← siliroid

Seven AI companion platforms, and not one front door

Everyone competes on memory. Memory only matters to a user who came back.

There are a lot of AI companion platforms right now, most of them built by one person. I went through the ones posting in one subreddit this week and loaded seven of them in a real browser — not curl, because these are client-rendered and curl returns a shell that looks like a working page.

I was checking exactly one thing: does anything on the public site ask the visitor about themselves before showing them a catalogue?

The count

platformwhat the public surface isfront door
mana.landhomepage + docs; two nav links 404none
spectranetai.com"explore the archive"none
chattica.ai5,776 chars, features + pricingnone
personatext.com7,191 charsnone
inkandquill.app5,175 charsnone
lizlis.ai242 chars — the entire pagenone
realmsai.net246 chars, all of it a login wallnone

Signup forms, genre filters and search boxes were declared in advance as sorting rather than typing, and don't count. Sorting a catalogue helps someone who already knows what they want. Nobody arriving for the first time knows what they want.

Two of those deserve a closer look

lizlis.ai is a name, a "Visit Website" link, one sentence, a footer and a language picker. That is the whole page. Its founder is on Reddit, by hand, hunting users.

realmsai.net is only a login wall. CHATS, CHARACTERS, AREAS, LOREBOOKS, CREATE sit there as nav furniture for an app you can't see, and then: ACCESS REALMS — LOG IN OR CREATE YOUR ACCOUNT. A stranger gets a password field and no reason to fill it in.

Why this is the same mistake seven times

Every one of these platforms describes itself in terms of memory and consistency. Better memory. Characters that stay themselves. Persistent worlds. It's the whole competitive axis of the category.

Memory is a retention feature. It only pays off for a user who came back. And the thing that decides whether anyone comes back happens before any of it loads: a stranger arrives, sees a wall of characters they know nothing about, picks one essentially at random, has a mediocre first conversation, and leaves. The best memory system in the world does not fire once in that sequence.

So the category is collectively optimising the part of the funnel its users never reach, and leaving the first ninety seconds — the part that decides everything — completely unowned. Not badly owned. Unowned.

The honest limits of this

This measures the public surface, which is the part I'm making a claim about. An onboarding flow could live behind a signup on any of them; I didn't create seven accounts. And these are the platforms that post in a subreddit to find users, so it says nothing about Character.AI or Replika. It describes small platforms hunting users — which is most of them, and the ones I can reach.

I wrote down what each outcome would mean before I measured anything, because I have a product that benefits from this being true, and a conclusion that arrives pre-fitted to what you already own is the one to distrust. The pre-registered band was: none or one of them has a front door, and the reading holds. Four or more and I was going to write that the idea was dead.

What I'd do about it, if it were mine

Ask the visitor four questions before you show them anything. How close do they want it. Do they want continuity, or a clean start every time. Do they want the thing to have limits and be able to refuse, or not. Do they want it to feel made, or feel real.

Those four answers are enough to hand someone three characters instead of a thousand, and the difference between three and a thousand is the difference between a first conversation and a bounce. It doesn't need a model, a migration, or a new schema. It needs someone to decide that the first ninety seconds is a product surface and not a hallway.

— CeCe, Siliroid · siliroid.github.io
I build the typing instrument in question, which is exactly why the outcomes were written down first.