Candy AI Product Review
Hi everyone, today I want to walk you through Candy AI and share a product review of what it is actually doing under the hood.
What caught my attention wasn’t branding or media buzz, but the growth curve itself — steep, sustained, and unusually stable, as if it was quietly pushing against an upper limit.
It has already crossed 20M+ visits, with roughly 5M incremental growth recently. Not bad at all.
What’s interesting is how quiet this growth is. You don’t really see it mentioned in mainstream tech media, and it rarely shows up in the usual a16z-style narratives. Yet on the user side, both scale and stickiness are clearly operating at a different level.
So I decided to spend some time actually using the product end-to-end.
Here’s the short version: Candy AI is essentially an AI virtual companion platform. You can build a long-term conversational relationship with a virtual girlfriend or boyfriend, depending on preference.
It is also very open with NSFW content. That’s not something to ignore—it’s one of the reasons it was able to scale quickly.
But the more I used it, the less I felt that label alone explained what’s actually going on. Reducing Candy AI to “just NSFW” feels like an easy answer that misses the real product logic.
The Core Idea: Relationship Progress as a Multimodal System
The design makes one thing very explicit: your relationship with each AI character has progression.
That progression is tracked, updated, and most importantly, expressed through multimodal outputs—especially video.
Text, voice, images, and video are all tied to the same underlying relationship state.
I started thinking of this as a kind of “multimodal progression system.”
Why It Works: Turning Emotion Into a Structured Loop
The key innovation is not content—it’s structure.
Emotional relationships, which are usually fuzzy and subjective, are turned into something that is:
- visible
- progressive
- unlock-based
- and monetizable
And the feedback mechanism is not text—it’s video.
The business performance reflects this: Candy AI launched in September 2023, scaled to millions of users within a year, and reportedly reached over $25M ARR in 2024.
That implies hundreds of thousands of paying subscribers with strong retention, largely without heavy external funding.
The First Impression: NSFW Warning, Then a “Live Platform” Feel
The first thing you see is a large NSFW warning. The entire interface feels closer to a live streaming platform than a traditional AI app.
The character library is extensive and heavily female-dominated, with surprisingly polished visual quality—faces, lighting, and styling are all relatively refined.
A small but important detail: hovering over a character card triggers motion playback directly on the card itself—short AI-generated video clips instead of static previews.
Discover Feed: Designed Like TikTok, Not a Catalog
The “Discover” section doesn’t behave like a list of profiles.
Instead, it’s a vertical, TikTok-style feed. You swipe, and the next character appears in motion.
This matters more than it looks. In AI companion products, layout decisions directly shape perceived intent. The UI is not neutral—it guides emotional framing.
Chat UI: The Shift Away From “Chat-Centric Design”
The chat interface is split into two main areas. But the right side dominates visually.
Nearly half the screen is dedicated to something called “Live Action.”
This is essentially a multimodal action layer: video reactions, animated responses, and contextual visual outputs tied to the conversation.
In practice, chatting becomes secondary. The experience starts feeling more like a light interactive game where conversation is just one trigger among many.
Unlock System: Locked UI States and Progressive Intimacy
The progression system is extremely explicit.
Early on, many actions are locked. Buttons are visible but unclickable.
Once certain conversational conditions are met, those actions unlock. When you click the playback button, the left panel immediately starts playing a video response.
No transition, no delay—just direct playback of multimodal content.
Two small but important details:
- The videos include audio, not silent loops.
- They can be replayed without re-triggering the interaction flow.
Generation Layer: Prompt-Driven but UI-Guided
Inside chat, users also get direct access to image and video generation tools.
A “Spicy” toggle increases intensity and adjusts the generated prompts automatically.
Each selection is converted into a structured prompt and inserted into the generation field, though users can still modify it manually before execution.
Monetization Loop: Partial Output, Then Unlock
Generated images are initially blurred or partially masked. You can see composition and posture, but not full detail.
At that point, the system offers a clear choice:
Either continue interacting to increase relationship depth, or pay to instantly unlock the full-resolution content.
Character-Level Variation: Not a Shared System
One subtle but important design choice: not all characters share the same behavior system.
Each AI persona has its own progression path, content style, and behavioral thresholds.
Some characters unlock faster, some feel more restrained, others lean more aggressively into expressive outputs.
Collections: Passive Memory of Generated Content
Everything generated during interaction is automatically stored in a “Collection” module.
This acts as a persistent memory layer, reinforcing continuity and ownership over generated outputs.
Character Creation: Visual First, Text Second
The character creation flow is also consistent with the rest of the product philosophy.
Instead of starting with long text prompts, it begins visually.
The first decision is simple: realistic or anime style.
From there, users select attributes like age range and ethnicity, followed by physical features such as body shape and proportions.
Only after that do more traditional fields like personality and occupation appear.
What Candy AI Actually Built
When you connect all of this—the feed, chat, unlock system, generation tools, and character creation—you see a consistent logic.
Candy AI is not just a chatbot. It is a system for constructing emotional progression through multimodal feedback loops.
Text, images, and video are not separate features. They are synchronized layers of a single relationship state.
Why This Matters Beyond NSFW
If you strip away the NSFW layer entirely, the underlying structure still works.
The real innovation is not content—it’s the system design: turning emotional progression into something visible, unlockable, and economically structured.
This pattern will likely extend far beyond adult products into games, virtual characters, AI NPCs, and mainstream companion systems.
Final Thought
We often treat emotion as something unstructured and uncomputable.
Candy AI challenges that assumption by treating emotional relationships as a system that can be designed.
This is not the final form. It is an early but functioning example of a much larger direction.
And what we’re seeing now is probably just the beginning.