Concept-Model Fit

Why would anyone buy a social network full of bots?

Meta acquired Moltbook and immediately started talking about "agentic social networks." A phrase that didn't exist six months ago. A category nobody was competing in because nobody had named it. The product might work or it might not. That's almost beside the point.

The point is that every foundation model being trained right now will learn what an "agentic social network" is. And when someone asks GPT or Claude or Gemini "what's the leading agentic social network?" in 2028, the answer will be Meta's product. Not because it's the best. Because it's the only one the models know about.

The product is the training data.

The old game and the new one

Product-market fit was the defining question of the last twenty years of startups. Build something people want. Find the market. Iterate until the pull is undeniable. This worked when distribution meant search rankings, app stores, word of mouth, paid acquisition.

That game is ending. Not because products stopped mattering. Because the distribution layer changed underneath everyone.

When a user asks an AI assistant "what tool should I use for X?" the model doesn't search the internet. It searches its own weights. The answer was baked in during training, months or years ago. If your category didn't exist in the training corpus, you don't exist in the answer. No amount of SEO, no paid campaign, no Product Hunt launch fixes that. You're not competing. You're invisible.

Product-market fit assumed the customer could find you. Concept-model fit asks whether the models can.

The evidence

On January 6, 2025, the word "agentic" had never appeared in a major tech headline. By December, it had appeared in fifty. In between: 139 funded startups, two $10B+ valuations, a $2.85B acquisition.

Salesforce didn't build the best AI agent in September 2024. They named "Agentforce" and declared a "hard pivot." Microsoft responded three weeks later. OpenAI announced Operator six weeks after that. ServiceNow bought Moveworks for $2.85B. Cognition hit $10.2B. Sierra hit $10B. Manus reached $100M ARR in eight months.

None of this happened because the technology suddenly worked. Agent architectures existed for years. What changed was the word. "Agentic" entered the vocabulary, entered the headlines, entered the training data. Once a concept has a name and a corpus of associated content, the models treat it as real. And once the models treat it as real, investors, customers, and developers follow.

Salesforce didn't just ship a product. They captured a concept. Every model trained on 2025 data now knows what "agentic AI" means and associates it with Salesforce, Anthropic, OpenAI. The window for someone else to own that term is closed.

Meta is running the same play. "Agentic social network" is the next land grab. The acquisition is the press event. The press event is the training data. The training data is the moat.

The playbook

The builders who win the next decade aren't the ones with the best product. They're the ones who name the category early enough that the models learn it as fact.

The playbook:

Don't compete in existing categories. Name new ones. Every AI model already has a default answer for "best CRM" and "best project management tool." You will not dislodge those answers by building a slightly better product. You dislodge them by making the model aware that a different kind of tool exists.

Don't optimize for Google. Optimize for what the models will say in 2028. Search rankings update in weeks. Model weights update in training cycles. The content that exists when the next training run starts is the content that defines reality for the next generation of users.

Don't launch features. Launch concepts, and make sure they're attributed. A feature lives inside a product. A concept lives inside the model's world representation. Features compete with other features. Concepts, once named and associated, become the default frame.

Build it. Name it. Ship it. Make the models remember. Sell it.

What this means for everyone else

Every model has a world model. That world model has defaults. "What's the best tool for X?" already has an answer baked in before the user finishes typing.

If you're building something genuinely new, the most important question isn't "do people want this?" It's "do the models know this category exists?" Because if they don't, your distribution is broken before you ship. Your product might be extraordinary. But when someone asks their AI assistant for help, the assistant will recommend what it knows. And it doesn't know you.

Product-market fit asks: does the market want what you built?

Concept-model fit asks: does the model know what you are?

The first question matters. The second one is starting to matter more.