New research suggests AI models often cite sources from search snippets without opening the page, which makes titles and meta descriptions part of AI visibility. Paris Childress covers the week's GEO news: ChatGPT Shopping's separate US and European systems, a ChatGPT attribution drop seen by Webflow, new ChatGPT Ads measurement partners, personalized ads in Google AI Mode, exact match domain experiments, GEO startup funding and the GEOforge MCP connector launch.
New research suggests AI models often cite a page from its search snippet alone. Paris Childress, founder of Hop AI and co-founder of GEOforge, covers this week's GEO news and what it means for brands that want to be found in AI answers.
In "Retrieved but Not Read", researchers Owen Walters, Lihu Chen and Francesca Toni found that models with web search often answered from search snippets without opening the underlying pages. Paris's takeaway: optimize titles and meta descriptions for AI retrieval, not only for clicks.
Peec AI found two systems: in the US, ChatGPT Shopping draws on OpenAI's own product index built from product feeds, while in Europe it relies on Google Shopping results through a search API.
Webflow observed an average 47% decline in ChatGPT attributed sessions from 8 October, linked to missing tracking parameters, with attribution recovering about a week later.
ChatGPT Ads is adding third party measurement partners and testing more prominent visual formats, and Google is testing personalized ads in AI Mode for users who voluntarily connect their apps. Paris argues that agents closing sales inside the chat would solve attribution.
A Peec AI customer experiment showed an exact match directory appearing in 38.6% of responses for one head term prompt, but in fewer than 5% for prompts that resemble real buyer questions.
Value Add VC estimates about $604 million has gone into 10 GEO startups, with Profound accounting for more than half. ALSA extended its seed round, Semrush opened MCP access, and Paris announced the launch of the GEOforge MCP connector.
Paris: Hey everybody, and welcome to episode 64 of The GEO Show, brought to you by GEOforge: full self-driving for AI visibility. It's Saturday, October 17th. We've got some big stories to cover today, and let's start with the biggest overarching theme for this episode, which is that AI discovery is fragmenting into different regional retrieval systems, while ad platforms in this space are racing to make AI-driven commercial outcomes measurable.
Paris: So we've got some big stories coming from Peec, some new academic research and more. So let's get right into the top story for today. The top story today comes from ChatGPT Shopping. Peec finds two different ChatGPT shopping systems in the US and Europe.
Paris: Peec has published an investigation based on approximately 150,000 lines of diagnostic retrieval data per tested prompt. Its findings suggest that ChatGPT Shopping currently uses different architectures across regions. The US uses OpenAI's Labrador product index supplied by product feeds, while Europe uses Google Shopping results retrieved through a search results API.
Paris: Peec also observed that its tested US retrieval system excluded sellers without a particular verification flag. Product descriptions were truncated at approximately 50 tokens during scoring. So geographic differences in AI visibility here may reflect fundamentally different retrieval infrastructure, not merely local consumer preferences.
Paris: So I think it's interesting how we can observe from the advertising side of AI and make inferences about the organic side. So clearly shopping is behaving very different. Shopping ads for ChatGPT is behaving very different in the US versus Europe. They're relying much more heavily on Google Shopping results in Europe, but that is a big risk because Google could take that away at any time.
Paris: So that does imply maybe different systems for GEO as well in the US versus Europe Okay, onto the next story. Academic research finds that AI assistants sometimes cite sources they never fully read. So this is researchers Owen Walters, Lihu Chen, and Francesca Toni published "Retrieved but Not Read: Understanding Why Search-Augmented LLMs Fail to Use Available Evidence."
Paris: The researchers evaluated one proprietary model with deployed web search and six open-weight configurations under controlled retrieval conditions. They found that models sometimes relied on search snippets only without opening the underlying pages. The behavior was associated with outdated answers and citations that did not faithfully support the claims.
Paris: All right, interesting stuff here. And, what to make of this? I think that because of the need for speed here, these models are sometimes through search retrieval process, only grabbing the snippet, which is technically the meta title and the meta description only, and then trying to synthesize an answer only from these snippets.
