Always-on AI agents like OpenAI's new Dots learn preferences, carry context, and work in the background, so one-off prompt tracking will capture less of a brand's AI visibility. Paris Childress also covers Writesonic's 9 million AI answers (82% of citations on bottom-of-funnel commercial prompts went to third parties, via Search Engine Land), Perplexity's evidence-aware embedding model, Profound's 11-day citation half-life, Informa TechTarget's MCP server, 10Fold's AI visibility reporting survey, Brandpoint's Optimize360, Zen Media's industrial automation benchmark, and Findrix's Australian AI Source Index.
Always-on AI agents will outgrow prompt tracking. OpenAI's new Dots learn preferences over time, carry context between tasks, and work in the background on their own, so a brand's AI visibility will depend on more than the answer to a single tracked prompt. Also this week: third parties win most AI citations on buying prompts, AI citations fade fast, and source authority depends on the market you're in.
In this solo roundup, Paris Childress, founder of Hop AI and co-founder of GEOforge, covers nine stories.
OpenAI launched Dots on September 29: always-on agents powered by GPT-6 Astra, each with its own cloud computer, that learn preferences over time and connect to more than 4,000 apps. They are rolling out to ChatGPT Pro and Business Premium users, with an Enterprise beta, and Pro users in the EEA, Switzerland, and the UK don't have them yet. As agents carry history and start conversations on their own, a recommendation depends on more than the current prompt. Stateless prompt testing becomes less representative, and share of voice based on prompt tracking is a starting point, not the whole picture.
Perplexity released pplx-embed-v2-context-9b-preview on September 30. Instead of training retrieval around one gold passage, the contextual embedding model learns to retrieve both the answer chunk and the supporting context needed to verify it. The unit being optimized for retrieval is becoming an evidence bundle, not an isolated quotable paragraph.
Not long. Profound measured 883,000 pages across ChatGPT, Perplexity, Claude, Google AI Overviews, Google AI Mode, Gemini, and Microsoft Copilot over twelve months. The median page loses half its citation share within 11 days of peaking, and 78% of cited pages fall to half their peak within two weeks. Engines decide independently: cross-engine correlations run from just 0.03 to 0.09. A falling citation rate can signal aging content, so watch citation persistence and refresh or replace pages as they decline.
"This is a really critical exercise in GEO, is to observe citation persistence and be poised and ready to update and refresh pages as they start to decline." (Paris Childress)
Mostly from third parties. Writesonic analyzed 9 million AI answers across nine platforms for 400+ enterprise brands and their competitors. On bottom-of-funnel commercial prompts, 82% of citations went to third-party sources and 3% to the brand's own pages. When owned content was cited, AI visibility rose 5 times, and owned citations lasted 3 to 9 times longer, while up to 58% of citations never reappeared after their first showing. The data comes from Writesonic customers, so it is not a market consensus. For Paris, it confirms that third-party citation building is the most important GEO tactic right now.
Subscribe to The GEO Show wherever you get your podcasts, and watch the full episode on YouTube: https://www.youtube.com/watch?v=t6YnwMMyMe4
Paris Childress: Hi, everybody. Welcome to episode 53 of The GEO Show, brought to you by GEOforge: Full Self-Driving for AI Visibility. And today is October 1st, the start of Q4, and we've got some great stories leading off with the biggest announcement in quite a while from OpenAI about Dots. So let's get right into that.
Paris Childress: OpenAI Dots makes AI discovery persistent and personalized. OpenAI launched Dots on September 29th. These are always-on agents powered by GPT-6 Astra with their own cloud computer. They learn preferences over time, work continuously, carry context between interactions, and connect to four thousand plus apps. OpenAI is initially rolling them out across Pro, Business, Premium, and Enterprise, and unfortunately, we do not have them here in Europe yet. So this is the biggest news, the biggest buzz in all of AI right now.
Paris Childress: Grok kind of got this trend started, and then Meta Muse followed with a wildly popular product, and now OpenAI is jumping in the game. So what does this mean for GEO? Well, this stateless prompt testing, I think, is gonna become less representative of the real user experience. A recommendation can now increasingly depend on accumulated history, the connected tools, preferences, and previous work, and not just the current prompt.
