AirOps research finds local websites account for 15.5% of citations in European AI answers vs 0.95% in its US dataset, about 16x more, and that identical questions in English and local languages share only about 10% of cited sources. Paris Childress covers the week's GEO news under one theme, AI discovery moving from dashboards to infrastructure, standards and outcomes: ChatGPT Ads in 89 countries and at a $1 billion run rate, China's GEO industry standards, Perplexity's multilingual retrieval models, data void manipulation, Adobe Express agentic retrieval research, formatting loss in document ingestion, long-tail bias in generative recommenders, and Gemini in Search Ads 360.
Local sites get 15.5% of European AI citations vs 0.95% in the US, about 16x more, according to AirOps research. Paris Childress covers this week's GEO news, with one theme: AI discovery is moving from dashboards to infrastructure, standards and outcomes.
AirOps announced its expansion into Europe, the Middle East and Africa with a London team led by Cecile Meuguet, whom Paris interviewed at brightonSEO. Its proprietary research reports that local websites account for 15.5% of citations in European AI answers, compared with 0.95% in its US dataset, roughly 16x more. Identical questions asked in English and in local European languages shared only about 10% of their cited sources.
OpenAI expanded ChatGPT Ads to 26 more countries, 89 in total, including Argentina, Colombia, Nigeria, South Africa, Serbia, Ukraine, Bangladesh and Pakistan. More than 100,000 advertisers have run campaigns in about eight months. Ads only reach free and Go users, so Paris warns B2B marketers that this may not be their target audience.
OpenAI disclosed that ChatGPT ads reached a $1 billion annualized revenue run rate less than 200 days after launch, a run rate estimate rather than recognized revenue. A specialized model keeps conversation context, identifies commercial intent, retrieves candidates, evaluates relevance and runs auctions, separately from organic answers. Paris expects sponsored agents invited into conversations to eventually change attribution.
Three group standards coordinated by the China Advertising Association took effect, covering GEO service requirements, monitoring and verification of GEO results, and classification and selection of content sources, with Frost & Sullivan helping draft the last. They are not globally binding, but Paris notes China has adopted GEO as the name for the industry.
Perplexity researchers published pplx-embed-v2-late, a family of multilingual and multimodal retrieval models: a 9 billion parameter model for document indexing and a 0.6 billion parameter model for low latency querying, sharing one embedding space.
Researchers found that in controlled experiments with ChatGPT, Gemini and Sonar Pro, attacker-controlled content appeared in responses 31% of the time when credible information was scarce, and a single well-placed page could sometimes sway an answer. This does not mean 31% of ordinary answers are manipulated. Paris recommends publishing an entity map from entitymap.org.
Adobe Express researchers proposed LARA, the Layered Agentic Retrieval Assessment, which separates retrieval quality, engagement and outcomes. Users reformulated queries 44.4% of the time in agentic chat vs 33% in gallery search, and traditional retrieval metrics say less when an agent consumes results first.
Using 76 real policy documents, researchers found that converting documents to plain text can discard meaning carried by formatting. Required identification F1 fell from 0.690 to 0.150 with formatting-blind ingestion, with color and highlighting especially important. Paris says knowledge bases should preserve structure and serve structured data to AI crawlers.
Researchers found candidate generation can drop niche products before ranking and ranking can favor popular items. Paris says brand mention co-occurrence gives well-known brands a big head start in GEO.
Google's first monthly marketing platform update redesigns the Search Ads 360 home with Gemini-driven insights and recommendations, and adds AI-curated media opportunities in Display & Video 360. These are paid features, not organic citation changes.
Paris: Hi, everybody, and welcome to episode 69 of The GEO Show, brought to you by GEOforge: full self-driving for AI visibility. I'm your host, Paris Childress. Today, we're gonna be talking about how AI discovery is moving from dashboards to infrastructure, and from dashboards to infrastructure standards and outcomes, to be more precise.
Paris: Let's get into it. We'll start off with news from AirOps. AirOps expands into Europe with a London-based operation. AirOps announces its expansion into Europe, the Middle East, and Africa, establishing a London team led by Cecile Meuguet. And I met Cecile and interviewed her in person at brightonSEO a couple of weeks ago.
Paris: This announcement includes several findings from AirOps proprietary research. Most notably, AirOps reports that local websites account for 15.5% of citations in European AI answers, compared with just 0.95% in its US dataset. It also reports that identical questions asked in English and local European languages shared only approximately 10% of their cited sources.
Paris: So two things going on here. They're entering into Europe, and the system for basically fetching citations seems to be very, very different and not matching up with the US dataset hardly at all, with more prominence for local websites. That's very interesting. On to the next story. ChatGPT Ads is also expanding into 89 countries.
