Google AI Mode is citing far fewer sources per answer, according to one dataset: Maeva Cifuentes's independent analysis of Scrunch tracking data finds about 16 cited sources per answer in the summer fell to about 4 by mid-October, about 74% fewer. Paris Childress stresses the figure is one unreplicated dataset, then explains how OpenAI says ChatGPT picks ads separately from answers, and what Sundar Pichai's agentic layer across Search, AI Overviews and Gemini means for websites.
Google AI Mode may be citing far fewer sources per answer. One new dataset finds AI Mode went from about 16 cited sources per answer in the summer to just about 4 by mid-October, about 74% fewer. Paris Childress's point is that if this holds, the competition for an AI Mode citation gets much narrower, and the pages that keep their citations look like the ones that already win in organic search. He is just as clear that it is one dataset, not yet replicated. The episode also covers how OpenAI says ChatGPT picks ads, and Sundar Pichai's description of one agentic layer across Search, AI Overviews and Gemini.
One dataset says yes. Maeva Cifuentes ran an independent analysis using tracking data from Scrunch and found:
Paris sees AI Mode as the likely future Google default, and says Google may be experimenting.
Not reliable enough to treat as settled. Paris flags four caveats:
"Regardless, if this is true, with fewer sources per answer, the competition for a citation link is much narrower."
The pages that keep their citations look like the ones that already win in organic search and get cited consistently across engines.
Through a pipeline that OpenAI says runs separately from the system that writes answers. In an engineering post, OpenAI describes the flow:
Ad sessions keep context across turns and conversations. Personalization uses non-sensitive information from past conversations, and only if the user opts in. OpenAI says ChatGPT has no dependency on the ad system, so a slow or failed ad never delays the answer. The post also repeats earlier figures: a one-billion-dollar annualized run rate in under two hundred days, and more than one hundred thousand advertisers since launch.
Only partly. The separation between ads and answers is OpenAI's own description, and nobody has audited it independently. OpenAI does not disclose how much weight each signal gets, a run rate is a projection rather than collected revenue, and there were no CPA, ROAS or retention benchmarks. Still, Paris calls it the clearest public description yet of how a conversational ad auction works. Relevance comes from the whole conversation and the user's history, not keywords, so paid and organic teams will both have to think about intent that way. And even if the systems are separate, an ad shown next to an answer still shapes what users trust and click. His practical advice: write ad creative for the problem the user is working through across several turns, not for a single query.
It is how Google CEO Sundar Pichai described search, AI Overviews and Gemini coming together, in a podcast interview with Alex Heath. He said AI agents will reach websites in two ways:
Pichai said the agentic layer gives a unified product experience, but did not say the products would merge. He said agent features now mostly for Ultra and Pro subscribers will reach more people over the next few months, and that task-style agentic queries are rising steadily, without giving figures. He did not explain how ads would work when an agent is doing the browsing.
Make the site work for both ways in. Paris's checklist:
"If search becomes one box that does tasks, what does a ranking really mean anymore?"
Fewer citations, ads that sit beside answers, and agents that act instead of read all point the same way: the pages and actions that are easiest for machines to trust and use are the ones that will keep showing up.
Paris: Hi, everyone, and welcome to The GEO Show, brought to you by GEOforge, full self-driving for AI visibility. I'm your host, Paris Childress, and today we have a light day, but still high impact. And the headline here is a new dataset showing that Google AI Mode is citing far fewer sources per answer.
Paris: So why don't we jump right into that top story here? One dataset finds AI Mode now cites about 74% fewer sources per answer. Maeva Cifuentes ran an independent analysis using tracking data from Scrunch. She finds that Google AI Mode went from about 16 cited sources per answer in the summer to just about four by mid-October.
Paris: Only around 9% of pages that were cited before kept their citation for the same prompt. Pages ranking in Google's top three kept theirs far more often than pages outside the top 20. All right, so a lot of shakeups happening in Google AI Mode, which I believe is the future of Google, the future Google default, and Google is shaking things up, maybe experimenting a little bit. But let's keep in mind this is coming only from one source.
