How buyers find a new model, and where yours is missed

Four doors, and each one reads the same few facts

Buyers find a new model through four doors: a web search, an assistant's answer, a directory of the stack, and a price table that lines up token rates. Each of them can only work with facts that are stated somewhere, chiefly the price per million tokens, the context length, the license and the release date. A model whose page does not state those plainly can be dropped when the buyer filters, long before anyone reads what it is good at.

That is the uncomfortable part for a lab with a strong model and a thin page. Quality is hard to filter on, because it is the thing a buyer cannot read off a page. A filter works on what can be read, and testing comes afterwards, on whatever survives.

Where the filter actually happens

A buyer with a real job to do has constraints before they have preferences. The context has to hold their documents. The license has to allow their product. The price has to fit a budget somebody else set. Each of those is a yes or no, and each is decided by a field, not by a demo.

Search engines and assistants face the same limit, as far as anyone outside them can reason. To answer a question like which models take long documents, they need a page that states a context length in tokens, next to the model's name, in a sentence. A figure that only appears inside an image, a tab or a playground setting is hard to quote, and a model whose figure cannot be quoted is easy to leave out of the answer.

Directories are blunter still. A row with an empty context field does not sort, and a filter for a license type cannot include a model whose license is a link to a long PDF with no summary line.

What a directory row needs from you

The models layer holds 144 live products today, and the large language model category is where a buyer narrows that down. Every fact on a row there carries the page it was read from and the date it was read. A field we could not find is shown as no answer on record, which is different from no.

That distinction cuts both ways. It protects a maker from being marked as lacking something it has, and it also means a buyer scanning a column sees a gap beside your name and a value beside someone else's.

Categories matter as much as fields. A model listed only under a broad heading competes with everything in it, while a buyer who has already narrowed to speech or image work looks in a narrower category and never sees it. Check that your model sits where a buyer with its job would look, and that each fact on the row points to a page you control rather than a reseller's summary of it.

Price tables only list what has a price

Price tables are the narrowest of the four doors. Our token price table covers 467 distinct models from 16 providers, and it is built from rates sellers publish per million input and output tokens, so a model needs a seller with a public price page to appear there.

Models priced only by quote, or only inside a console, are invisible to every comparison of that kind. That may be a deliberate sales choice, and it is a legitimate one. It should be made knowing that the buyer who compares by cost will never see the model at all, rather than see it and decide it is dear.

A model sold by several hosts has a second problem. Each host may publish its own name for it, its own context limit and its own price, and a buyer comparing them is reading the host's page, not yours. The maker cannot set those prices, but it can publish the facts the hosts should be copying, so that a buyer who spots a disagreement has a source to settle it.

Assistants repeat sentences, not impressions

When an assistant names models for a task, what it can repeat is a statement: a name, a number, a unit and, ideally, a source. Nobody outside the labs that build those assistants knows how sources are weighed, and we do not claim to. What can be said is that a sentence like the model accepts 128,000 tokens of input can be quoted, and a hero banner that says massive context cannot.

Launch posts are a common gap. They are written for the day of release, full of charts and adjectives, and they can be the only page that mentions the model by its exact name. A plain model page, kept current, gives a search or an assistant something to point at after the launch post has scrolled off.

A check to run on your own page this week

Open your model page in a private window, signed out. Find the price per million input tokens and per million output tokens, the context length in tokens, the license in one sentence, the release date, and the exact string a buyer would pass to the API. If any of those takes a click, a scroll into a footnote, or a login, that is the fact a buyer is missing.

Then search for the model by name and see which page comes back first. If it is a reseller's page, or a third party's summary, that is the page buyers are reading your facts from. Claiming your listing here is free and is how a maker corrects a value we hold or adds one only it knows.

Repeat the check whenever the model changes. A point release that moves the context length or the price makes the old page wrong in exactly the fields buyers filter on, and the old figure is the one that keeps getting quoted until somebody replaces it.