Getting your business recommended by AI is one job, not eight.
That is the most useful thing I can tell you about marketing right now, and it runs against how most of the industry operates. The standard playbook treats ChatGPT, Perplexity, Claude, Gemini, Grok, Copilot, and Google's AI Overviews and AI Mode as eight separate games: eight algorithms to decode, eight sets of tactics to maintain, eight line items on a proposal. They are not separate games. Every one of those engines is reading from what amounts to the same map of meaning. Define your brand sharply on that map and every engine inherits the definition, including the engines that have not launched yet. Leave it blurry and every engine passes over you for a competitor whose definition is crisp.
I can say this flatly because of a research result that deserves to be far better known: every AI model, whatever its architecture and whatever it was trained on, converges on the same underlying structure of meaning. Rival companies, different data, different designs, one geometry. When the researchers who proved it needed a name for the principle, they did not reach for a computer scientist. They reached for Plato.
I came across this research recently, and it supported, then considerably expanded, the argument I published four months ago in "The Philosophers Already Knew": understanding is not a thing stored in a system. It is a structure, and it lives between the parties in a dialogue rather than inside either one. The conclusion you need is in the paragraphs above. What follows is why it is true, and why the philosophy turns out to be the practical part.
I. The Cave, Properly Told
Everyone has heard of Plato's cave. Fewer people remember how it actually goes.
In Book VII of the Republic, Plato asks us to imagine prisoners chained in a cave since childhood, facing a wall. Behind them burns a fire, and between the fire and the prisoners runs a walkway where people carry objects: statues, figures of animals, tools. The prisoners cannot turn their heads. All they have ever seen are the shadows those objects cast on the wall, and all they have ever heard are echoes. So they do what any intelligence does with the only data it has ever received: they build a science of shadows. They learn the patterns, predict the sequences, honor the prisoner who forecasts the next shadow best.
Then one prisoner is freed. He turns, sees the fire and the carried objects, and understands that everything he took for reality was a projection of something more real. Plato's point was never about caves. It was his Theory of Forms made vivid: the particular things we encounter are imperfect, shifting copies, and behind them stand the Forms, the stable structures that make the copies intelligible at all. Every particular chair is a shadow. The Form of the chair is what all the shadows have in common, and it is more real than any of them.
For twenty-four centuries this was metaphysics, which is to say, unfalsifiable. There was no experiment you could run on the Forms.
In 2024, researchers at MIT noticed something in AI models that needed a name. As models grew larger and more capable, their internal representations of the world were converging, even across different architectures and different training data, as if every model were approximating one underlying representation that none of them had been shown directly. They called it the Platonic Representation Hypothesis, and the cave was the reason. Each model is a prisoner. Its training data is its wall of shadows. No model chooses what it sees, a condition Heidegger called thrownness and I wrote about in April. But the shadows in every cave are cast by the same objects. Train a system deeply enough on its own wall, and what it recovers is not the shadows. It is the structure of the things casting them.
The MIT paper observed the convergence. A Cornell group, in a paper called "Harnessing the Universal Geometry of Embeddings," did something stronger: it used it. And the difference is the difference between believing there is one territory behind all the maps and actually navigating by it.
II. Two Maps of the Same Terrain
When an AI system reads text, it converts the text into what engineers call an embedding: a long list of numbers that places the text as a point in a space of meaning. Meaning becomes location. Texts about similar things sit near each other. Unrelated texts sit far apart. Everything built on AI, search, recommendations, chat assistants, runs on these locations.
Each model builds its own space, and the spaces are mutually unintelligible. The same sentence gets completely different numbers from different models. For years this was treated as proof that each model's "understanding" was private, an artifact of its particular training. The Cornell group showed the opposite. Working with nothing but a pile of raw numbers from one model, no access to the model itself, no examples of how its numbers line up with any other model's, they translated those numbers into a different model's space almost perfectly. Their method learned the shape both models share and used it as the bridge. Their formal statement of what they call the Strong Platonic Representation Hypothesis is a philosophical claim in an engineer's vocabulary: models trained on the same kind of task "converge to a universal latent space such that a translation between their respective representations can be learned without any pairwise correspondence."
