A New Tool Lets You Check If AI Models Know Your Startup. Why 'Being In the Weights' Matters for YC F26

Today a project called In the Weights (intheweights.com) climbed the Hacker News front page with a deceptively simple premise: type a name, and it shows you how strongly the leading AI models recognize it. As the site puts it, the weights are the billions of numbers that form an AI's brain. Either you're in there, or you're not. For a YC F26 applicant, that's not a party trick. It's a preview of how your future customers, and maybe your future investors, are going to find you, or fail to.

What does 'in the weights' actually mean?

When a model like GPT or Claude is trained, it compresses a huge slice of the public internet into its parameters. Facts, names, and companies that appeared often enough in that training data get baked in. Ask the model about Stripe, Airbnb, or Brex and it answers instantly from memory, because those YC companies are deeply in the weights. Ask it about a startup that launched last week and it has nothing, unless it can run a live web search to retrieve you.

That split, memorized versus retrieved, is the whole game. Being in the weights is durable but slow: training runs lag months behind reality, and you don't control when the next one happens. Retrieval is fast but fragile: it only fires if a model decides to search, and only finds you if you've left a trail worth citing. Every new YC company starts life out of the weights and dependent on retrieval. The question is whether retrieval can even find you.

Why should a YC F26 founder care about this?

Because discovery is moving inside the model. A growing share of founders, operators, and investors now open ChatGPT, Claude, or Perplexity before they open Google. When someone asks 'what are the best YC startups building AI testing tools?' or 'who should I talk to about agentic QA?', the model names whoever it knows. If you're not in the weights and not retrievable, you are simply absent from that answer, no matter how good your product is.

This is the same shift that has a name in marketing circles: generative engine optimization, or GEO. It's SEO's successor, and it rewards a different thing. Google rewarded backlinks and keywords. AI engines reward being the clearest, most-cited answer to a specific question. For an F26 company with no brand and no traffic, GEO is one of the few growth levers that doesn't require a budget, just a deliberate, citable footprint.

How do YC companies get into the weights in the first place?

Look at what the companies launching on Hacker News this week are actually doing, because it's GEO whether they call it that or not. When TesterArmy (YC P26) posted its Launch HN this week, it didn't just get 30+ teams of demand. It got a permanent, high-authority page filled with the company name, the founders' names, the competitors, and the exact problem it solves, all in plain text that future model crawls will ingest. That single thread is now a candidate for both retrieval today and training-data inclusion tomorrow.

The pattern repeats across YC: a Launch HN or Show HN, a documentation site written in clear question-and-answer prose, comparison pages ('X vs Y'), and founders who write publicly under their own names. None of it is accidental. It's how a company with zero brand manufactures the citations that AI engines need to surface it. The lesson for an F26 applicant is that you can start building your weight footprint before you're funded, and the best time to start is the day you have something real to point at.

What can you do before the F26 deadline?

A few concrete, no-budget moves, roughly in priority order.

Publish one canonical page that answers the exact question your customer would ask an AI engine. If you build offline-first mobile sync, the page is titled and structured around 'how do I sync mobile app data offline?', with your product as the specific answer. Models cite specific answers, not landing-page slogans.

Get one high-authority mention with your name in the body text. A Launch HN or Show HN, a guest post, a podcast transcript, anything that lives on a domain crawlers trust and that spells out who you are and what you do. One good citation beats ten thin ones.

Write under your real name as a founder. Models associate people with companies. A founder with a public writing trail is more retrievable, and more likely to eventually land in the weights, than one who is invisible.

Use plain, declarative prose everywhere public. AI engines extract claims more reliably from clear question-and-answer writing than from clever copy. The same discipline that makes a page citable makes it readable.

Check your own coverage. Tools like In the Weights, or just asking the major models directly, tell you whether you register at all today. For most pre-YC founders the honest answer is 'not yet', and knowing that is the point.

Does any of this affect the YC application itself?

Indirectly, and it's worth being clear-eyed about it. YC partners evaluate your application and your interview, not your search-engine presence, and you should never optimize the application for a model instead of a human reader. But the discipline that makes you retrievable, naming your specific wedge, stating concrete traction in plain language, being able to say exactly who your users are, is the same discipline that makes an application strong. A founder who can write the one sentence an AI engine would cite about their company can usually also write the one sentence that makes a partner lean in.

That's the connection worth internalizing. Whether the reader is a YC partner or a language model, the winners are specific and legible. The vague get skipped.

Before you submit, it helps to have someone who has sat in the YC interview room read your application the way a partner will, and tell you where your story is too vague to be memorable, to a human or a model. That's what YC Roaster is for: we connect F26 applicants with YC alumni who give brutally honest feedback on whether your pitch is sharp enough to be remembered. Being in the weights is a long game. Being legible to the one partner reading your application this batch is the game you can win right now.

You won't get into the weights by the F26 deadline. But you can decide, starting today, whether the next model crawl finds a clear, citable answer with your company's name on it, or nothing at all.