What Is an AI Companion App
An AI companion app is a consumer application whose product is a conversational character you resume rather than a task you finish. You talk, it answers in a consistent voice, and something carries over to the next session. That single design goal — resumability — is what separates the category from a general-purpose assistant, and almost every other property these apps share, including the ones that are hard to evaluate, follows from it.
The category has no formal boundary and the marketing vocabulary is loose, so it helps to be precise about what is actually being offered before comparing anything.
The three properties that define the category
An app is in this category if it does all three of these. Missing any one of them and it is something else with a chat window attached.
A stable character. The thing you are talking to has a name, a manner of speaking, and traits that persist across replies. This is not a technical achievement so much as a product decision: the app supplies a standing set of instructions with every request, and that is what keeps the voice consistent. Change those instructions and you have a different character.
Continuity across sessions. You close the app, come back tomorrow, and the conversation resumes rather than restarts. How much continuity, and of what kind, varies enormously between products and is the single most oversold feature in the category — see what “memory” means when an app claims it for how that claim breaks down.
Open-ended interaction as the point. There is no task the conversation is trying to complete. An assistant that has answered your question has succeeded; a companion app that ends a conversation has failed. This inverts the usual measure of a good software interaction, and it is worth noticing that the product’s definition of success is time spent.
What is adjacent but not in the category
Drawing the boundary is more useful than listing features, because the boundary is where the marketing gets fuzzy.
General-purpose assistants can be prompted to behave like a companion, and many people use them that way. They are not in the category because continuity and character are things you have to maintain yourself rather than things the product provides.
Character platforms are a genuine sub-category rather than an exclusion: the app supplies the machinery and other users supply the characters. That changes who is responsible for what you are talking to, which is a substantial difference and not a cosmetic one.
Scripted chat products — dating-sim style apps, choose-your-reply narrative games — look similar in a screenshot and are not generating their replies. If the same input always produces the same output, it is a game.
Therapy and wellbeing apps sometimes use a conversational interface. Where they make clinical claims they are a regulated category with different obligations, and treating them as companion apps understates what they are supposed to be held to.
Voice assistants in hardware share the mechanism and not the purpose.
Why the category is unusually hard to evaluate
This is the part that matters if you are choosing between products, and it is structural rather than anyone’s fault.
The product is a feeling, and feelings are not specifications. Conversational quality cannot be listed on a comparison table in any way that survives contact with actual use. Two apps with identical feature lists can be completely different to talk to, and no amount of reading tells you which is which.
Nothing you can see is stable. The model behind the app, the instructions that shape the character, the limits on how much you can send, and the rules about what it will discuss are all server-side settings that change without an app update and without notice. A description of how an app behaves is a description of one week.
Almost every published comparison is untested. The economics of the category reward publishing lists quickly, and reward affiliate links more than accuracy. Reading those lists usefully is a skill in itself, covered in how to read an AI companion app review.
The switching cost is asymmetric. Getting into one of these apps is free and takes a minute. Getting out means abandoning whatever continuity you have accumulated, because none of it transfers.
What is actually checkable
There is a shorter list of things that are genuinely verifiable, and it is worth knowing that it is short.
Provenance. Who publishes the app, what else they publish, and whether the policy pages belong to the same operator. This is store-listing work and takes about a minute.
The billing channel. Whether you are paying a platform store or the operator directly determines whose cancellation and refund process you get. This is knowable before you pay and matters more than the amount.
What the app asks your device for. Permission requests are observable, they are enumerated by the operating system, and they can be compared against what the stated features actually need.
Whether an export exists. If you can get your history out, the switching cost is lower and the operator’s continued existence matters less.
What the published policy says. Not what a reviewer says it says — the document itself, on the operator’s own domain, which is the only authority on the operator’s own practices.
What this does not tell you
A definition does not rank anything. Knowing precisely what the category is tells you what questions to ask and does not answer any of them, and specifically it does not tell you whether any given product is good, honest, or durable.
It also will not tell you whether one of these apps is a reasonable thing for you to use. That is a question about your circumstances rather than about software, and the useful version of it starts from what you would actually do with the app instead of from which app is best. The mechanism behind all of them is the same in outline, which is set out in how AI companion apps work — and understanding the mechanism is, unhelpfully but honestly, most useful for lowering your expectations of the marketing rather than for choosing between products.