Researcher, Magician, Assistant

I was on a run last week thinking about why some AI features compound and others vanish. Came back with three buckets.¶

Researcher. You ask, it answers. Magician. It does a trick. Assistant. It does the thing before you ask.¶

All three are legitimate. Only one of them builds retention.¶

The big platforms are combining all three at once. ChatGPT runs Researcher through chat, Magician through image gen, Assistant through memory and personalization. Apple Intelligence does the same shape: Writing Tools as Researcher, Image Playground as Magician, Smart Notifications as Assistant. Google. Microsoft. Whoever’s next. The platforms are running the three-way combo because they can afford to.¶

Most other teams pick one. Most pick wrong.¶

Researcher is the lane for information work. Deep research, knowledge retrieval, sense-making across documents.¶

It’s also the lane I’m most skeptical of as a destination product.¶

For most users, traditional search already maps cleanly to how they think. Type the keywords, scan results, click the link, read the page. The mental model is decades old and it works. It’s hard to justify replacing it with a conversation when someone just wants to find a thing. Conversation is friction for transactional intent. Most queries are transactional intent.¶

Where Researcher genuinely wins is the small slice of users doing hours-long research. Comparing primary sources. Synthesizing across documents. Pulling threads through dense material that a human reader would burn an afternoon on. Real lane. AI saves real time. But it’s not most people.¶

The other place Researcher shows up is the “I don’t want to read” version. WebMD-style. Symptom in, answer out. AI can technically solve that. But responsible product teams should ask if it’s a problem we should be solving in the first place. Sometimes the friction is the work.¶

In most products, Researcher is a feature, not a destination. And for smaller teams, Researcher products are structurally hard to make retentive. You’re competing with the frontier model platforms for the same query market. They own the models. They own the distribution. They own the verb. Most startup Researcher products spend their first 18 months building the thing OpenAI is going to ship for free in the next API release. It’s an expensive way to learn what your real product was supposed to be.¶

Magician is the demo lane. Image gen. Voice clones. Avatars. Virtual try-on. “Watch what the model can do.”¶

Magician is photographically perfect. The trick happens, the audience claps, the screenshot gets shared, the launch hits TechCrunch. Then the user logs off. The trick was the value. There’s nothing underneath.¶

Most consumer AI launches right now are Magicians. Features designed for the launch-day demo. Features that win Product Hunt. Features that look like the future. None of them retain. At consumer-AI margins, retention IS the unit economics. Magicians don’t have a retention story. They have a launch story.¶

Siegfried and Roy is the right metaphor for this lane. Vegas spectacle. Tigers, lights, smoke. Built the most successful magic show in the world. Sold out the Mirage for thirteen years. Got the prime-time TV specials and the screenshots-shared-everywhere energy. And then the retention curve crashed. The thing they built only worked while they were doing it. Nobody else could replicate the act. There was no flywheel. The show was the product. When the show ended, the product ended.¶

That’s the Magician pattern. Whatever happens on stage is the whole thing. There is no underneath.¶

Assistant is contextual and proactive. It knows what you were already doing and gets there first.¶

The smart compose finishing your sentence in your voice. The IDE autocomplete that knows your codebase. The calendar that catches the conflict before you do. The notification that surfaces the right thing at the right moment. The product that quietly removes a step you didn’t realize was a step.¶

Assistant is the lane that builds a moat. Every personalization signal makes the next interaction better than the last. The context infrastructure compounds. The data improves the product, which improves the data, which improves the product again. That’s a real defensibility curve, and it’s the only one of the three buckets that has one.¶

The Assistant lane is also harder. It requires real context infrastructure, real personalization signals, a real opinion about what users were going to do anyway. It doesn’t demo as well. The launch-day screenshot of an Assistant feature is boring. The retention curve six months later is the whole point.¶

This is also where the clarifying question lives. Assistants don’t surface five hundred options. They surface the right one, at the right moment, because they actually know the user. The interface narrows on the user’s behalf. That’s the service.¶

Most consumer AI right now is still optimizing for Magician. Features that win launch day. Demos that retweet well. “Look what we built” energy.¶

I get the pull. Magicians get the cover stories. Assistants get the long compounding curve nobody sees from outside. The fundraising deck wants the Magician. The retention curve wants the Assistant. Most teams optimize for the deck and then wonder why the curve didn’t follow.¶

But the long compounding curve is the only one that builds a company.¶

Magic is what gets you in the door. Assistance is what keeps you there.¶

Less magic. More assistance.¶