Only about 38% of URLs cited in Google AI Overviews rank in the top ten for the original query, down from 76% a year earlier, according to an Ahrefs analysis of 1.4 million ChatGPT prompts and 4 million AI Overview citations. In episode 4 of The GEO Show, Paris Childress works through that finding and seven other stories, including Google's new Preferred Sources button, free AI visibility tracking from Llumo, and Tollbit data showing European publisher sites absorb one human AI referral per 179 bot visits.
Ranking number one on Google used to be the whole game. Ahrefs analyzed 1.4 million ChatGPT prompts, attributed 88.46% of citations to the general search index, then looked at 863,000 Google result pages and 4 million AI Overview citations. Only about 38% of cited URLs ranked in the top ten for the original query, down from 76% a year earlier. The link between classic rankings and AI citations is coming apart, and query fan-out is why: the retrieval process behind an AI answer runs its own searches and reaches well past page one.
That is the anchor story in episode 4 of The GEO Show, and the other seven point the same way. The metrics the industry inherited from SEO, rank position, crawler volume, a single blended score, are measuring the wrong surface.
It matters less than it did. Only 38% of AI Overview citations rank in the classic top ten for the query that produced them, against 76% a year earlier. Paris is direct about it: page one is not the end-all, be-all it used to be in SEO.
What has not changed is the floor. You still need to be indexed, because the real-time search step behind an AI answer has to find you at all, and you still need to be competitively ranked. Paris puts the practical threshold at the first two or three pages rather than the first position. The work then shifts from position to coverage: if fan-out generates its own evidence questions, the page that gets cited is the one that answers them, at whatever rank it holds.
By platform, not as one number. Fractal ran 96 prompts 15 times each across ChatGPT 4o, Gemini 2.5 Flash and Claude Sonnet 4.6, producing 4,320 responses and more than 8,500 brand references. Only 11% of brands appeared across all three models. 77% appeared in only one.
Paris says the GEOforge data agrees, with dramatic differences in mention rate and citation rate between ChatGPT and Google's AI properties. The one strong relationship is inside Google: citations and sources correlate highly between Google AI Mode and AI Overviews. Add Claude and Perplexity to a blended average and it gets muddier still.
So really the right way and the best way to report on AI visibility and share of voice is at the platform level, and I really do believe this. You have to dig in and find the devil in the details rather than making any broad conclusions about a blended KPI.
No, and the gap is widening fast. Tollbit analyzed AI bots from 40 vendors across 3,906 publishers and found median scraping on European sites four times higher than on North American sites, with one human AI referral per 179 bot visits. The European scrape-to-referral ratio worsened from 150:1 in Q1 to 227:1 in Q2, close to twice as bad in one quarter.
If AI-referred human traffic is your primary GEO success metric, crawler volume is neither a proxy nor a predictor: bot volume is growing much faster than the visits it sends back. Paris has no explanation for why Europe gets crawled so much harder, offering privacy rules only as a wild guess.
Less reliably, and sometimes at your own expense. Search Engine Journal and Ahrefs tracked 9,886 AI answers across 34 self-promotional listicles published on five domains. AI systems frequently used the list as a source without recommending its publisher. In one experiment promoting Ahrefs' own conference, a competing event was recommended in 43% of the answers.
The tactic is familiar in SaaS: publish a best-ten-tools list, put your own brand at number one, collect the citations. Two things are eroding it:
Everything above the reporting layer. Llumo launched on August 20 with free AI visibility software on a bring-your-own-API-key model, tracking mentions, competitors, citations, prompt responses, query fan-out, sentiment and share of voice across ChatGPT, Gemini, Perplexity, Copilot, AI Mode and AI Overviews. Basic prompt and citation monitoring is now a commodity game.
Paris says GEOforge was built on that assumption from the start. SignalForge covers competitive AI visibility tracking, the part that is commoditizing. The value sits above it in ContentForge, which produces content grounded in a proprietary knowledge base, and SiteForge, which organizes citation building opportunities and lets users act on them inside the platform. The routes up the stack: content creation, content strategy, citation building, or analysis broad enough to cover SEO alongside GEO.
The through line across all eight stories is that GEO is outgrowing borrowed measurement. Rank, crawl volume and a single averaged score each describe something real, and none of them describes whether a brand is being cited.
Hi, and welcome back to The GEO Show. This is episode 4, and let's get right into the top stories for today. First, we have Google just created a user-controlled visibility signal inside AI search. Google has launched something called an embeddable Preferred Sources button on August 20th. Readers who select the site via this button can then find it more easily in Top Stories, AI Overviews, and AI Mode. More than 600,000 unique sources have already been selected. Publishers can promote the selection via the selection link, via email, or through social.