Paris: And if that is true, I think that even puts more of an onus on us as marketers to really optimize those snippets, so optimizing the page title and the meta description, not necessarily for keyword optimization the way we have been doing in SEO for the last twenty years, practically, but for AI bot retrieval.
Paris: So that might be different. We might want to think differently about that title and meta description in the snippet if we think that that is all that the AI bots are going to retrieve in this real-time search and retrieval process. Interesting stuff. All right, next story which the whole industry is following closely right now.
Paris: Webflow confirms ChatGPT referral attribution recovered after a significant disruption. This was observed by an investigation on October 16th from Webflow. They investigated a sudden decline in ChatGPT-attributed website traffic beginning October 8th across customer websites using Webflow Analyze or Optimize.
Paris: The company observed an average 47% decline in ChatGPT-attributed sessions. Webflow attributes the change to missing tracking parameters on ChatGPT links, and the company reports that those parameters returned on October 15th, and attribution subsequently recovered. All right, good news for those who are worried about this.
Paris: I think there was a bit of a panic that OpenAI was going to take away the, that referral attribution traffic, but it appears that it was just some sort of a temporary bug, and now it's back, so we can all breathe a sigh of relief. Next up, more from ChatGPT. On the ad side, ChatGPT Ads expands conversion reporting to third-party measurement.
Paris: OpenAI is expanding ChatGPT Ads measurement capabilities. The report changes include integrations with Northbeam, Rockerbox, and Incrmntal, alongside additional mobile measurement parameters. A redesigned conversions interface consolidates reporting, optimization status, and diagnostic warnings.
Paris: Advertisers can also retrieve weekly performance summaries. Separately, OpenAI is testing more prominent visual ad formats. These formats remain experimental and no general rollout or performance uplift has been confirmed. So I think ChatGPT Ads realizes that in order to continue to attract more and more investment in its ad platform, they will have to do better on conversion tracking in order to prove and demonstrate ROI.
Paris: So they're going deeper and deeper with more third-party measurement tools, which is great. They're testing more prominent visual ad formats and ultimately the success and viability of ChatGPT Ads once it moves past this novelty phase that we're in right now, it's going to be on measurable ROI, and ChatGPT will have to demonstrate that they are influencing downstream conversions, whether that be lead gen or whether that be purchases.
Paris: And I think that the way this is going to get resolved ultimately is when an agent representing the advertiser gets introduced into a conversation by an assistant, that agent then closes the sale or captures the lead inside of ChatGPT's ad platform without any click to an advertiser's landing page or website.
Paris: And then that really solves attribution because that entire conversation then which ends in a sale or conversion can be fully attributed to that advertiser's agent who comes in to sell and close the deal Which brings us to the next related story here, which is about Google Ads testing. Google Ads begins testing personalized ads in AI mode.
Paris: Google confirmed that it is testing advertising inside AI mode for users who voluntarily connect applications through personal intelligence. Connected information may help personalize the advertising experience alongside the user's question and surrounding AI-generated response. Google says advertisers cannot directly access connected applications.
Paris: It also describes restrictions on sensitive information and direct targeting from Gmail or Google Photos content. All right. So Google AI mode is, I think, waiting in the wings to take over as the default Google experience, and probably the only thing holding it back right now is Google's ability to monetize that through Google Ads.
Paris: So they are testing it, and this is, this is high stakes for Google. So for users who are voluntarily connecting applications through personal intelligence, that means, the ads are going to start probably looking at your Gmail history and your conversation history, search and conversation history with Google AI mode and classic search.
Paris: I don't know about Google Photos, but I think that the really, the more a user is willing to share their Google apps across Google's ecosystem and their behavior across Google Chrome, YouTube, Gmail, et cetera, then the more personalized and the more relevant ads that they will be served in AI mode.
Paris: And again, I think this could become an agentic experience, just like with ChatGPT ads. So let's watch this space carefully All right, the next story again is from Peec. Peec customer experiments reveal weakness in head term citation strategies. Peec published two customer experience experiments examining how GEO tactics perform under different circumstances.
