Welcome to episode 38 of The GEO Show, the briefing on Generative Engine Optimization and AI visibility. Brought to you by GEOforge.
In this episode, Paris Childress, founder of Hop AI and co-founder of GEOforge, covers ten stories: Profound's Google AI mode citation instability, Azoma's agentic commerce citations, Nexi's merchant catalog push in Europe, Mastercard Agent Pay credentials, BrandGhost's 11.8% cross-engine overlap, Machine Relations' long-tail citation map, AppScribed's live-vs-API tracking gap, Cloudflare's Disallow AI Training setting, the NYT "doom loop" filings, and Pyra's Primey race engineer.
In this episode:
🤖 Profound flags Google AI mode citation variability
- Profound's status page says Google AI mode citation data remains affected by variability. One of the category's largest vendors is publicly acknowledging that AI search measurement can go unstable when the model layer moves. Treat those numbers with a grain of salt.
🛒 Azoma: 86.5% of Alexa shopping citations are earned or social
- Dunnhumby Ventures invested in Azoma. Q2 analysis: 86.5% of citations behind Alexa shopping recommendations came from earned or social media (76% for Walmart Sparky), not brand.com. Off-site citation work is not optional in commerce GEO.
🇪🇺 Nexi × Refi Buy: catalogs as AI infrastructure
- European payments giant Nexi is helping merchants structure product catalogs for AI agents. A forthcoming survey of ~28,000 consumers across 11 countries found more than half used AI in the prior month to search, find, or buy. Machine-readable catalogs are becoming table stakes.
💳 Mastercard gives agents one-time payment credentials
- Mastercard Agent Pay into Alchemy's Agent Card means approved AI agents can use one-time tokenized credentials inside user-set constraints. Agentic commerce is moving from recommending to buying. Trust will follow the same path early online shopping did.
📊 BrandGhost: only ~11.8% cross-engine source overlap
- Across 7,902 citations on ChatGPT, Claude, Gemini, and Perplexity, source overlap is about 11.8%. In 41.8% of comparable recommendation sets, brand overlap was zero. Winning one engine does not win another. GEO is at least a two-platform game.
📈 Half of AI citations need 1,357 domains
- Machine Relations' index: top 10 domains are only ~7% of citations. Reaching 50% of citation share takes 1,357 domains; 80% takes 6,643. Reddit leads at 1.81%. Paris's read: this is a volume and long-tail game, not a single New York Times link.
🔬 Live ChatGPT vs API tracking can disagree completely
- AppScribed: one buyer query hit 5/10 live logged-out ChatGPT runs and 0/30 API runs. Broader set: 196,554 citations over 91 days, including pages ranked #80 or outside Google's top 100. Prompt tracking is probabilistic. You still need to be indexed. You do not need top 10.
🛡️ Cloudflare: reject training without killing search
- The Disallow AI Training setting publishes a no-training preference while qualified mixed-use crawlers keep indexing. Defaults often block training bots. For most brands maximizing AI visibility, Paris's call is clear: let the training bots in so models learn accurate brand facts.
📰 NYT filings and the "doom loop"
- Unsealed filings warn that AI products can weaken publisher economics and eventually degrade the content models depend on. For B2B brands the practical move is sharing proprietary knowledge without giving away the secret sauce.
🏁 Pyra Primey: race-engineer diagnosis from $69/mo
- Pyra's Primey layer explains competitor citation wins and next actions. Diagnosis and recommendation are moving down market. GEO, like SEO, remains a competitive game.
Subscribe to The GEO Show wherever you get your podcasts, and watch the full episode on YouTube.
💬 Question: If Cloudflare defaults to blocking training bots, did you check your settings this week, or are you still flying blind on what the models can learn about your brand?