Paris Childress: So I think that the introduction of these agents will steadily replace a lot of the prompting and the back-and-forth chatting that's happening now that all is currently initiated by the user in classic AI chatbots. But now when you have an AI assistant that is proactively initiating conversations, there's gonna be less actual prompt tracking. So I do think that this share of voice, which is based on prompt tracking, I think it's short-lived really.
Paris Childress: It is the natural starting point, but things are gonna develop and a brand's AI visibility is going to be much broader in the age of these agents than simply tracking when a user initiates a specific prompt. So it's gonna be very interesting to see how this develops.
Paris Childress: Next up, Perplexity trains retrieval to return evidence, not just the answer passage. Perplexity released pplx-embed-v2-context-9b-preview on September 30th. Instead of training retrieval around one golden passage, the new method retrieves both the answer-bearing chunk and supporting context needed to understand or verify it. Perplexity describes this as a new contextual embedding training method and benchmark for evidence-aware retrieval. So Perplexity, as it has done many times before, is really leading the innovation around the retrieval process, answer retrieval.
Paris Childress: And what it appears now is that the unit that's being optimized for retrieval is becoming an evidence bundle, not just simply an isolated quotable paragraph. All right.
Paris Childress: This next story builds on yesterday's announcement about Profound citing AI citations half-life. Profound finds the median AI citation half-life is just 11 days. Profound analyzed eight hundred and eighty-three thousand pages across seven answer engines, creating roughly 1.2 million page engine life cycles. Its September 30th study found that the median passage loses half its peak citation share within 11 days. Seventy-eight percent fall below half peak within two weeks.
Paris Childress: Citation lifecycles were also largely independent across engines, with cross-engine duration correlations of only zero point zero three to zero point zero nine percent. This is Profound's own dataset and methodology. Profound is doing some really interesting analysis here about how long a citation sticks and what is the half-life of a citation. And what we're finding is that citations stick for very short periods. They have very short half-lives, and that makes it difficult for people that are building citations.
Paris Childress: But it also presents a great opportunity for new entrants to come in and get citation visibility because there's only really, on average, an eleven-day half-life. So citation persistence is now a new competitive measurement, and what we're seeing is that falling citation rate can really reflect content that's aging more so than a retrieval change or a source replacement or a model change or any other kind of changes happening on the model side.
Paris Childress: So if your citations are aging and dropping off, it's a strong signal to either look at those pages and do refreshes and update the published dates on those pages, or perhaps just replace those pages with new pages on the same topics that are more up-to-date. But this is a really critical exercise in GEO, is to observe citation persistence and be poised and ready to update and refresh pages as they start to decline.
Paris Childress: Next story. Nine million AI answers show commercial citations overwhelmingly come from third parties. Search Engine Land did an analysis of roughly nine million AI answers and found that third-party sources accounted for eighty-two percent of citations on bottom-of-funnel commercial prompts, while owned pages account for just three percent. Yet when owned content was cited, it was associated with a five times visibility lift, and owned citations reportedly persisted three to nine times longer than third-party citations. Up to fifty-eight percent of citations never appeared again after their initial appearance.
Paris Childress: This underlying data comes from Writesonic customers, so it should not be treated as a market consensus. This underscores really that the most important tactic in GEO right now is third-party citation building, which is getting your brand mentioned on third-party sources, getting your content placed, cited, republished in third-party sources because this analysis really makes it clear eighty-two percent of the citations for the strong high-intent prompts, bottom-of-funnel commercial prompts are pulling from third-party sources. Owned pages, your brand, your website is just three percent.
Paris Childress: And yes, those are stickier, they last longer, but it's still only three percent. Next story. Informa TechTarget makes proprietary buyer intent directly callable by AI agents. Informa TechTarget launched an MCP server for buyer intelligence on September 29th, making its person-level intent data accessible inside Claude, ChatGPT, Copilot, Gemini, and custom AI systems. This person-level intent data contains fifty-eight million-plus permissioned first-party B2B audience members and says that its network produces millions of exclusive intent signals daily.
Paris Childress: So what we're seeing now with this announcement is another example of high-value proprietary datasets that are making themselves available to AI engines via MCP, and this is bifurcating this retrieval through search alone process. I think increasingly this retrieval that happens in real-time when a user's prompt triggers basically an answer that requires some type of real-time research.
Paris Childress: Right now, predominantly that goes to search, but I think increasingly they're gonna start to tap into these other additional large datasets that are now making themselves available, and these agents are gonna be able to query them directly.