Paris: In one case, UX agency Bricks tested a newly created directory with little traditional search authority. The directory appeared in 38.6% of responses to one exact head term prompt, but in fewer than 5% of responses, of responses to prompts more closely resembling constrained buyer questions.
Paris: The experiment involved a narrow set of conditions and does not demonstrate that exact match domains generally outperform authoritative websites. Okay, I've been hearing a lot of buzz lately that one of the successful winning strategies or tactics in GEO is EMD or exact match domains. And this is really refuting or rebutting that strategy, because the head terms, which exact head term prompts really are the ones that match up to those exact match domains.
Paris: They did well, but once you move to prompts that really more closely resemble real buyer questions, longer tail prompts with much longer strings of words, then those exact match domains and those exact head term prompts plummeted to just 5% of visibility. So beware if you're thinking that an exact match domain approach could be a silver bullet in GEO. It might be, a strategy to win visibility for, these exact match head terms. But that really doesn't represent the long tail of prompts in the same way that it used to in the, in the SEO-driven, keyword world. Next up, new funding analysis estimates $604 million invested across 10 GEO startups.
Paris: Value Add VC published a financing overview of 10 selected GEO and AEO companies. Its compilation estimates that approximately 604 million in disclosed cumulative funding in just those 10 companies. Profound, of course, number one in that category, accounts for approximately 335 million, more than half or 55% of that total, and others in that bucket of 10 include AirOps, Bluefish, Brandlight, Peec, and other competitors.
Paris: So actually last week in the How to Web conference in Bucharest where I was pitching, I reported that there was approximately 550 million invested in this category. Here, that estimate has risen to 604 million, and that is just the top 10. If we probably look at a longer tail here, I'm sure we can get closer to 700 million, and that is astounding given the fact that this category is just barely two years old.
Paris: What's interesting, though, is that Ramp data has shown that less than 1% of its software buying audience that's on Ramp, fewer than 1% have bought into this category yet. So you've got massive investment flowing into the category, but less than 1% of what I think are qualified buyers have yet to buy in the category.
Paris: So that is just massive green space, in my opinion. Okay, and speaking of investment, ALSA, A-L-S-A secures additional investment for trusted agent access and transactions. ALSA announced a seed extension involving Tether and Hiven as strategic investors. The company reports cumulative funding now has exceeded $9 million.
Paris: The amount raised specifically in the extension was not disclosed. Its technology connects enterprise AI agents to external information models, APIs, and services through permission-controlled infrastructure. The architecture includes source provenance, authorization, spending limits, transaction processing, and auditability.
Paris: So an example here of an AI agent infrastructure that is expanding beyond model execution and into trusted access to information and commercial services. And yet another example of smaller investments that are going into this category or into related closely adjacent categories. So it's just really on fire All right, let's wrap it all up with some news from Semrush.
Paris: Semrush documents agent-compatible access to its marketing intelligence. Semrush's updated documentation describes MCP access for compatible AI assistants. Depending on the customer's subscription and permissions, assistants can retrieve SEO, traffic, competitor, and project data. Supported clients include ChatGPT, Claude, Microsoft Copilot, and Perplexity.
Paris: Authentication uses OAuth or API keys and requests consume the usual API allowance. So Semrush is following a lead with other major SaaS players in going increasingly headless and allowing their customers to do more with MCP so that they never actually have to enter the Semrush platform at all. They can look at their SEO statistics, traffic, competitor data, project data, all, through an intermediate layer of, let's say, Claude or ChatGPT or Microsoft Copilot.
Paris: And this is certainly the direction I think of most of SaaS and even most of the GEO category as well. And with that, I would like to also proudly announce that GEOforge has just launched our own MCP connector. And so we're following suit here, and we will allow our customers to do the same thing, to use, basically use our product in any way they would like, access the data, take actions, build reports, et cetera, analyze data all through Claude or ChatGPT or their LLM of choice via an MCP connector.
Paris: All right. That will wrap us up here for episode 64. Thank you for listening and see you all on the next one.
