From Chatbot to Companion: How Fast the Category Has Moved

How fast is "fast," actually?

I get asked some version of "has this really changed that quickly" a lot, so I want to answer it with actual comparison points rather than just asserting it. The clearest evidence is video generation: it went from being a feature that essentially didn't exist across this category to 22% adoption across the 129 platforms I currently track, in a genuinely short window. That's a much faster adoption curve than most consumer software features experience, and it happened specifically because it rode on top of infrastructure (image generation, and the broader generative AI wave) that was already maturing at the same time.

I've written a separate, more foundational piece on the full history of AI girlfriend apps, from scripted software through today's large language model era, which lays out the three broad phases this category passed through. This piece is narrower: it's specifically about how compressed the timeline between those phases actually was, and what that speed itself tells us about where the category is headed.

The compressed timeline, phase by phase

The earliest "virtual girlfriend" software relied on keyword matching and decision trees, a genuinely limited technology with a hard ceiling on how varied a conversation could ever be. That gave way to a second generation with more flexible chatbot-style conversation, still short of true generative AI. The current, third phase, built on large language models capable of genuinely dynamic conversation, is where essentially all 129 platforms I track sit today.

What makes the pace of this category notable isn't just that those three phases happened. It's how tightly compressed the transition into the third phase was relative to how long the earlier phases lasted. Once large language models became capable and affordable enough to run at consumer software scale, the shift into today's dynamic, LLM-driven companion app happened quickly across the industry, not gradually company by company.

What actually changed in each phase, beyond just "got better"

It's worth being specific about what actually improved at each transition, rather than treating "AI got better" as a sufficient explanation. Moving from phase one to phase two meant conversations stopped feeling entirely scripted, gaining real, if limited, flexibility. Moving from phase two to phase three meant conversations became genuinely generative for the first time, with every message producing fresh, contextually informed text rather than matching against a fixed set of pre-written responses.

That third shift is also what made voice, image, and eventually video generation technically feasible as add-on features at all, since all three depend on the same broader wave of generative AI capability that made dynamic conversation possible in the first place. That's a big part of why this category's feature set expanded so quickly once the core conversational engine caught up: the same underlying technology wave unlocked several features nearly simultaneously, rather than one at a time over a much longer period. Our best AI girlfriend ranking tracks exactly how each platform has kept up with this pace, feature by feature.

How this compares to how other consumer tech categories usually move

I think it's fair to say this category has moved faster than most comparable consumer software categories did at an equivalent stage. A lot of that comes down to distribution: smartphone-native apps can reach a mass audience essentially overnight compared to earlier consumer technology that required dedicated hardware or installed software, and AI girlfriend apps benefited from that same distribution advantage that reshaped a lot of software categories over the past couple of decades.

The other accelerant is falling AI inference costs, which made it economically viable to offer broad free tiers and iterate quickly on new features without the kind of prohibitive per-user cost that would have made a slower, more conservative rollout the only sensible option in an earlier era of AI technology.

What hasn't kept pace, and why that matters more than the speed itself

Here's the part I think gets lost in "look how fast this moved" commentary: speed has been wildly uneven across features. Memory and voice are the clearest laggards. Only 21% of platforms document real cross-session memory today, and voice interaction averages just 1.81 out of 5, the weakest score of any category I measure. Both are much harder engineering problems than generating a photo or a video clip, and the pace of this category's overall growth hasn't changed that underlying difficulty at all.

I think that unevenness is actually the more important story than the headline speed. A category that moves fast on easy problems and slowly on hard ones isn't uniformly "ahead," it's specifically ahead on visuals and behind on the substance that determines whether a companion relationship actually feels continuous over time.

Why moving this fast cuts both ways

Fast growth also explains this category's real instability. In a single re-audit pass, at least 23 platforms, about 18%, went dark, got sold, or quietly rebranded within just one year. A market that expands this quickly, to 129 competing platforms, inevitably includes a lot of companies that won't survive the shakeout, which is part of the cost of how fast this category scaled in the first place.

I think that's a reasonable trade to have made as an industry, faster iteration and access clearly benefited users overall, but it's also a real reason to stay a little skeptical of any single platform's long-term durability, and to recheck a review's "last tested" date more often than you would in a slower-moving software category.

A side effect of moving fast that doesn't get discussed enough

One underdiscussed side effect of this pace is how much it's shortened the effective "shelf life" of any single review or comparison in this category. A platform comparison written even a year ago, in a category moving this quickly, can already be meaningfully out of date on pricing, features, or even whether the platform still exists at all. That's part of why I think a review's last-tested date matters more here than it would in a slower-moving software category, and it's a direct consequence of how compressed this category's growth has actually been.

It also means the skills required to evaluate a platform well have had to evolve alongside the category itself. Testing an AI girlfriend app three years ago mostly meant judging chat quality. Testing one today means judging chat, voice, image generation, video, memory, and pricing structure all at once, simply because the feature set itself expanded this quickly in a short window. That's a real shift in what a thorough review even has to cover now compared to what it needed to cover earlier in this category's history.

What this means going forward

My honest read is that the pace of this category isn't going to slow down on its own, but the areas where speed shows up next are likely to shift. Chat, image, and now video have already had their fast-growth moment. I'd expect memory and voice to be next in line for a similar leap, once the specific engineering problems underneath them get solved, rather than staying permanently behind the way they've been through this category's first few years. I cover where I think that leap is most likely to come from in my broader look at where the AI girlfriend industry is actually headed, and you can read more about how I test and score every platform or my background as a researcher in this space.

Further reading

• A Short History of AI Girlfriend Apps: From Chatbots to Companions
• The Future of AI Girlfriends: What's Next for Virtual Relationships?
• The best AI girlfriend apps, ranked