Google changed its privacy policy so that, for users who connect Gmail or Google Photos to Search, the context it draws can personalize ads, including in AI Mode, with no advertiser controls or reporting yet. Paris Childress also covers OpenAI data showing 20.2% of consumer ChatGPT messages are now information seeking, Shopify's agentic knowledge base, Visa's AI agent partners, and Semrush confirming poisoned Google results in SEO tools.
Google changed its privacy policy so that, for users who connect Gmail or Google Photos to Search, the context it draws can personalize ads, including in AI Mode, with no advertiser controls or reporting yet. Paris Childress also covers OpenAI data showing 20.2% of consumer ChatGPT messages are now information seeking, Shopify's agentic knowledge base, Visa's AI agent partners, and Semrush confirming poisoned Google results in SEO tools.
Yes, for users who connect them. Google changed its privacy policy so that, when users connect Gmail or Google Photos to Search, the summaries and inferences Google draws from that content can personalize ads, including in AI Mode.
Advertisers get no controls or reporting for this yet. Paris's advice: break out AI Mode performance wherever Google allows, so you can see what changes.
OpenAI's own data points that way:
The change is gradual, about three and a half percentage points over two years. The scale is not: at about 1.2 billion weekly active users, roughly 20% information seeking makes ChatGPT a massive new search engine, which raises a fair question about why AI referral traffic still looks so small.
Shopify built the store knowledge base that AI shopping agents use to answer customer questions directly into the merchant admin:
What an AI agent says about a store is becoming a data layer that merchants approve. The knowledge base is the new product feed, so review the auto-drafted FAQs before publishing.
On its earnings call, Visa said nearly 50 AI agent partners are live in production and that it expects that number to double in the coming months. Its focus is confirming an agent's identity, linking it to a real customer, and checking that its spending follows the customer's instructions. Early use cases are B2B, such as automated bill payments and invoice settlement.
Agent checkout is real but early. Watch which AI assistants connect to payment networks' agent programs, because that is where an AI recommendation is most likely to turn into an AI purchase first.
Semrush confirmed that Google is serving some deliberately wrong results to data collection tools. In one example, a Brazilian search for a baby stroller returned a dental supply page. Earlier vendor estimates put the problem at about 11% of DemandSphere's daily results pages and 8% for another vendor.
Third-party ranking and AI Overview data now has a known but unmeasured error rate. Paris sees 8% to 11% as noise, not collapse, and advises checking any drop against Search Console impressions before you report it.
Paris: Hi, everybody. Welcome back to "The GEO Show," brought to you by GEOforge, full self-driving for AI visibility. I'm your host, Paris Childress, and we've got some great stories on the docket today. Let's get right into it first up, Google lets connected Gmail and Photos context shape ads, including in AI Mode. Google changed its privacy policy. If you connect Gmail or Google Photos to Search, the summaries and inferences Google draws from them can now be used to personalize ads. That matters most in AI Mode, where Google is testing ads for people who use personal intelligence.
Paris: Advertisers don't see your inbox, and sensitive categories are excluded. So I think we're starting to see glimpses here of the future of Google Ads in AI Mode. And I think also Google realizes that they're going to have to compete with ChatGPT ads in terms of the degree of personalization that gets brought in. So they're now allowing users who opt in to personal intelligence to connect their Gmail emails and Google Photos in order to provide more context for personalized ads in AI Mode. And this policy change took effect on October sixteenth So this is pretty major.
Paris: Google's most personal user data can now help pick which ads appear inside AI answers, and advertisers get no controls or reporting on it, at least not for now so does this make AI Mode now the best targeted ad service on the internet? And also the least transparent one, perhaps. My advice here is if you run Broad Match, AI Max, or PMax, break out AI Mode performance wherever Google lets you. Don't compare it to classic search until there is proper reporting. Next up, ChatGPT is slowly but steadily shifting from a writing tool to a search engine.
Paris: OpenAI's data confirms that about one in five ChatGPT consumer messages now seeks information. OpenAI's latest signals data shows that twenty point two percent of sampled consumer ChatGPT messages in September were looking for information, up from sixteen point seven percent in July of twenty twenty-four. Writing requests fell from thirty-four percent to twenty-one percent over that same period, and now the two are nearly level. So a little more detail here. In August, information seeking was twenty point seven percent and writing was twenty point six percent.
Paris: OpenAI's definition of seeking information includes questions about products and services for sale. The methodology, about one million messages a month from adult users on the free Go Plus and Pro plans classified automatically. Enterprise and Codex are excluded here. And practical guidance, such as how-to advice and tutoring, was the largest category at 28.1%. So what we are seeing is ChatGPT drifting from writing tool towards search, and product questions fall inside the information category. So this begs one question, if one in five messages now is a search, why do AI referral traffic numbers still look so small?