Paris: OpenAI announced that ChatGPT Ads is expanding into 26 additional countries, bringing availability to 89 countries worldwide. The company also reports that more than 100,000 advertisers have run campaigns since advertising launched approximately eight months ago. New markets include Argentina, Colombia, Nigeria, South Africa, Serbia, Ukraine, Bangladesh, and Pakistan.
Paris: And ads remain restricted to free and Go users. OpenAI says advertising does not influence ChatGPT's organic answers. Well, great news. I don't think Bulgaria is yet on the list here, but they are now at 89 countries and expanding very rapidly. I think their revenue now, their annual recurring revenue is above $1 billion in less than a year from launching.
Paris: So this new ad platform is really on a tear. And I think also what's interesting here with this expansion is that ads are only showing to free and Go users, but not to paid users. I would actually prefer to see some ads. I guess as a marketing leader, I wanna see what it looks like actually, but because I have a free plan, I'm not really able to see that.
Paris: So I would love to have an option to turn ads on in ChatGPT if possible. And there is a consideration here that if you're a B2B marketer, you should consider that running ChatGPT ads is only gonna be showing to free subscribers, well, non-paying subscribers, free users of ChatGPT, and this might not really be your target audience. Okay.
Paris: As I had just previously mentioned, OpenAI reveals a $1 billion ChatGPT advertising business. OpenAI disclosed that ChatGPT ads reached $1 billion in annualized revenue run rate less than 200 days after launch. Its engineering team also explained how the advertising system works. Unlike traditional search advertising, ChatGPT's advertising infrastructure maintains context as conversations develop.
Paris: A specialized model identifies commercial intent, retrieves advertising candidates, evaluates relevance, and runs auctions. Importantly, this advertising system operates separately from organic answer generation. The disclosed revenue figure is a run rate estimate, not recognized annual revenue. So very interesting and very innovative advertising system here, where you have, basically maintenance of context as the conversations develop.
Paris: So when ChatGPT detects that there is some commercial intent based on the conversation, it retrieves ad candidates and evaluates their relevance, and it runs the auctions. And I think that right now we're seeing display ads showing up when this commercial intent is triggering the ad auctions.
Paris: But I think in the future, these could be replaced by actual agents, conversational agents that are sponsored and brought in or invited into the conversation by the primary agent, and that could be across ChatGPT Dots and the classic ChatGPT experience. And I think when that happens, and when a lead is captured or a sale is made, within ChatGPT, ushered by one of these sponsored agents that gets invited into the conversation, that's gonna change entirely the attribution problem that we're now facing.
Paris: The next story comes out of China. Three GEO industry standards take effect in China. Three group standards coordinated by the China Advertising Association take effect today. They address, number one, requirements for generative engine optimization services. Two, monitoring and verification of GEO results.
Paris: And three, classification, grading, and selection of content sources. Frost & Sullivan participated in drafting the content source standard. These are Chinese industry group standards. They're not globally binding regulations. Their impact on commercial practices remains to be established.
Paris: But it's interesting now that China has recognized GEO as an industry and is starting to apply industry standards implementation. So GEO is beginning to develop formalized service and measurement frameworks all around the world. And standards around result verification could eventually influence enterprise procurement expectations.
Paris: And just as a side note, it seems that China has adopted GEO as the primary naming convention. So that's interesting. That's good news for us as well. The next story, Perplexity. Perplexity publishes new multilingual retrieval models. Perplexity researchers published details of pplx-embed-v2-late, which is a family of multilingual and multimodal retrieval models.
Paris: The release describes two models. First, a 9 billion parameter model for document indexing, and a 0.6 billion parameter model suitable for lower latency querying. Both models share an embedding space, enabling asymmetric retrieval architectures. The researchers report strong results across text, image, and visual document benchmarks.
Paris: All right. So Perplexity is still pushing hard on innovation here, despite their very, very low market share, and pushing into new multilingual retrieval models. So that retrieval infrastructure is improving its ability to process different document formats and languages, and the technical systems determining which evidence an AI assistant encounters will continue to evolve.
Paris: All right. This next story is looking pretty important. Academic research finds AI search is vulnerable to data void manipulation. Researchers investigated how malicious information can influence AI assistants when credible information about a subject is scarce. Across controlled experiments involving ChatGPT, Gemini, and Sonar Pro, attacker-controlled content appeared in responses 31% of the time.
Paris: The researchers found that a single strategically placed page could sometimes influence an answer. The study examines deliberately constructed experimental attacks. Its results do not establish that 31% of ordinary AI answers are manipulated. All right.
