Decoding AI for Marketing

The episode surfaces a handful of genuinely useful concepts - share of prompt as a GEO metric, chunk-size optimisation for RAG, and the AI-slop delisting risk - but roughly half the runtime is consumed by host nostalgia (Larry Page anecdotes, MMA great debates history, IAB alumni bonding) and circular affirmations that dilute the useful-ideas-per-minute rate substantially.

“83% of AI citations originate from pages that do not rank in the traditional Google top 10”

“the number of words that you use in describing it can be an issue, because if the AI is doing the rag pattern, it's looking for a chunk of probably about 40 to 60 words”

'Share of prompt' as a distinct AI-era metric and the observation that cross-platform narrative coherence (not just on-page SEO) drives LLM citations are genuinely fresh framings; however, the RAG-vs-training-model distinction and the Google-parallel comparisons are already standard industry talking points, limiting how far the episode pushes into genuinely contrarian territory.

“share a prompt and it's really our version of share of voice for AI surfaces”

“The LLM thought it was a violent thriller. And the, the comp for it was John Wick... the real con for it was actually more of a Marty Supreme”

Justin Inman is a directly relevant practitioner - running a platform purpose-built for AI visibility, sitting on the IAB AI board writing guidelines, and working across entertainment, biotech, and pharma verticals - but he is the founder of what appears to be an early-stage startup with limited disclosed scale, and his claims (98% box office accuracy, 20% LLM hallucination rate in entertainment) go unchallenged and unsourced, suggesting emerging rather than proven authority.

“I said on the IAB AI board we're writing the kind of rules and guidelines for AI visibility”

“we just hit 98% for the latest Super Mario movie”

The episode offers several concrete anchors - the Seer Interactive 83% stat, the 40-to-60-word chunk-size heuristic, two-to-three-week lag estimates by platform, and the John Wick hallucination case - but key performance claims (98% prediction accuracy, 20% entertainment hallucination rate, predicted 'lift' figures) are asserted without sourcing, and broader claims about publisher economics and LLM retraining cycles remain hand-wavy.

“83% of AI citations originate from pages that do not rank in the traditional Google top 10”

“It might take a couple days for ChatGPT or Claude or Gemini to actually have those newly cited kind of sources pop up”

Rex Briggs earns credit for genuinely probing follow-ups on RAG mechanics, vectorisation, and the control/treatment experiment problem, and Greg pushes on proof-of-performance; however, the hosts frequently answer their own questions, let big numerical claims pass unchallenged, spend considerable time on mutual admiration and organisational plugs, and close with a transparent funding-announcement fishing expedition.

“How do you actually know that you've sort of accomplished like what becomes proof of performance and all this?”

“How does Ambrose know that better than somebody else, by the way?”