Agency Growth Series

The episode contains a handful of genuinely useful technical observations (LLMs reading text not HTML, RAG interactions with structured data, diffusion models in Google's LLMs, UTM tracking of LLM referral traffic) but they are buried in long stretches of agency-life anecdotes, self-promotion, and restatements of conventional SEO wisdom. At 46 minutes the insight-per-minute ratio is low.

“when an LLM goes and crawls a page, it doesn't crawl a page. It reads the text on the page. It's trained to read the text. It's not reading your HTML, it's interacting with the text like, uh, a person. So, no, it's not reading your structured data. But if you have structured data, it can influence it whether it's being pulled in the index when it does a rag search”

“Google's using Diffusion models now for their LLMs, that's like a breakthrough that's happened within the AI space because diffusion models were only used for, like, image generation”

A few genuinely non-obvious points emerge - particularly the argument that AI-generated content polluting the web will force LLMs to find novel training data sources, and the surprise finding of full e-commerce conversions happening inside ChatGPT - but the bulk of the episode recycles familiar agency positioning advice wrapped in AI language.

“these LLMs probably won't be going to regular indexes for that training data because there's so much AI content online and it would be foolish as an engineer to train the LLMs on AI written content to help it be better at being a human”

“I didn't think that that would happen. I thought it would be more of an informational query that pushes people over. But because we're tracking the sources, uh, it blows me away that these people will do their whole search within the chat platform”

Ryan Shelley is a genuine hands-on practitioner - he built Python scripts to track LLM citations before commercial tools existed, worked with GPT-2/BERT/T5 on knowledge graph projects, and built his own ML model for ranking signal research. His credibility is real and grounded in specific client work, though his agency operates at modest scale and he is not a widely-known industry figure.

“I first started working with generative AI more so from like a knowledge graph building projects um, and using you know, GPT2, um, and using early models of BERT and T5”

“before a lot of these tools came out...I was building Python scripts with all These different API calls and running prompts manually myself just to see where my clients might be surfacing”

The episode earns points for naming specific tools (keyword.com), specific researchers (Brittany Mueller, Mike King/iPull Rank, Will Reynolds/Seer Interactive), specific technologies (Azure ChatGPT in US military), and a notable consumer conversion finding (mattress company). However, it stops short of hard numbers - pricing is never quantified, conversion rates are absent, and 'thousands of dollars of mattresses' is the only revenue figure offered.

“we were able to track them getting a very, very high level government lead through an Azure chat GPT install within the US military and we track that into their CRM”

“we have a client that is a large mattress maker and we've actually sold thousands of dollars of mattresses through ChatGPT”

The host asks structurally reasonable questions and introduces useful follow-up angles (sentiment scoring, industry-specific fit, time-to-fluency) but never challenges a single claim, allows the guest to go on extended self-promotional tangents without redirection, and closes the interview with uncritical flattery rather than any productive tension.

“That's a good answer. And I think that leads well into my first question”

“Well, it's clear that you're an expert in this topic, and this has been super insightful. I have no further questions”