Paris: This is a bit troubling actually, because in cases where there's a void of credible information about a subject or about a brand, then there is a risk of attacker-controlled content being able to influence that answer. And in this controlled experiment, it looks like that the attacker-controlled content was pretty successful in getting into responses 31% of the time.
Paris: We've been doing some experiments recently with entity maps, and I think that this is becoming a really important thing, to submit, which is a one-page document that you can create at entitymap.org. And that document contains all of the important and critical and credible information about your brand and all of the entities that are associated with your brand.
Paris: And I would strongly suggest you all listening to check out entitymap.org and create this single-page entity map and upload it to your website according to those instructions. Next up, new research identifies an attribution gap in agentic retrieval. Researchers examining retrieval in Adobe Express propose a framework called Layered Agentic Retrieval Assessment, or LARA for short.
Paris: The study distinguishes retrieval quality, user engagement, and outcomes, and overall product success. In its observed environment, users reformulated queries in agentic chat 44.4% of the time, compared with 33% in gallery search. Reformatting their queries. Interesting. The researchers argue that traditional retrieval metrics become less informative when an AI agent consumes search results before presenting generated content to users.
Paris: And as we know, retrieving information, incorporating it into an answer, and influencing an eventual outcome, these are increasingly becoming separate events, and this retrieval process is now starting to behave differently with these AI agents. And I think that we're quickly heading into the next big phase where AI agents are really gonna dominate the usage of AI.
Paris: And it seems that users are reformulating their queries a lot more often when they're using AI agents as opposed to what is called a gallery search, or I guess that is just regular, just a regular LLM behavior. And that's gonna be really interesting to keep a close eye on. The next story is also a little bit troubling.
Paris: Enterprise document ingestion can erase important information. Researchers investigated whether AI agents correctly interpret information contained in enterprise documents. Using 76 real policy documents, they found that converting documents into plain representations could discard meaning communicated through formatting.
Paris: In one reported comparison, required identification F1 performance fell from 0.690 to 0.150 when using a formatting-blind ingestion approach. Color and highlighting were particularly consequential. These findings concern a specific document understanding task, and they do not establish equivalent performance losses across all retrieval-augmented generation systems.
Paris: All right. Accurate knowledge retrieval really begins before an AI system searches the resulting knowledge base, and document ingestion can introduce errors that later reasoning cannot reliably correct. So I think most of us are assuming that there's accurate understanding and ingestion of documents, and this is pointing to some troubling evidence that that is not necessarily the case, especially with formatting, like different colored and highlighted formats can really, perhaps jumble the AI agent's understanding and accurate interpretation of that information. I think this underscores why knowledge bases in general should really preserve structured information, and they should also serve structured data whenever possible to AI crawlers so that they can ensure the highest possible accuracy of ingestion and understanding of those documents.
Paris: Next story, generative recommendation models can disadvantage niche products. Researchers examined how generative recommendation models handle less popular products. They identified two potential disadvantages for long-tail items. First, candidate generation methods can eliminate niche products before ranking occurs.
Paris: And second, the ranking model can favor popular items even when less common alternatives are relevant. The researchers proposed a method to restore information about item co-occurrence, and reported improved long-tail recommendation accuracy in their experimental settings. This study concerns particular recommendation architectures, not directly observed behavior in ChatGPT shopping or in Google AI mode.
Paris: Okay, I think this is also kind of bad news if you are one of these niche products or niche brands. If you look at the number of brand mentions and the co-occurrence of the most popular brands all around the web, this, I think, does give an unfair advantage by default to those larger brands.
Paris: And this is why I think that if you're already a really, really popular brand and well-known, you have such a big, big lead when it comes to GEO, just because of the co-occurrence factor of the sheer volume of your brand mentions all around the web is giving you a massive head start here. And onto our last story of the day, Google adds Gemini-driven recommendations to Search Ads 360.
Paris: Google introduced its first monthly marketing platform product update. Among the changes is a redesigned Search Ads 360 home experience. Google says Gemini models now help consolidate marketing insights, generate recommendations, and streamline campaign execution. The announcement also includes AI-curated media opportunities in Display and Video 360.
Paris: These are paid advertising platform capabilities, not new organic AI search citation features, just to be clear. So Gemini and Gemini-driven recommendations are now being brought into Google Ads and Google Search Ads, and that makes a ton of sense. Well, I think this is a clear signal of AI automation continuing to get stronger and stronger within Google Ads.
Paris: All right. That's a wrap for episode 69. Thank you all for listening. Be sure to follow us and subscribe on YouTube. Check us out on Apple Podcasts and Spotify, and we'll look forward to seeing you on the next episode.
