ChatGPT referral traffic is getting harder to measure because links inside ChatGPT answers are reported to reach websites without their referral data. In this news roundup, Paris Childress covers that attribution gap and nine more GEO stories, from citation research and shopping-button bias to agentic orders and agent-to-agent commerce.
Links inside ChatGPT answers are reaching websites without their referral data. In this news roundup, Paris Childress, founder of Hop AI and co-founder of GEOforge, explains why that matters for attribution and covers nine more stories shaping Generative Engine Optimization this week.
Many links inside ChatGPT answers, not the citation source panel, are reported to arrive without their referral data, so tools like Google Analytics can't attribute the visit to ChatGPT. Reported ChatGPT referral traffic is down as a result. It isn't clear whether this is a bug or a permanent change, but either way it makes an already hard attribution problem harder.
Research by Chen and Koudas across 158,000 AI-generated answers found retrieval behavior explained about 32% of citation variation, against 2.8% for individual page characteristics. ChatGPT also sometimes inserts familiar brand names into its fan-out queries. Paris's read: make brand names more prominent in page titles, meta descriptions and H1s.
Wislr found 56% of ChatGPT shopping buttons across four monitored brands linked to major retailers and about 5% to brand-owned stores. OpenAI has not confirmed the selection behavior.
Forter reports agentic orders up more than four-fold since August, alongside rising automated fraud. Product.ai's survey of about 1,500 US consumers found almost 90% of AI shoppers verify recommendations elsewhere, and 21% have let an assistant complete a purchase. Paris compares it to the early days of e-commerce, when trust took time to build.
Spangle Halo puts a merchant-side agent in conversation with shoppers' agents over UCP, MCP and the browser. DeepLumen, now on nearly 2,000 Shopify stores, charges commission on AI-attributed orders rather than relying only on subscriptions.
Hi, everybody, and welcome to episode 61 of The GEO Show, brought to you by GEOforge, full self-driving for AI visibility.
Let's get into today's top stories, starting with some news from Peec. Peec AI launches customizable dashboards built by AI agents. Peec introduced Views, which allows users to build and share customized AI visibility dashboards. The release includes 80 available widgets spanning brand perception, citations, competitor performance, search queries, and shopping analytics.
Users can create dashboards manually or instruct Peec's AI agent to build them. Dashboards can be shared internally or externally through password-protected read-only links. All right. So this is a pretty interesting step by Peec, and it's really leveraging, I think one of the great powers of AI is to build custom dashboards.
We've seen recently that Claude has launched a couple of new artifact types. One of them is motion for motion graphics, which is really cool, and the other, I believe it's called dashboards, and so you can create custom dashboards as well. So I think there's always going to be some, kind of tension or internal competition from any SaaS tool, any third-party SaaS tool that has a native dashboard feature or functionality, and also has an MCP connector because I think that dashboards in general for SaaS tools are moving into Claude primarily because they're, they're so good at it.
So I'll be interested to see how successful the Peec dashboards are, given the fact that most people are going to be creating them also through a Claude MCP connector.
Next story. Academic research finds retrieval explains far more citation variation than page characteristics. Researchers Mahe Chen and Nick Koudas examined 158,000 AI-generated answers containing 468,000 citations. Their experiments found that retrieval behavior explained about 32% of citation variation in the reported setting, compared with 2.8% explained by individual page characteristics.
They also found that ChatGPT sometimes incorporated familiar brand names into its search queries, and controlled experiments demonstrated that brand names could influence product rankings even when product information was identical. Hmm. All right So I think what we're seeing here in query fan-out more and more is the insertion of brand names by LLMs.
So what it's saying here is brand names perhaps could influence product rankings even when the product information is identical. So maybe what we want to do here, what this might be implying, is to lean more heavily on brand names being included on those pages to increase their likelihood of winning citations.
So this could be inserting them into the page title, meta description, H1 tags, etc. But getting a stronger, let's say, keyword density of brand, brands appearing on those pages, it seems could have a positive impact on citations because the query fan-outs are increasingly including brand names in those queries Next up, ChatGPT shopping links disproportionately favor major retailers.
Wislr, spelled W-I-S-L-R, examined newly appearing shopping buttons inside ChatGPT answers. Across four product brands it monitored, 56% of buttons linked to major retailers, while approximately 5% directed shoppers to brand-owned stores. The researchers also tested outbound links and observed missing referral information.
Now, that's a big theme, the missing referral information, which I'll come back to. The sample's relatively small, and OpenAI did not independently confirm the reported retailer selection behavior or its causes. So it appears here that there is some bias in OpenAI's shopping buttons to display major retailers at the expense of smaller brand-owned stores.
And also, one of the big stories that's going around now is missing referral information from OpenAI, from ChatGPT. it appears now that many of the links that are included in answers, not in the citation source panel, but links in the answers, have been stripped of their referral UTMs, meaning that they're not passing that referral onto, let's say, Google Analytics, and that's making it much more difficult to observe your traffic, your referral traffic from ChatGPT.