Paris: Another thing to point out here is that this is not anything explosive. This is three and a half percentage points over a two-year period, so let's keep that in mind. And there's one other thing I want to point out here, which is that OpenAI's overall usage and user growth, that's what's explosive. Over this two-year period, I don't know where they were two years ago, but now they're at about one point two billion weekly active users. So it is interesting to look at the ratio of using ChatGPT as a writing tool versus using it for information seeking.
Paris: But the big story is that still, twenty percent, roughly twenty percent information seeking from one point two billion weekly active users means that this is a massive new search engine. Our next story comes from Shopify. Shopify moves its AI agent knowledge base into the merchant admin section. Shopify has built the store knowledge base that AI shopping agents use to answer customer questions directly into the merchant admin. Shopify drafts FAQs from the storefront, and agents only see them once the merchant publishes them. The old standalone app shuts down on fifteenth of January 2027.
Paris: More details here. This lives under the sales channel's agentic knowledge base section, and draft FAQs are unpublished until the merchant reviews them. It holds up to one thousand FAQs. Stores with more than a thousand FAQs were not migrated automatically and must re-enter the FAQs they want to keep before before the fifteenth of January. Published FAQs are available to API apps, such as the Shop app through Shopify's storefront MCP connection. All right. Very, very interesting here as Shopify and e-commerce is essentially going more and more headless.
Paris: What an AI agent says about a store is becoming a data layer that merchants approve rather than just crawling a website. So essentially, the knowledge base of a store is becoming the new product feed. And of course, if you are a Shopify merchant, definitely check the auto-drafted FAQs before publishing them. Our next story comes from Visa. Visa says about fifty AI agent partners are in production, and trust is the bottleneck. On its earnings call on Tuesday of this week, Visa said nearly fifty AI partners are now live in production and that it expects that number to double in the coming months.
Paris: CEO Ryan McKearney said, "Trust is the biggest barrier to scaling payments made by AI agents." Visa's focus is confirming an agent's identity, linking it to a real customer, and checking that its spending follows the customer's instructions. Early use cases are business-to-business, such as automated bill payments and invoice settlement. So agent checkout is real for sure, but still very early. And interestingly here, it's starting more in B2B, more so than in consumer shopping So what's gonna be very interesting here to watch is which AI assistants connect to the payment network's agent programs like this Visa payments program.
Paris: That's where an AI recommendation is most likely to turn into an AI purchase first. Next up, Semrush confirms poisoned Google results are affecting SEO tools. Semrush published a post confirming that Google is serving some deliberately wrong or poisoned results to data collection tools. Semrush says it filters out clearly invalid results. Google hasn't commented, and Semrush didn't say how much of its own data was affected. Semrush gave an example. A Brazilian search for a baby stroller returned a dental supply page. All right, so they're not even trying to even jumble up the results here.
Paris: They're producing entirely different category of results. Earlier vendor estimates put the problem at about eleven percent of DemandSphere's daily results pages and eight percent for No-Nozzle. Neither described its method according to PPC Land. So now this really changes how we see these rank tracking tools. Ranking and AI Overview data from third-party tools now has a known but unmeasured error rate, which leads us to ask how much of the industry's AI Overviews on X percent of queries research can we still trust? That said, eight percent or eight to eleven percent error rate is still really just noise.
Paris: It's not really collapse of the data set or the category. So my advice here is to check a ranking or AI Overview drop against Google Search Console impressions before you actually report on it. All right, this next story is a correction from something we reported on yesterday. Google says that disabled Discover source taps were a bug. Yesterday, we reported a Discover test that disabled taps on the sources in AI summary cards. Rajan Patel, Google's VP of engineering for search, says it was a bug in a Discover experiment, and Google has turned the experiment off and taps work again as intended it is an interesting response though, because it was a bug in an experiment.
Paris: So then what was the experiment actually testing? All right, let's finish up with some quick lightning round headlines. On October twenty-seventh, another unconfirmed Google ranking volatility spike, which follows earlier spikes around the sixteenth, twentieth, and twenty-first, and twenty-fourth of October. Next, iOS twenty-seven point two is out, and in it, Siri AI can now remember past conversations. It draws on more sources, including Safari history and reading lists, and it adds five languages. It's still unavailable, though, in the EU and mainland China, sadly.
Paris: Last lightning round update, John Mueller reports on Search Console delays. Delays in the page indexing report rarely matter, and real changes show up first in the performance reports, he confirms. Okay, that will do it for today's episode. Thank you all for listening, and see you on the next one.