So this is a very interesting development. There is now a CTA button that allows any user to select a site as a preferred sources site, and I think this is a very good practice to put on your blogs, starting with that. And what that will do is create a preference for users to start to see your pages more in citations, in AI Overviews, and AI Mode. So to me, that is kind of a no-brainer. I think it can only help.
Moving along to the next story, a new AI entrant is attacking the price of GEO monitoring, and this entrant is called Llumo, and that's with a double L. Double L-U-M-O. They launched August 20th with free AI visibility software using a bring your own API key model. It tracks mentions, competitors, citations, prompt responses, query fan-out, sentiment, and share of voice across ChatGPT, Gemini, Perplexity, Copilot, AI Mode, and AI Overviews.
Yet another competitor has joined this already very crowded category, and they're offering a free product. So I think this was a matter of time before we started to see the reporting-only layer of this new tool category, AEO or GEO, AI visibility tools, let's call them. But now there's a free layer, so this is really becoming more and more commoditized by the day. Basic prompt and citation monitoring effectively is already a commodity game.
And then the goal is for the existing players of the category to go up the stack and to offer something more, to offer some differentiation. So this could be content creation and/or content strategy. It could have something to do with citation building, or it could be deeper analysis, broader analysis that encompasses SEO along with GEO. So I think this is a very, very fast-changing software category, and it's very interesting to watch.
One of the things that we decided with GEOforge from the very beginning was that we would have a competitive analytics AI visibility tracking module, which we called SignalForge. But that the real alpha, the value add above that would be ContentForge, which produces content that is grounded in a proprietary knowledge base, and also SiteForge, which allows users to organize citation building opportunities and take direct action on them directly inside of the platform. So differentiation is key in this now hyper-competitive category of AI search visibility tools.
Moving on to the next story. Creator marketing is officially entering GEO budgets. Digiday reports that Zoom and multiple agencies are experimenting with creator campaigns specifically to influence AI visibility. Agencies including Trevant and Crispin are auditing which creators and content receive large language model citations and then using that data in their creator selection.
So that is pretty fascinating because now GEO is starting to influence influencer marketing itself, and influencers are being selected in part based on criteria around their citations. So I imagine a new workflow being that you would engage in a trial with an influencer, and if that influencer's content, of course, you wanna look at its distribution, its reach, its conversions, and its ROI. But separately now, there are new judgment criteria around the degree of citation surface area that that influencer's content may have. So GEO budgets are now overlapping with influencer marketing, and GEO is now influencing influencer selection within influencer marketing.
Next story. Google AI Overviews are pulling social content now at meaningful scale. This is a study from BrightEdge, which analyzed more than 300 million monthly US searches and found Facebook cited 19.5 million times in AI Overviews, Instagram roughly 877,000 times, and TikTok 78,000 times. So interestingly, Facebook is overwhelmingly number one, about 10x or 20x larger than Instagram at number two, which is also 10x larger than TikTok at number three. Regardless, this analysis argues that exact answer relevance matters more than follower size.
So apparently what they're seeing here is that long-tail social media posts, not only from mega-influencers, but regular long-tail social media posts from users that don't have huge followings but that have very, very hyper-relevant posts are starting to appear now within AI Overviews at some significant scale. So what this is suggesting to me is that social posts are increasingly part of the retrieval layer in GEO, not just distribution.
So we've for a while thought about social media as a way to distribute repurposed content that initially gets created and optimized for SEO and GEO. So an example would be that you have an optimized blog post that gets repurposed, let's say, into a video, and then that video gets published on YouTube, Facebook, Instagram, TikTok, etc. And that is a distribution strategy. But now, as we see more of these social media assets appearing in AI Overviews, and these are not only YouTube, which of course dominates, but also now Facebook, Instagram, and TikTok, then it becomes also a retrieval layer in addition to a distribution game.
So what does that mean for video creators? LinkedIn video community content, this should contain concrete facts, explanations, benchmarks, and technical answers that can stand alone when they're retrieved by AI. So that actually suggests a slightly different approach in creating those videos because you think of them as not only distribution assets, but as citation sources that need to stand on their own, let's say, in a list of sources within an LLM response.
All right, moving on to the next story. This is about listicles, my favorite topic. Writing your own best tools article or listicle can actually help your competitor more than it helps you. And this is analysis coming from Search Engine Journal and Ahrefs. Ahrefs published 34 self-promotional listicles across five domains, and they tracked 9,886 AI answers. And what they saw was AI systems frequently used the list as a source without recommending its publisher. And in one experiment that promotes Ahrefs' own conference, a competing event was recommended in 43% of the answers.
