I asked my kitchen speaker a genuinely follow-up question recently — something building on what I'd just asked it, the way you'd naturally continue a conversation with a person — and it actually handled it, which would have failed outright on the same device two years ago. That kind of upgrade is real, and it's worth being specific about exactly what changed and what didn't.

The clearest improvement across Alexa, Siri, and Google's assistant is conversational understanding: all three have moved from rigid, keyword-triggered commands toward genuine back-and-forth, where a follow-up question that references something you asked thirty seconds earlier actually works, instead of requiring you to restate the entire request from scratch every single time.

The upgrade that's gone further than most people realize is the assistant taking multi-step action rather than just answering a question — booking a specific restaurant reservation, rescheduling a calendar event around a conflict, or composing and sending a text message based on a spoken description of what it should say, not just looking up a fact and reading it back.

What hasn't fully caught up is reliability on exactly that multi-step action category — a request spanning several apps or several steps still fails or does the wrong thing meaningfully more often than a simple single-step request like setting a timer, which is worth knowing before trusting a voice assistant with something that actually matters, like a real reservation or a payment.

The practical read: the conversational layer is genuinely much better and safe to lean on for everyday back-and-forth, while the action-taking layer is real progress but still worth double-checking for anything with a real consequence if it gets it wrong — the gap between "understood what I asked" and "did exactly the right thing" hasn't fully closed yet.