it again on August 24th. It says the database currently contains over seventy-two thousand observed ads across nine countries. It's searchable by advertiser, by message, by country, and by date.
LLMPulse explicitly states that this is sampled, observational data, but not official OpenAI campaign data. Bottom line: competitive intelligence around conversational ads led by ChatGPT Ads is arriving extremely quickly, and I wouldn't be surprised if ChatGPT themselves launch their own official ads library for research purposes.
We've had this for many years with Google Ads and the Transparency Center. We have it as well with LinkedIn Ads, with Meta Ads, and I do think the industry has come to expect this as a sign, or a signal, of full transparency. It definitely assists with intelligence research and competitive research.
So I do think LLMPulse was the first out of the gate here, but I think there will be an official ChatGPT Ads library, sponsored by OpenAI, in the coming future. And of course, that kind of thing will spur even more investment in their ads when you can see the ads that are running and which ones are getting the most success.
All right, next story: daily AI citations are far noisier than conventional SEO rankings. This isn't really breaking news, but it further solidifies something we strongly suspect. The vendor research here comes from a company called Chloro, who compared identical prompts across consecutive daily runs from August 1st through the 16th.
A previously cited domain remained cited the next day only 37.8% of the time in Google AI Mode, 49.9% of the time in ChatGPT, and 68.5% of the time in Perplexity. That's actually interesting: first, that citations are won and lost effectively on a day-to-day basis. It means you're never really out of the game, no matter how small your brand is or how early you are into the category.
There's also a big difference across Google AI Mode, ChatGPT, and Perplexity. Perplexity has the highest citation stickiness, at 68.5%. Google AI Mode has the lowest stickiness, or maybe the highest slipperiness, at 37.8%. ChatGPT sits right in the middle, at about fifty percent. So saying you lost a citation today is basically just statistical noise.
It's not evidence at all that your optimization has failed. I think more than ever, when it comes to reporting on GEO, we need to take a long-term view and look at the trends from statistically relevant data. That means you can't just run a prompt once and say you were cited, call that a win, and expect to be there again tomorrow.
You actually need to run it, in our experience, probably in the order of thirty to forty times in a row to get a statistically relevant, high-confidence AI visibility observation out of that. And then that needs to be repeated. In our case with SignalForge, we repeat that every week, and all thirty of the prompts we track run on a cadence of forty runs every seven days, which gives us a very tight margin of error, at around plus or minus two percent.
So the takeaway here is: never report GEO from single executions. Always use repeated runs, rolling averages, and confidence bands, especially for executive dashboards. Otherwise, you're really just misrepresenting the data.
On to the next story: Muck Rack's twenty-five-million-link dataset reinforces that AI source strategy must be platform-specific. Muck Rack's August 24th analysis draws on more than twenty-five million cited links across ChatGPT, Gemini, and Claude. It says citation behavior varies materially by platform and by query type, and more than half of journalism citations came from content published within the previous twelve months.
Its underlying research found industry trend queries cite journalism at more than twice the rate of how-to queries. That's a big win for classic journalism. This adage of "get more PR" is really too crude. I think we have to look more deeply at the relevant publication, plus topic, plus model combination, which is a lot more nuanced, in fact.
This is a very interesting study, and as we've seen and discussed many times throughout several episodes of this show, there's enormous variability by platform, enormous variability by query type, and then, of course, journalism publications versus non-journalistic publications, which also carries a lot of weight.
One of my takeaways here is that if you can get cited in reputable journalism sources, this typically carries a much greater weighting, especially if it's published within the last twelve months. If it's fresh, that gives you an outsized advantage, but it's something that needs to be analyzed on a case-by-case basis. Publication, plus topic, plus model combination.
One last note here is about domain authority. We really don't look at domain authority or domain rating anymore at all as a signal, because AI doesn't care about the link profile. As far as we know, when AI chooses citations, it doesn't look at a domain-level authority score or a domain rating.
That's really something from the SEO days that hasn't carried over to GEO. It's all about relevancy. If you have a niche publication that's super relevant, it's better to get cited in that source rather than to go for something purely on a high domain rating or domain authority score alone.
All right, now let's wrap up with the last story of the day, and that's about agent experience, which is emerging as the discipline immediately downstream of GEO. RedMonk's August 24th framework distinguishes Generative Engine Optimization from Agent Experience, or AX.
GEO gets a product discovered or recommended, whereas AX determines whether an agent can actually use it. It frames GEO as upstream discovery and AX around access, context, tools, and orchestration. So we're moving from a world of "can AI recommend us" to a world of "can AI operate us." We're starting to see this and talk about this with some e-commerce clients.
I think this is a longer-term view here: agents are now coming and crawling en masse. They're crawling websites at a much higher rate now than human visitors, and that's been proven. But what comes next after the crawl? Agents are going to start using these websites, and we're seeing this now with tools like Grokbot, which I've been testing, where every single Grokbot agent has its own computer in the cloud, and it's using websites like humans.
Claude is doing this much faster with every passing month. The Chrome extension that allows Claude to operate Chrome as a user is working really well. ChatGPT's browser control is working very fast. So more and more, we're moving into this era of agent experience, and we need to think about websites not only through the lens of human user experience, but also agent experience.
The agent comes and crawls, and you have to make sure it can crawl efficiently and effectively. But also, if an agent is going to fill out forms, transact, or take specific actions on a website, like doing a search, creating filters, or generating a result, that experience for the agent also needs to be very smooth.
This flows logically downstream from GEO, because it flows downstream from an agent's crawl visit to discover the information as the first step. So that's something we'll be watching a lot in the future. That brings us to the end of this episode. Thank you all for tuning in, and we'll see you on the next one.
