A small subset of Netflix viewers just got a new way to find movies and TV shows. The streamer recently started to test an AI-powered native voice search feature with some subscribers, prompting them to press the Netflix button on their remote control to ask for viewing recommendations.

Once viewers press the button, they’re presented with a few search suggestions, including phrases like “I need a good cry,” “watch in the background,” and “help me stay awake.” Each of these suggestions leads to a set of viewing recommendations, but there’s also an “Ask” button with a waveform icon. Select it, and you’ll start an AI-powered voice search that delivers viewing recommendations in response to natural language prompts.

I’ve had access to this new voice search feature for a few days now, and found it remarkable for two reasons: While still in beta, it’s impressive in its ability to serve up appropriate and interesting viewing recommendations to even the most esoteric requests. It also squarely circumvents the voice assistants and search features built into smart TVs and streaming devices, highlighting the power struggles between TV OS platforms and the streaming services running on them.

Netflix is currently testing its new voice search feature with select members in the U.S. I got access to the feature on a Chromecast with Google TV streaming dongle and a TCL Google TV, but wasn’t able to access it on Roku or Fire TV devices. Even users who have multiple devices running the same smart TV OS have reported that they were only able to access it on a subset of those devices.

Playing with the feature for a few days, I was impressed by its ability to find relevant picks for a wide variety of requests, ranging from the rather obvious (“date night movies from the ’80s” served up The Breakfast Club) to the very specific (“I like the music of Brian Eno, what should I watch” resulted in a suggestion for Abstract: The Art of Design).

I stress-tested the feature more than people usually would during day-to-day use, but it was able to make solid recommendations in response to unusual requests. I honestly thought asking for “fun kids TV shows about death” wouldn’t get me any results. Instead, it served up A Series of Unfortunate Events and Raising Dion, which both were on point.

“I had too much coffee today, what should I watch” resulted in Netflix suggesting a “laid-back” Sheng Wang comedy special and the Headspace Guide to Sleep. When I simply said “hurt people hurt people,” the Netflix app responded with “Hurt people do hurt people. Here are some of the stories behind that,” followed by recommendations for 13 Reasons Why and It’s Okay to Not Be Okay.

Interestingly, Netflix’s voice search doesn’t appear to be tapping into the service’s personalization engine right now. When you ask for recommendations based on what you’ve recently watched, the app’s response is “We can’t answer that one yet, but we’re working on it!” (Speaking of which: All responses are shown as written text. Netflix’s voice search doesn’t use any voice output of its own, which I didn’t mind at all.)

Responses are often accompanied by suggestions to further narrow your results, including “more unhinged” and “more bittersweet.” However, narrowing down recommendations this way doesn’t always work. A search for Cantonese comedies yielded many relevant results. But once I followed Netflix’s suggestion to narrow those results by pressing the “more quirky” button, the streamer suddenly tried to sell me BoJack Horseman as a “Cantonese comedy with a quirky twist.”

There are some things that can trip up Netflix’s voice search. When I asked for Blaxploitation movies, the app’s voice recognition system instead heard me ask for “Black exploitation movies,” and declined to provide me with any results. Which, yeah, probably the right choice.

A request for “TV shows about porn” was also declined, but Netflix did happily serve up plenty of suggestions when I requested “steamy TV shows.” That’s notably different from Google’s Gemini assistant on Google TV, which informed me that it could not fulfill that request, adding passive-aggressively: “I can help you find appropriate shows if you’d like.” That alone shows why a service like Netflix might want to have its own voice search instead of relying on the one provided by your TV or streaming device maker.

Another major reason for Netflix to build its own voice search: The streamer wants viewers to remain in its own app, and exclusively recommends Netflix content. Smart TV OS platform operators like Google, Roku, and Amazon on the other hand would like consumers to use their own, universal search that also recommends results from their own services and services from partners they can monetize.

This power struggle between platforms and publishers has been going on for years, and has also long extended into voice search. For consumers, the results have been confusing: When you press your remote’s mic button while using Hulu or Disney Plus, you’ll most likely access the universal search function built into your TV or streaming device, with results from the app you’re in intermixed with those of other publishers. Do the same while browsing Netflix or YouTube, and you’re only searching the catalogs of those respective apps.

Thanks to their market share, YouTube and Netflix have so much power that they have been able to demand voice routing privileges not available to smaller publishers. Both companies initially just used that privilege for speech-to-text functionality, essentially feeding your voice queries into the same search fields you can also access with your remote to search for a show by name.

With the emergence of LLMs, both services now extend those privileges for more full-featured voice assistance. YouTube recently launched its own conversational AI tools, allowing viewers to access Gemini-powered voice features on devices made by Amazon, Roku, and others.

Just like YouTube’s implementation, Netflix’s take on AI voice search shows that consumers can benefit if apps have access to the mic. It just shouldn’t take massive market share — or a branded button on your remote — for publishers to get to build such experiences.