Paris Childress: Next up, 10Fold folds AI visibility into B2B communications intelligence. 10Fold's MetricsMatter five point oh combines earned media, social web activity, competitive data, and visibility across five major large language model surfaces. In its survey of four hundred US and European B2B technology marketing decision-makers, fifty-eight percent said AI search or LLM visibility is already included in integrated reporting, while only thirty-eight percent connect communications and visibility metrics to pipeline or revenue influence.
Paris Childress: The platform also introduces what they call share of answer, which is a new metric which is probably similar to share of voice. So AI visibility is migrating into communications and executive reporting. It's not just remaining an SEO specialist metric. And what is most interesting about this survey is that the majority of people are looking at their LLM visibility now in some sort of an integrated report, but only thirty-eight percent can really connect the attribution all the way through to revenue or pipeline.
Paris Childress: And I think the important point here is that I think the best way to do this is that in the initial intake form, when you capture a lead, however that happens, you need to ask the question, "How did you hear about us?" And then have one of those selections be AI chat or AI chatbot like a ChatGPT or Gemini, and have people self-report how they found you.
Paris Childress: And if that makes it into the CRM, then later you can attribute, at least directionally attribute how much of your pipeline, how much of your forecasted revenue was influenced most probably by a first touch by an AI chatbot. And that's gonna get you a lot closer to understanding the pipeline or the revenue influence of your investments in GEO.
Paris Childress: Next up, Brandpoint launches GEO measurement with a built-in distribution network. Brandpoint launched Optimize360 on September 30th for PR and marketing communications teams. The platform measures AI-powered search visibility and competitive positioning using more than 30,000 prompts per month per brand, then connects those insights to Brandpoint's content and distribution infrastructure. Brandpoint separately advertises a network of 3,000-plus publications and 800-plus guaranteed placements. So another great example of PR muscling its way into GEO and leveraging the distribution part.
Paris Childress: As we discussed earlier on this episode, off-site citation building is the most important and the most impactful thing you can be doing to improve your AI visibility for GEO and PR agencies with large partnerships and publication networks that can guarantee placements like Brandpoint here are in a really strong position because now they really need to just add the content layer, the measurement layer and the content generation layer.
Paris Childress: But they have the hardest thing already in place, which is they have the distribution network pretty much solved for.
Paris Childress: Next up, industrial automation data shows why vertical GEO needs vertical logic. Zen Media analyzed one thousand buyer prompts and four thousand responses within the industrial automation sector across ChatGPT, Gemini, Perplexity, and Grok. They all had to do with industrial automation. The average tracked company appeared in only about four percent of responses versus twenty-one percent for Siemens. Roughly half the prompt set carried direct purchase intent. Manufacturer sites and authorized distribution channels played a particularly prominent role in the observed answers.
Paris Childress: So here the important source classes change with the market. In industrial categories, the specs, the distributors, certifications, support coverage, and integration compatibility can matter more than generic media authority alone. And it's important that we don't just slap a generic GEO strategy on all sorts of industries because they really differ. And this is a great example coming from industrial automation.
Paris Childress: The GEO visibility and AI visibility here is not just about reaching out and getting media authority, but in really thinking about how and where you publish accurate specs, information about your distributors, your certifications, your support, what's covered, what's not, integration compatibility, et cetera.
Paris Childress: All right, and our last story coming from Australia supports what we just talked about here. Australia study shows source authority is local, category-specific, and engine-specific. Findrix's Australian AI Source Index covers twenty-three thousand three hundred and ninety-six buying short list prompts across nine markets and three AI surfaces. Company retailer business websites accounted for sixty-four point six percent of retrieved sources overall. But source composition varied materially by vertical and by engine.
Paris Childress: More than eighty-five percent of retrieval in clinics, home services, and finance and accounting came from local .au domains versus just thirty-two point nine percent in B2B software. Google AI Mode also cited sources far more often than ChatGPT in the measured set. So again, a universal best GEO sources list is not really defensible. Authority is a function of your market, your category, which engine you're measuring, and the buyer's question. So we always need to put this in the right type of context. What market are we in?
Paris Childress: What language is the user speaking? And what particular industry are we in? And that's really important to consider the category and the context of the industry that you're tracking for.
Paris Childress: All right, that'll do it for episode fifty-three. Thank you all for tuning in, and we'll see you all on the next one.