Two details tell you how deep the shared shape goes. The translators were trained on general text, then handed embeddings of medical records, and they preserved meaning down to rare diagnoses they had never encountered in any form. The shared geometry is not a list of things the translator memorized. Even unseen concepts have a determinate place in it. And because the researchers are a security lab, they drew the uncomfortable corollary: meaning survives translation so faithfully that the numbers alone leak their contents. From bare vectors, no text, no model, they reconstructed the substance of most of a set of corporate emails. If your company treats its vector database as anonymized, it is not, and that finding alone should reach every CTO. You cannot extract an email from coordinates unless meaning really has a geometry, and unless every model has found the same one.
Two maps, drawn by cartographers who never met, of the same terrain. That is what the convergence of AI models amounts to, and the terrain is human meaning itself. - Drew Thomas Hendricks
Strip away the architectures, the training runs, the corporate rivalries, and every one of these systems has been forced to learn the same shape. The prisoner in every cave, working only from his own wall, draws the same objects.
III. What the Models Are Actually Converging On
The analogy strains in one place, and the strain matters.
Plato's Forms are eternal, mind-independent, and unconditioned. They would exist if humanity had never existed. The universal geometry is none of those things, and the researchers are careful on this point: the convergence holds for systems trained on the same kind of task with the same kind of data. Change what the systems are asked to do and the guarantee dissolves. Whatever the models are converging on, it is not a timeless heaven of meanings.
So ask what it actually is. What is the one thing every one of these models, whoever built it, was given to learn from? Us. Human language, produced across centuries, carrying the accumulated structure of human distinction-making: what we treat as alike, what we oppose, what we group together, what follows from what. The models converge because they are all reconstructing the same object, and the object is the structure of human meaning as it has been deposited in language. The Form these systems approximate is not above us. It is us.
The Form these systems converge on is not above us. It is us: the structure of human meaning, deposited in language, recovered independently by every system that reads deeply enough. - Drew Thomas Hendricks
Readers of the April essay will recognize whose territory this is. A structure of meaning that precedes any individual encounter with it, that is historically sedimented in language, that conditions every act of understanding without belonging to any single mind: that is not Plato's realm of Forms. That is Gadamer's tradition, the shared horizon out of which all understanding happens. Plato was the first to insist that behind the shifting particulars stands a common structure that makes them intelligible, and he earned his place in the hypothesis. But the philosopher whose account fits the actual finding is the one who located that structure not in a heaven of ideals but in the historically accumulated space between minds. The researchers measured that space. It has coordinates now.
IV. What I Was Arguing in 1993
In the spring of 1993, in front of the Gonzaga University Philosophy Club, I argued that Churchland and Sejnowski's phase space, the high-dimensional landscape a neural network navigates as it learns, and Gadamer's fusion of horizons were two descriptions of the same phenomenon. Understanding is a trajectory through a space of possible states, not a content stored at an address. In April I made the argument again with three decades of hindsight and pointed at large language models as the vindication.
Buried in that argument was a prediction. If understanding were storage, then two systems with different architectures, trained on different data, should end up with different and incompatible contents, the way two libraries with different acquisition histories hold different books. If understanding is structural, the systems should converge, because they are not accumulating contents at all. They are mapping the same relational territory, and the territory constrains every accurate map the same way.
That test has now been run at scale. Models sharing no data, no architecture, and no lineage arrive at geometries so nearly identical that meaning passes between them almost losslessly. Storage would have diverged. Structure converged.
The fusion of horizons comes out of this strengthened. Gadamer's claim was that understanding is an event between two parties, produced by the encounter, located in neither. When I wrote in April that understanding "exists between the system and the user, not in either one," I meant it as a philosophical description. The measured, mapped, navigable existence of a common semantic space makes it an engineering fact. The between is not a metaphor. It has a shape, and the shape can be surveyed.
