Why a chat box is the wrong front door for a health product

An open chat box is the wrong front door for a health product. By front door I mean the first screen, the primary way in, the thing a new user meets before they've learned anything about how your product thinks. A blank prompt in that spot assumes a person who already knows what to ask. Most people who arrive at a health product are in the other situation: someone newly prescribed a GLP-1 drug, or just told their fasting glucose is creeping up, who has a worry rather than a query. The friendly line above the cursor says ask me anything, and that is exactly the problem. It takes the hardest job in the exchange, working out the right question, and hands it to the person least equipped to do it, on the day they can least afford to get it wrong.

In brief

• An open chat box hands the hardest job — working out the right question — to the user least equipped to do it, on the day they can least afford to get it wrong.
• ChatGPT Health makes the chat box look like the settled default, but feeling natural and being usable are two different things.
• A better front door guides the question: it shows people where to start, and keeps free text for stating intent rather than carrying the whole interaction.

But 230 million people ask ChatGPT about their health every week. Why wouldn't that work for us?

Because it looks settled, and settled things stop getting questioned. When OpenAI launched ChatGPT Health in January 2026, it put a medically-reviewed health assistant inside the app that already fields roughly 230 million health questions a week. What that scale means for your product is a separate question. Here the point is narrower: that reach makes the open chat box look like the settled, obvious choice.

So the chat box arrives pre-approved. If the most-used AI product on earth answers a health question with an open prompt, and hundreds of millions of people reach for it that way, copying that interface feels less like a decision than like reading the room.

This is the part founders get wrong. They copy ChatGPT's front door because its naturalness reads as proof that it is what users want. But naturalness and ease are two different things, and the gap between them is where the cost hides.

Why the blank box is harder than it looks

The cost has a name now. A 2026 paper, "The Keyhole Effect" (arXiv:2602.00947), argues that chat is the wrong default for any task with more than one step that builds on the last: you end up reading a whole situation through the keyhole of a single message, holding everything you've already seen in your head because each new reply pushes the previous one out of view. Chat is good at two things, stating what you want and reading back a result, and weak at the exploring and comparing in the middle.

Put a worried, non-expert person in front of that and it gets worse, not better. The work that pits chat against a plain menu of tappable options finds chat leaves people with less sense of control and more mental effort, even when it feels more natural; the numbers are in the companion piece on designing the question. Someone confident and well is fine typing into a void. Someone anxious and newly diagnosed, unsure whether her question is even a reasonable one to ask, is exactly who the blank box punishes hardest, and she is a large share of who a health product serves.

What does a better front door do instead?

It guides the question. Instead of opening on a blank box and waiting, a better front door meets the user with a few legible ways in, shaped around what people in her situation usually come to ask. It turns ask me anything into here is where people like you tend to start, and it lets her recognise her question rather than compose it from nothing. The chat can still be there. Natural language is good at stating intent, so keep it for that. What you build around it is structure: something to look at, choose from, and point at.

There is a name for assembling that structure on the fly: generative UI, where instead of returning a wall of text the system builds interface to fit the question, say a chart or a set of comparison cards, rendered as the answer forms. Vercel open-sourced a library for it this year called json-render; a company called Thesys ships an API, C1, that returns interactive components instead of prose. That machinery is worth its own discussion.

The design move underneath the machinery comes first, and it is smaller than it sounds: decide what your product asks the user before you decide how it answers. That is the subject of a companion piece, on designing the question, not the answer.

The blank box only looks generous. An invitation to ask anything is really a request that the user supply the right question, and a health product makes that request of the people least able to meet it, on the day they can least afford to get it wrong. So before you design how your product answers, design what it asks. For the person who opens your app not yet knowing what to say, that first question is the whole front door. Build it for her.