And it's been reported that these numbers are down. Referral traffic from ChatGPT is way down as a result of this. So we're not sure if this is a bug, if they're planning to fix it, or if this is some permanent move, which would be unfortunate really, because it makes an already hard attribution issue even harder On to the next story.
Forter reports agentic orders increasing more than four-fold. Forter introduced new capabilities for identifying authorized AI shopping agents, managing their access, and enabling secure checkout. The company reports that agentic orders across its network have increased more than four-fold since August.
They also report increasing automated fraud, reinforcing the need to distinguish legitimate agents from malicious automation. All right, so the trend is very clear here. Personal agents from Grok Bot, Muse, Dots, Instinct, and this is just the beginning. There's going to be more. They are more and more increasingly going and shopping on behalf of their users.
And Forter is reporting a more than four-fold increase here in, actually this is a four-fold increase in the orders from agents. So this is a trend that is really drawing a line in the sand because we've seen Shopify opening its doors to agents while Amazon has shut the doors to agents. It's a very important and interesting trend to keep our eye on here, especially if you're in e-commerce.
On to the next story. Nearly 90% of surveyed AI shoppers verify recommendations before buying. 1,500 US consumers were surveyed by Product.ai, including 550 who had used AI assistants for product research.
And among those respondents, almost 90% verified recommendations through another source before purchasing. Just 21% of those AI shoppers reported allowing an AI assistant to complete a purchase from start to finish. I think it's just the beginning here. These numbers will grow as we trust these AI assistants more and more.
I kind of liken it to the beginning of e-commerce shopping on the internet, where most people were, not trusting enough to put their credit cards down, but that changed. I think this is going to change as well. People are going to increasingly trust their AI assistants to shop on their behalf, the more that they see that they can do it responsibly and accurately.
And our next story, G2 Research finds AI agent supervision consuming significant employee time. G2 released its 2026 AI Agents Report, and it found that 31% of surveyed B2B software buyers said employees are spending six or more hours weekly supervising AI agents. Another 84% reported that agent capabilities would influence their software purchasing decisions in 2027.
Very interesting. I think on the last point, 84% of software buyers saying that AI and agent capabilities would influence their software purchasing next year, no surprise there, of course, and 31% of people reporting a heavy workload of supervising AI agents. I think that is really a reality of the world we're living in, especially for heavy AI usage.
I can say that I probably spend at least 10 hours a week, supervising my Grok Bot, supervising Claude and ChatGPT, but I still think it's maybe much more productive overall The next story, Work4Flow launches a closed-loop knowledge optimization for ServiceNow. Work4Flow introduced a ServiceNow native knowledge article optimizer.
The system analyzes search behavior to identify knowledge gaps, recommend article improvements, supports human review, and allows teams to remeasure the results. It sounds like Content Forge in GEOforge. The product connects enterprise search relevance with knowledge management and AI agent readiness.
All right, so we've got a closed-loop knowledge optimization emerging here in an enterprise software as enterprise software capability. It's moving beyond just a specialized GEO tool. ServiceNow is a major software platform, and now they have, built in this capability for a knowledge article optimizer, so they're taking it very seriously Next up, Spangle launches an AI agent representing retailers to shopping agents.
Spangle introduced Spangle Halo, a merchant-side AI agent. Rather than simply exposing product information, Halo uses retailer-specific commercial intelligence to interact with external shopping agents. The product supports Universal Commerce Protocol, UCP, also MCP, and browser-based interactions. All right, we are finally getting there, agent to agent.
My agent will go and shop on a website, and it will interact with that website or that e-commerce store's own agent. And I think this is really the true future of online retail. It's going to be agent to agent. My agent will interact with the merchant's agent, and they will, between the two of them, figure out how to make the transaction as optimally beneficial for both sides.
All right, now the last story of the day, DeepLumen approaches 2,000 Shopify merchant installations. DeepLumen announced that its agentic page product has been installed by nearly 2,000 Shopify merchants, approximately doubling its August milestone. Its system structures merchant information for AI discovery and tracks orders attributed to AI referrals.
The company's commercial model charges merchants commissions on attributable AI-generated orders rather than relying exclusively on software subscriptions. that is a really interesting pricing model, and they're taking some risk here. They are basically charging only for, commissions on attributable AI-generated orders.
Really interesting here. And I think that AI-generated orders are going to continue to increase. They hit a big milestone here with 2,000 installations, and, this is going to be a really interesting space to watch. they're obviously optimizing the whole experience, and they are focused on making, the e-commerce merchant information more discoverable by AI agents and then charging a commission for that.
I think that's a very savvy move. And that will do it for today's episode 61. Thanks for tuning in, and we'll see you all in the next one.