Paris: No one else has replicated this yet. Google hasn't commented. And one brand made up 42% of the prompt-page pairs. Regardless, if this is true, with fewer sources per answer, the competition for a citation link is much narrower. The pages that keep their citations look like the ones that already win in organic search and get cited consistently across engines. Next story, OpenAI explains how ChatGPT picks ads and says ads are kept separate from answers.
Paris: OpenAI published an engineering post on how ChatGPT selects ads. An orchestrator model reads the conversation and nominates candidate ads. A separate auction ranks and prices them, and a final check can still drop them. OpenAI says this runs in a different system from the one that writes ChatGPT's answers.
Paris: All right. So this is a post-trained language model which runs the ad flow. It writes retrieval queries that capture commercial intent, filters the candidates, and nominates them. OpenAI says this model doesn't set prices or pick the winner. The auction handles admission thresholds, pacing bids, quality-adjusted rankings, and pricing.
Paris: An ad only counts as delivered once ChatGPT approves and displays it. And showing no ad is a valid outcome. Ad sessions keep context across turns and conversations. Personalization uses non-sensitive information from past conversations and only if the user opts in. OpenAI says ChatGPT has no dependency on the ad system.
Paris: A slow or failed ad never delays the answer. The post also repeats earlier figures, a one-billion-dollar annualized run rate in under two hundred days by ChatGPT ads and more than one hundred thousand advertisers since launch. Very interesting and very appreciative that OpenAI has disclosed all this about their ad platform.
Paris: A few caveats. The separation between ads and answers is OpenAI's own description. Nobody has audited that independently. OpenAI does not disclose how much weight each signal gets. A run rate is a projection. It's not revenue that's already collected, and of course, there were no CPA or ROAS or retention benchmarks given. Still, this is the clearest public description yet of how a conversational ad auction works.
Paris: Relevance comes from the whole conversation and the user's history, not keywords. And both paid and organic teams will have to think about intent that way.
Paris: So this is the ads system that can't touch the answer, and this is a promise that marketers should be able to rely on. Of course, the counterpart to that is even if the systems are separate, an ad shown next to an answer still shapes what users trust and click.
Paris: So some practical advice, if you run ChatGPT ads, write creative for the problem the user is working through across several turns, not for a single query. Next story. Sundar Pichai of Google says that Search, AI Overviews and Gemini are moving toward one agentic layer, and agents will reach sites through pages and protocols.
Paris: Google CEO Sundar Pichai gave a podcast interview to Alex Heath. He said AI agents will reach websites in two ways, by browsing pages like a person and through agent protocols. He described an agentic layer that brings Search, AI Overviews and Gemini together. He also said agent features that are now mostly for Ultra and Pro subscribers will reach more people over the next few months.
Paris: Here are some more details on that. Pichai called browsing a very elegant way that keeps users in the loop. He called protocols a far more scalable approach that will take longer to build. He pointed to MCP, payment protocols, and Google's Universal Commerce Protocol, UCP. He also said that the agentic layer gives a unified product experience across Search, AI Overviews and Gemini, but he did not say that this meant the products would be merging.
Paris: He also said that Google's data shows task-style agentic queries rising steadily, but he gave no figures on that. On the ad front, he said that the long-term answer is to put users first. He didn't explain how ads would work when an agent is doing the browsing. Okay, why is this important? Well, Google's CEO here is describing two ways into a website, agents reading pages like humans do, mimicking human behavior, and second, agents calling structured endpoints.
Paris: And websites in the future will likely need to work for both. This begs the question for SEOs, if search becomes one box that does tasks, what does a ranking really mean anymore? I think the way that most site owners can prepare for this is first list the key actions on your site, such as a quote, a booking, or product data, and check which ones an agent could complete through a feed, an API, or a protocol without a screen built only for humans.
Paris: Okay, and that's all we've got for today. Thank you all for tuning in, and hope to see you all in the next one.
