Google AI Mode Loses 62% of Its Citations Every Day

About this episode

Welcome to Episode 7 of The GEO Show, where we break down the biggest developments in Generative Engine Optimization (GEO), AI search, and the rapidly changing world of SEO.

In this episode, Paris Childress, founder of Hop AI and co-founder of GEOforge, runs through seven stories reshaping AI search this week, including new daily-tracking data showing just how volatile AI citations really are from one day to the next.

In this episode:

🔍 Google says there's no special technical GEO layer for AI Search John Mueller confirmed there's nothing website owners need to do specifically for Google's AI Overviews or AI Mode beyond strong SEO fundamentals: crawlability, indexation, clear information architecture, and entity clarity. It's a Google-only position and doesn't mean AI visibility equals rankings, but it does undercut the mysterious "Google GEO hacks" some vendors are selling.

🧬 Pew Research finds AI fingerprints on 10% of the entire web, 35% of pages published since ChatGPT Pew analyzed 490,000 English-language pages from Common Crawl's July 2026 sample. AI authorship or editing shows up in 10% of all pages, and in 35% of pages with a publication date after ChatGPT launched in November 2022. AI-generated content isn't a differentiator anymore. It's ambient infrastructure, and the only moat left is information gain: subject matter expert interviews, proprietary data, benchmarks, and firsthand experience.

💻 NVIDIA and Perplexity launch Portable Computer for local-first AI Following earlier rumors of an NVIDIA investment valuing Perplexity at $30 billion, the two companies have introduced Portable Computer, which runs Perplexity entirely on-device and escalates to the cloud only when a task requires it. It's a serious signal for privacy-conscious companies not yet ready to send proprietary data into cloud-hosted models, and an early look at what AI on the edge looks like in practice.

📊 The first public, searchable ChatGPT Ads library has launched LLMPulse launched an independent ChatGPT Ads library on August 21 and updated it again on August 24, now covering more than 72,000 observed ads across nine countries, searchable by advertiser, message, country, and date. It's sampled, observational data, not official OpenAI numbers, but it's the first real crack at competitive intelligence for conversational ads.

📉 Daily AI citations are far noisier than classic SEO rankings Vendor research from Chloro tracked identical prompts across consecutive daily runs from August 1 to 16. A previously cited domain stayed cited the next day just 37.8% of the time in Google AI Mode, 49.9% in ChatGPT, and 68.5% in Perplexity. Losing a citation today is statistical noise, not proof your optimization failed. Report GEO from thirty to forty runs at a time, not a single execution.

📰 Muck Rack's 25-million-link dataset shows AI source strategy has to be platform-specific Muck Rack's analysis of more than 25 million cited links across ChatGPT, Gemini, and Claude found citation behavior varies materially by platform and query type. Industry trend queries cite journalism at more than twice the rate of how-to queries, and more than half of journalism citations came from content published in the last twelve months. Domain authority doesn't factor in at all. Relevance does.

🤖 Agent experience is emerging as the discipline downstream of GEO RedMonk's new framework draws a clear line between GEO, which gets a product discovered or recommended, and agent experience (AX), which determines whether an agent can actually use it. Agents are already crawling sites at a far higher rate than human visitors. The next question isn't just can AI recommend you, it's can AI operate you.

Subscribe to The GEO Show for ongoing analysis of the news, experiments, strategies, and emerging tactics shaping visibility across ChatGPT, Google AI, Perplexity, Claude, and the rest of the AI search ecosystem.

💬 Question: If you reported a lost AI citation as a loss this week, would it still look like a loss after forty runs instead of one?