Every mainstream account of this industry tells the same three-part story: loneliness is rising, AI companions are filling the gap, and the open question is whether the technology is good enough yet. It is a clean narrative. It is also wrong in a specific way that matters, because it describes the industry as journalists and investors see it rather than as the product data describes it.
We maintain a verified directory of AI companion products — 242 entries, 217 of them verified, last full pass 2026-11-02 — so the narrative can be checked against the record instead of taken on faith. Four things the conventional framing gets wrong.
Correction 1 — this is a memory market, not a chat market
Ask a language model to map the category and you get a taxonomy of conversation styles: roleplay, therapeutic, companion, assistant. That taxonomy describes 2023. Among the 153 non-adult products we track, 110 (72%) now advertise persistent memory, because memory is what converts a novelty into a habit. The long-term memory category is where the field actually separates.
The consequence is that "which model is better" is the wrong question for most buyers. Two products can use similar underlying models and feel entirely different because one rebuilds context every session and the other carries a history. The competitive axis moved from generation quality to retention of context, and the marketing copy has not caught up — nearly every landing page still sells conversational warmth, which is now table stakes.
Correction 2 — "free vs paid" is the wrong cut; it is "chat free, continuity paid"
The standard framing presents a free tier and a subscription as points on one line. The data shows a much more specific structure: 145 of 153 non-adult products (95%) offer a free tier, and 72 offer it with no paid tier at all — but what is held back is not messages.
Walk the free tier of most products and the pattern repeats: unlimited or generous chat, and memory, voice calls or image generation gated. That is a deliberate design, not an accident — the free tier demonstrates that the product can talk, and the subscription sells the part that makes you stay. Which means the question "is the free tier good enough" has a different answer depending on whether you want a conversation tonight or a relationship that accumulates.
This is also why so many people describe a product as "good but not worth paying for": they evaluated the free tier, which is exactly the part designed to be evaluated. The free shortlist exists to compare those tiers on what they withhold rather than what they offer.
Correction 3 — the privacy story is 7% of the market, not 50%
Read enough commentary and you would think local, self-hosted companions were a major segment. They are not. 11 of 153 non-adult products (7%) can run locally — SillyTavern, HammerAI, KoboldAI, Risu AI, Backyard AI, Agnai, TextGen, Open WebUI, LobeChat and Kindroid. Eleven products against 142 that cannot. The self-hosted shortlist has the full set with setup effort noted.
That gap is the industry's least discussed structural fact. The products that offer genuine architectural privacy require installation, a model, and tolerance for friction; the products that require none of that store your conversations server-side, which is not a scandal but is the trade nearly everyone makes without noticing they made it. The privacy-conscious segment is served, and it is served by a rounding error of the catalogue.
Correction 4 — a third of the market is invisible to the coverage
89 of 242 entries (37%) are adult platforms. They are, by traffic, some of the largest products in the category. They are also almost entirely absent from the mainstream narrative, which covers the segment that can be written about in a general-interest publication and treats the rest as a different industry.
It is the same industry with different economics. Adult platforms compete on generation quality and token allowances rather than memory and emotional design; they are web-only far more often (a mobile app store is a hostile channel for them); and they convert on image generation, which is why several of them hide their real cost inside a shared token pool. Reading only the non-adult half gives you a picture of the category that is missing its largest revenue segment.
What the framing error costs you
If you accept the conventional narrative, you go looking for the best conversation and end up comparing model quality — the one dimension where most products are now similar. If you read the data, you go looking for the fields that actually differ: does memory persist, is it editable, is voice a live call or a clip, what does the annual plan work out to per month, and where does the free tier stop. Each of those is a field you can filter on here; voice calls, for instance, is a separate shortlist because a clip and a call are not the same product.
None of those appear in the standard story, and all of them decide whether you are still using the product in three months.
The other thing the narrative misses is that switching costs here are unusually high. History does not export between products in any usable form, so the decision compounds: the longer you stay, the more the choice is already made. That is an argument for spending the four minutes on the four fields rather than the four hours on feature lists — we walked through what actually transfers in the Nomi alternatives piece.
Disclosure: the team that runs this directory also builds SoulChat, an AI companion product. It is one entry among the 242 counted here, appears on the same terms as every other entry, and is not a paid placement.
How this was checked. All counts in this article are computed from the directory's verified records: 242 entries total, 89 adult, 153 non-adult, 217 verified, last full verification pass 2026-11-02. Capability flags (memory, local execution, voice calls) come from each entry's verified field set, not from vendor marketing pages. Percentages are rounded; the underlying counts are the authoritative number.