Google Discover Chatbot Feed Customization Experiment

Google may be getting ready to let you talk back and control the Discover feed, borrowing a concept from an ongoing YouTube experiment that allows users to conversationally fine-tune recommendations by chatting with an AI. Clues within the Google app data indicate that a chat-driven customization flow is in the works, indicating a deeper integration of Gemini across Google’s consumer-facing products.

What the clues reveal about Discover chat customization

Text strings in the Google app 16.49.59 have surfaced a dialog called “Discard conversation” that warns “any changes you’ve discussed for your feed will be lost” when resetting Discover settings. Curiously, the visible input box is nowhere to be found—a prime indicator that this feature is half-finished and being controlled on the server side. Clicking “Customize your space” reportedly causes a subtle glow animation in Google’s colors, suggesting that a new model of interaction is being wired up.

It’s the language of “conversation” that gives it away. Here’s how you do it today: not by interacting with the app via a dialogue interface, but by hiding stories and following topics. It has long had straightforward knobs for that kind of thing—follow topics, hide stories, mark your interests—but not this dialogue paradigm. Describing changes as something you “talk” about suggests an AI broker that understands natural-language requests such as “more women’s football coverage” or “less day-trading content.”

A page from YouTube’s playbook on conversational tuning

YouTube has also been evaluating a chatbot that allows people to steer the content on their homepage through prompts. Here the logic is clear: recommendations steer behavior, and conversation reduces friction. YouTube has said in the past that recommendations drive more than 70% of watch time, meaning that even small wins for preference signaling could have a huge impact on engagement. Applying that filter to Discover — Google’s evergreen interest-based feed — might be a more specific, less tedious alternative to scouring settings.

If Discover takes a similar approach, you might request “fewer transfer rumors” or “deep dives from long-form tech outlets” instead of being forced to switch off entire topics, or block specific sources altogether. The nuance is in how a model dissects intent and weights it against historical signals from Search history, location, app activity, followed topics.

Why this matters for Discover’s future personalization

Discover lives to the left of your Google app on most Android phones (as well as iOS), and is a prominent pane on many device home screens. Even minor UX changes to Discover are widely deployed, given that Android’s active device base is north of 3 billion users. Discover is a potent, if at times unpredictable, referral channel for publishers. Clearer, more conversational controls may also help users fine-tune the trade-off between breadth versus depth — possibly helping stabilize the content mix they receive and even easing swings in exposure that are based on interest.

For Google, a chat interface dovetails with its larger Gemini roadmap that posits natural language as the control layer for discovery and productivity. And it gives you a structured way to capture high-intent feedback (“show me more investigative climate reporting”) that’s richer than a binary “not interested.”

How it could work in the Google Discover experience

Though we’re not seeing the UI here, an outline based on breadcrumbs is something like this:

• Tap a new “Customize your space” entry in Discover.
• A chatbot features some speedy prompts (“More of this,” “Fewer short posts,” “Prioritize local teams”).
• Users write natural-language requests; the model proposes changes and presents a summary of what will change (topics, sources, frequency).
• Settings save, with the choice to return or “discard conversation,” like the string surfaced in the app.

Look for Gemini to fuel the interpretation layer, but with Discover’s ranking system still calling the shots at final placement. Google has a history of gradually enabling such features server-side, experimenting with A/B tests.

Impact on users and publishers if chat controls roll out

Used well, conversational controls could cut accidental over-filtering — in which one “not interested” click silences an entire topic. Users would be able to choose “less speculative headlines but maintain official team updates,” hinting at a balance of diversity with less noise. For publishers, more fine-grained preference-setting might help favor outlets tailored for niche but persistent interests, instead of ones that are optimized solely for viral flares.

The other side of the coin is transparency: users will want to see an explanation for what their prompts affect, how long it affects them, and how they interact with current activity signals. Underneath, summaries or visual chips showing what adjustments are currently active would prevent the “set it and forget it” issue where your feeds end up wandering off course.

Privacy and control questions raised by chat-based tuning

Conversational tuning leads to questions that are already familiar: Does one use prompts to train models? Do they share space with account history? Is processing on-device or cloud-based? With profiling and consent being a hot topic for regulators, assume explicit controls to view, edit or delete your chat-based preferences — and opt-ins made apparent. Google has been focusing on account-level controls for Web & App Activity and ad personalization; Discover’s chatbot should fit into that framework with uncomplicated ways to reset.

What to watch next as Discover chatbot tests expand

There’s no visible input box, so this is a bit early, but the conversational wording tends to be significant. Look for limited tests in the Google app (in response to server-side experiments) and parallel rollouts of YouTube’s chatbot experiments. Not every experiment ships (for Google or any tech company), but where this one is going seems all but inevitable: if you think about it, natural language starts to look like the interface for curation as much as search.