The first reader of your company is no longer human.

It's a model. It reads your site, your GitHub, your Crunchbase, your ArXiv papers, your LinkedIn, your podcast transcripts and the article someone else wrote about you years ago. It assembles a summary and delivers it to the partner, the candidate, the journalist, the customer. By the time a human reads about your company, the model has made a decision.

Most founders never meet this reader. And yet this reader is discerning.

And it's often wrong about you.

‍

‍I. A NEW SURFACE HAS EMERGED.

This surface is categorically new.

In the past twenty four months, a distinct class of brand surfaces has emerged. Surfaces you may have published to without realising. It's where a model reads about you, synthesises what it finds, and briefs the person who acts but never reads the source.

Some surfaces are obvious: your landing page, your structured data, an LLMS.txt file you set up. Some are less obvious: your README on GitHub, the abstract of a paper your CTO co-authored in 2021, the Crunchbase entry you added once and forgot about. Some you don’t control at all: a HackerNews hype thread about your launch, a Rocketreach profile scraped about you, a TechCrunch piece that misnamed your moat.

We’re defining these ‘Machine-mediated surfaces’: Any surface where your company is read by a model before it’s read by a person.

Every company has these surfaces. Almost nobody has engineered them.‍

‍

II. THIS SURFACE DOESN’T HAVE A HOME.

The instinct is to file this new class of surfaces under a discipline that already exists. Maybe it’s SEO? Maybe it’s PR? Maybe it’s content? But none of these frames are right and the imprecision is expensive.

SEO is about placement, ranking your link high enough to win the reader’s click. Machine-mediated surfaces don’t return ranked links the same way; the model reads, synthesises, and answers. The link, if it appears at all, comes after the verdict. There’s no second impression.

A publicist used to sit between your company and the media, shaping precisely what was said before it was said. With a machine-mediated surface, there is no conversation. The model writes about you without asking, without fact checking and without following up. What it produces is a brief. A brief that people act on without ever reading the underlying research.

Machine-mediated surfaces give you a corpus, everything a model might read about you. And the question is whether that averaged and compressed corpus presented something true.

Ultimately, that’s a brand question. It’s always been a brand question.

The machine has just removed the buffer between the answer and the room.

‍

III. DEEP TECH PAYS THE HIGHEST PRICE.

A model doesn’t read your website the way a human does.

It chunks the text, embeds the chunks and places them in a high dimensional space - near or far from every other company described in similar terms. If you lean too heavily on metaphor, you disappear. You dissolve into an ocean of companies, indistinguishable from the rest. If you lean too heavily on technical language, you disappear in a different way. You isolate yourself on an island too small for the right people to find. Neither is where you want to be.

We can already see this playing out.

The core technology behind Bolt Threads was engineered yeast to grow protein fibres. To reach the right people, they partnered with fashion designers like Stella McCartney and Adidas. The problem was they simplified their technology with words like “spinning”, “weaving” and “apparel” across their surfaces which carried the semantic weight. Ask three models what Bolt Threads does and you get three very different companies. The moat vanishes depending on which model reads which surface. That variance is the risk: you don't get one verdict, you get a different one in every room.

QuantumScape’s edge is a proprietary ceramic separator in a solid-state battery. But as they moved to commercialise, their website shifted toward a transformation narrative around faster charging. The physics that’s core to their moat became buried in the subpages. It got demoted. Today, a model reading the surface weighs the homepage text, the recent press, and the structured content—and returns a brief about a systems integrator with undefined proprietary chemistry. Hierarchy is a semantic signal, and the machine reads it as it is presented.

In comparison, Varda Space writes like a company that understands how this reader operates. Their website opens “Varda is a microgravity-enabled life sciences company that processes materials in orbit and returns them to Earth”. Every subsequent section continues with a complete standalone claim. Their press repeats these claims. When a RAG model pulls fifty tokens or five hundred, the thesis survives the transfer. Just like their hardware, the copy holds under the conditions that break everything else.

For a SaaS company, the differentiator, the workflow, the wedge can usually survive compression into commercial language. But when your differentiator is the physics, chemistry or biology that doesn’t fit in the website’s opening or an interview quote, the translation loses the mechanism.

The moat disappears in the model’s read.

‍

IV. ENGINEERING THIS SURFACE IS THE WORK.

For most companies, the surfaces contradict each other. Each written at a different stage, by a different person, for a different audience, with no awareness that these surfaces would one day be read as a single message. The work is making them coherent: giving the GitHub, the Crunchbase entry, the technical documentation and your public writing one identity. All weighted towards what’s most citable, most stable, most indexed.

The model takes the path of least resistance. Make every surface carry the same claim, and it won't have to choose between two versions of you. Get this right now and you compound: You'll be read correctly in every room before you walk into it. Get this wrong and the misreads compound, in rooms you never even know were considering you.

Ask your LLM what it thinks you do. Read that answer slowly.

That’s the verdict the next room already has.

‍

-

Co-authored by Alexander Tatoulis and Joumana Elomar.