|
Almost Timely News: 🗞️ A Better Mental Model of AI for GEO (2026-05-31) :: View in Browser
 |
The Big Plug
👉 This Friday, come check out a free Trust Insights webinar on competitive GEO.
Full disclosure, it is also a gentle sales pitch for our GEO 201 course, also coming this Friday.
Content Authenticity Statement
100% of this week’s newsletter content was originated by me, the human and reorganized by Claude Opus 4.8 from my voice recordings. Learn why this kind of disclosure is a good idea and might be required for anyone doing business in any capacity with the EU in the near future.
Watch This Newsletter On YouTube 📺
Click here for the video 📺 version of this newsletter on YouTube »
Click here for an MP3 audio 🎧 only version »
What’s On My Mind: A Better Mental Model of AI for GEO
This week, let’s do a different mental model of how AI works. I’ve seen a lot of hot commentary specifically around GEO that is completely disconnected from the reality of how AI models work, and that advice will inevitably lead to dissatisfaction when the desired results just don’t appear.
The same is true for any expectation for AI to behave in completely predictable ways. It can’t; by its very nature it is a probability engine.
Part 1: The Machine — Map, Server, Harness
Part of the challenge for marketers is that AI is full of jargon, and it’s jargon where it’s difficult to even infer what the jargon means - vectorization, embeddings, softmax layers, key value caches, etc. Even the basics, like what a model is and how it’s only part of the AI tools you use, can be confusing. While I often refer to AI models as giant databases of statistics, even this isn’t as accurate as it could be for helping folks understand what a model is.
Let’s start with this: a model is like a map.
Open up a maps app like Apple Maps, Open Street Maps, Google Maps, etc. An AI model is the map itself, and everything on the map has coordinates. In the real world, those coordinates are latitude and longitude, though we often refer to things in relation to other things, like telling people to take a right at the red barn on the corner.
Now, that map needs to be served up. You don’t just access the map itself - there has to be a server somewhere on the internet that serves up the map, AND there has to be an app or a website that allows you to access it.
Map, server, app. Make sense so far?
The map is the AI model. Every AI model has to have a server of some kind. You have to have some kind of app to interface with the server. We’re used to web interfaces, chat interfaces like ChatGPT, Claude, Gemini, etc. But there’s a ton of apps that also connect to models and servers, like Claude Code, Claude Cowork, OpenAI Codex, Google Antigravity, and so many others.
When you start talking super nerd speak like agentic harnesses, just substitute the fancy word for apps. An agentic harness like OpenClaw or Hermes Agent is just another app. It needs to talk to a server, and the server needs to have models available.
We’re extending our analogy of an AI model as a map. A quick word on terms before we go on: the app is sometimes called the interface, but we’ll call it the harness. Harness is a broader term than interface, because interface implies just the user-facing stuff, and harness also has rules built into it about how it interprets and sends and receives things from the model.
Now, what we need to make clear is that when it comes to using AI and getting results out of AI, the app, the harness, is just as important as the model itself. Just like a map application: if the interface is terrible, if it’s got all sorts of weird rules built into it, if it’s censored, and so on, it’s going to give you very different results.
This is critically important to understand for things like GEO. If a harness has rules that interfere with the model, even if the model gives you one response, it may not be the response that the user actually sees. All AI models have some basic training, the helpful, harmless, honest trifecta that Anthropic and others have posited for years about how to align models.
But on top of that, model makers and AI providers have created these harnesses that enforce rules more deterministically, or have guard models that silently inspect and reject responses that go outside the guardrails that model makers have made. For good or ill, those additional guardrails cause further interference in our ability to predict what AI is going to respond with.
Our capability of predicting AI responses is predicated on not just what’s in the model but how it is served and then how the harness interprets those responses. For example, a server sends and receives information from the model, right? The server enables the model. It has settings like top-p and top-k, temperature, repetition penalty. These are all technical adjustments that the server operator makes to serve up the model.
Probably the most famous harness right now is Claude Code. Claude Code is a massive harness that talks to Anthropic’s AI. It talks to their servers and it talks to their models through their servers. However, what many folks have recognized and realized is that the Claude Code harness is really terrific. It’s full of useful rules for constraining how a model behaves. The closer a model behaves to how Anthropic’s Claude models behave, the more useful Claude Code is. That means if you use it with a model similar in capabilities to Claude’s models - Opus, Sonnet, and Haiku - you will get similar results.
This is why you can use Claude Code with a model like Minimax or Alibaba Qwen and get fantastic results. Yet if you were to use that model natively in an interface that does not have a good harness around it, one that has fewer of the battle-tested rules, you will notice a difference pretty quickly, because it doesn’t have all of the guardrails that Anthropic has baked into Claude Code.
This is also why brands show up in different tools in different ways. For example, ChatGPT is a harness around OpenAI’s servers and OpenAI’s models. Its search mechanisms operate differently, where Claude and Gemini often try to discern, based on log probabilities - which is essentially a confidence measure - whether or not they need to invoke search. OpenAI tends to read the prompt up front and decide from the prompt whether it needs to do web searches, or whether it can answer the question from its own knowledge.
Every tool does things differently; every brand shows up differently in every tool. As underlying models change, so does what shows up in the server and in the harness. For example, in May of 2026, Google released Gemini 3.5 Flash. This model replaced Gemini 3 Flash, which is used in Google AI Mode and Google AI Overviews. Even though the harness, which is web search, did not substantially change, the underlying model did. The way it does things like its query fan-out changed dramatically. Gemini 3.5 Flash gathers more sources up front and has fewer iteration loops to get to the answer. This makes it faster, and it means the consideration set is larger.
For a marketer, that in turn means that if you’re trying to be recommended in the results, all other things being equal, you have a higher chance of being in the consideration set, because the consideration set is wider up front. However, it also means that the model, if it doesn’t know about you, is less likely to stumble upon you in subsequent iterations because they don’t exist.
Part 2: The Map of Language — Concepts Are Cities
In plain language, embeddings and vectorization become a concept’s coordinates because language models are built by ingesting huge amounts of information, huge amounts of data, trillions and trillions of words. This training occurs in several phases, but the non-technical explanation is that every word, and the words around it, are all predictable.
For example, if I say “I’m spilling the tea,” those words have meaning not only because of the words themselves, but the order that they are in. If I say “tea, the spilling I’m,” this makes absolutely no sense at all. When we convert words into these statistics that form the map, if you will, in a generative AI model, it is just as important to understand how different words relate to each other.
A lot of people are asking GEO to be deterministic. They want to know who’s the number one brand, et cetera, in AI responses. When we go back to o