So one of the most popular tactics now for citation building, and especially within SaaS and software, is the listicle. You create a best ten XYZ tools for fill in the blank, and you put yourself in that list. Often, you put your own brand number one in your own listicle. And these have tended to perform pretty well and get picked up well by LLMs and displayed well in LLM citation sources.
But what we're seeing is that this doesn't always show your brand in the best light because AI can take that same list and recommend your competitors over you, using your list as a source, but then not actually citing that source. So you really don't get any of the intended value there that you hoped for. And I have also seen recently that listicles as a share of total citations is now declining. So apparently, ChatGPT and others have started to see this as a quasi-manipulative way to just get more mentions of your brand on the web.
All right, moving on to the next story. AI visibility looks increasingly meaningless as a single blended KPI. A company called Fractal ran 96 prompts 15 times each across ChatGPT 4o, Gemini 2.5 Flash, and Claude Sonnet 4.6, generating 4,320 responses and 8,500 plus brand references. What they found was that only 11% of brands appeared across all three models. 77% appeared only in one.
And this is something that I have long suspected and we're steadily seeing more and more evidence, which is that a blended share of voice across models is really meaningless when your share of voice differs so drastically across the different models. And we now have enough evidence within GEOforge to see dramatic differences in your mention rate and your citation rate within, let's say, ChatGPT versus Google AI Overviews or Google AI Mode.
What we have seen is that there is a pretty high correlation of citations and sources between Google AI Mode and AI Overviews, but that we see very little correlation or similarity in brand visibility across ChatGPT and Google's AI properties. And when you add in Claude and Perplexity and other models, I think it gets even muddier. So really the right way and the best way to report on AI visibility and share of voice is at the platform level, and I really do believe this. You have to dig in and find the devil in the details rather than making any broad conclusions about a blended KPI or a blended AI visibility KPI.
All right, moving on to the next story. European sites may be paying a much higher AI scraping tax. Tollbit analyzed AI bots from 40 vendors across 3,906 publishers, and they found median scraping on European sites was four times higher than North American sites, with one human AI referral per 179 bot visits. So just let that sink in. One human AI referral for every 179 bot visits.
The European scrape-to-referral ratio worsened from 150 in Q1 of this year to 227 to 1 in Q2. So this ratio, it's scrape-to-referral ratio. It means how many pages does an AI bot scrape relative to how much referral traffic, human referral traffic that it delivers to that brand, to that website. So it got almost two times worse in one quarter from 150 to 1, meaning 150 scrapes per one AI-referred visit, to 227 to 1 in Q2. That's just, you know, phenomenal to see, and I think that is not gonna reverse anytime soon.
So what does this say? For one, that crawler volume alone is really a terrible GEO success metric because crawler volume is far surpassing and growing much faster than AI-referred human traffic. So if your primary GEO success metric is the traffic that you're getting referred from AI responses, LLM responses, don't look at crawler volume as a proxy or a predictor for that, because it's just growing much, much faster than those referred visits.
And that's particularly the case for European sites that had a four times higher crawl-to-visit ratio than the North American counterparts. I have no idea why Europe gets crawled so much more. It could have to do with privacy. That's just a wild guess on my part, though.
All right, let's move on to the last story of the day. New data reconciles that SEO still matters, with ranking number one isn't enough. So let's unpack this. Ahrefs analyzed 1.4 million ChatGPT prompts, and they attributed 88.46% of the citations to the general search index. So they took all the citations, and they wanted to see basically where they are ranking for equivalent searches.
What they found was that across 863,000 Google result pages and 4 million AI Overview citations, only about 38% of cited URLs ranked top ten for the original query, and that is down from 76% a year earlier. So let me try to explain that in simple terms. AI Overview citations only ranked in classic search results 38% of the time in the top ten. And a year ago, they ranked 76% of the time in the top ten results.
So this is a divergence of AI Overview citations and search queries. So that means that it no longer really matters to rank on the first page of Google if your goal is to maximize your citations within AI Overviews, because the query fan-out process can go much deeper than page one. You still definitely need to get indexed, and you need to be competitively ranked. I'd say you just probably should still be ranked on the first two or three pages.
But being ranked on page one is not the end-all, be-all like it used to be in SEO, because basically only 38% of those citations are ranked in the top ten for their original query. So keep those strong SEO fundamentals because you still need to get indexed so that the search retrieval process can find you when it goes out and does that real-time search. But also expand your topic coverage around the evidence questions that AI systems will fan out into.
All right, that's a wrap for today. We got through eight big stories, and we'll see you next time on episode five. Thanks for listening.