V. One Geometry, One Game: The Mechanism Behind GEO and AEO
At Nimbletoad we measure how often clients get mentioned, cited, and recommended across all eight engines. Our measurement data is what first pushed me off the platform-by-platform view. When a client's standing improved, it improved across engines together. When AI systems got a client wrong, confused them with a competitor, filed them under the wrong category, repeated a fact that stopped being true two years ago, the error appeared across engines too. We could see the correlation in the data long before we could explain it. The universal geometry is the explanation. If every model converges on the same structure of meaning, there are not eight indexes with eight entries for your brand. There is one semantic space, and your brand is a region in it. Every engine is reading from approximately the same map, because there is only one territory for the maps to be maps of.
Your brand is not eight entries in eight indexes. It is one region in one shared geometry of meaning, and every AI engine is reading from approximately the same map. - Drew Thomas Hendricks
This is what generative engine optimization and answer engine optimization actually are. The question that determines your AI visibility is not "how do I rank on this platform." It is: where does your brand sit in the shared geometry, and how sharply is it defined there? Is it a distinct region, tightly bound to the problems you solve, the category you lead, the places you serve? Or is it a diffuse smear that no engine can retrieve with confidence, so every engine reaches past you for a competitor whose region is crisp?
The work follows from the question, and none of it is platform tricks. Consistent naming and entity descriptions everywhere your brand appears, because contradiction blurs the region. Structured data that states facts unambiguously, because schema is how you hand the map-makers your coordinates directly. Content that answers real questions in self-contained, citable form, because citations are how a region earns its associations. Third-party corroboration repeating the same facts, because a region defined by many independent sources is sharper than one defined only by its owner. Every one of these moves is the same act performed through a different instrument: defining your region of semantic space so clearly that any system mapping the territory places you correctly. Do it once, well, and every engine that reads the map inherits it. When the next engine launches, and there will always be a next engine, you are already positioned in the only space it can draw from.
The inverse compounds just as fast. Contradictory descriptions across your web presence, category ambiguity, stale facts sitting in high-authority sources: in a single-geometry world these are not eight small problems. They are one blurred region that every engine reads. Incoherence used to cost you one ranking at a time. Now it costs you everywhere at once.
VI. The Practical Degree
Philosophy has spent fifty years as the punchline of career advice. Every philosophy student has heard the joke about the degree and the coffee shop.
Here is what the degree actually trains. Holding a concept steady while everything around it shifts. Noticing when two vocabularies are describing one phenomenon. Asking what a thing is before asking how to optimize it. Following an argument to conclusions you did not expect and accepting them. Those are not soft skills. In an economy where every business is trying to position itself inside a geometry of meaning that AI systems read, they are the hard skills. Defining what your company is, sharply enough that eight different AI engines place it correctly, is a conceptual analysis problem before it is a technical one. The people trained to do conceptual analysis have been sitting in philosophy departments the whole time.
I wrote in April that the tool is identical and the outputs are not, and that the difference is the horizon of understanding the operator brings to the dialogue. The universal geometry extends that claim from one model to all of them. The discipline behind the results our clients see, 33x click growth, 350,000+ AI citations in a single quarter, visibility that rises across engines together, is not a stack of platform tactics. It is the practice of defining meaning clearly in a space that every engine shares, which is to say, it is applied philosophy with measurement attached. Visibility that moves across engines at once is exactly what a universal geometry predicts and what a platform-by-platform theory cannot explain.
Thirty-three years ago I argued in front of a philosophy club that phase space and the fusion of horizons were the same idea. This year, researchers proving that every AI model converges on one structure of meaning needed a name for the finding, and the name they reached for was a philosopher's. The traditions are not converging anymore. They have converged. What remains is the work: taking what philosophy knows about meaning and using it, deliberately, in the places where meaning now gets read by machines.
The philosophers already knew. Now the instruments agree.