Ask YouTube: guides built from video

What works

• Ask YouTube is a named mode with its own landing, not a bolt-on search bar tweak. Dive deep into any topic sets scope before the first ask.
• Starter chips span jobs: learn a skill, plan a trip, compare gear, find a documentary. The grid teaches what belongs here better than a blank composer.
• The disclaimer sits above the composer where people see it before they send. Standard AI hygiene without burying the input.

What we would push on

• Nine equal-weight chips is a lot of scanning. Group by intent or rotate examples so repeat visitors see fresh jobs.
• No visible history or back affordance beyond the browser. A learning session may span days. Where do prior guides live?

Product bet

YouTube is testing whether people will ask questions instead of typing keywords. Starters sell depth over scroll.

Takeaway

Name the mode, set scope in one line, and ship starters that mirror real jobs on the platform. Do not open on a generic chat box.

What works

• Ask YouTube and All as a persistent toggle keeps the host product honest. You can bail to classic search without leaving YouTube.
• The user query renders as a right-aligned pill. The answer gets a guide title immediately, so the page shape is clear before prose arrives.
• Skeleton bars under the title signal synthesis in progress. People know to wait for a guide, not a video grid.

What we would push on

• Loading state shows title plus skeleton only. No partial sections or cited videos yet. Long waits feel empty.
• Follow-up composer appears before the first answer lands. That is fine for continuity, but New search should be harder to hit by accident mid-load.

Tradeoff

Takeaway

When the output is a guide, show the guide frame immediately and keep a visible path back to classic search.

What works

• The answer is a reading experience: title, intro, section heading, then evidence. That is the chat-artifact move for video.
• Bolded fold names in prose teach vocabulary before the clips. Text leads, video proves.
• Cited videos use native card chrome: thumbnail, title, channel, views. Citations feel like YouTube, not a chat footnote.
• Ask a follow-up and New search share one floating bar. Continuation and reset stay in thumb reach.

What we would push on

• Two side-by-side cards from the same channel with similar titles invite duplicate watching. Curate for contrast, not repetition.
• Intro copy can run long before the first clip. Learning guides should surface a video early for people who think in motion.

Product bet

YouTube would rather synthesize a guide from its catalog than rank ten thumbnails. The durable output is an article with receipts.

Takeaway

For video platforms, treat guides as first-class artifacts. Embed native video cards as citations inside the prose, not as a list at the end.

What works

• A comparison row with Difficulty and Key technique columns turns search into a buying guide. People pick a path without opening three tabs.
• Section copy frames the row: once you have the folds, try these models. Structure matches how learners actually progress.
• Three abreast with aligned metadata rows scan faster than stacked results. Comparison is the feature.

Takeaway

When you recommend videos, add compare metadata in a scannable grid. Tell people why these three belong in one row.

What works

• Thumbs down opens a modal with labeled chips: Inaccurate, Biased, Irrelevant, Too long. Fast taxonomy beats a blank textarea.
• Optional detail field plus privacy copy sets expectations about what gets sent. Standard for Google-family AI surfaces.
• Feedback attaches to the synthesized guide, not an individual video card. Quality loops target the model output.

What we would push on

• Chip set is generic chat feedback. Learning guides might need Wrong steps, Missing video, or Bad citation.
• No visible confirmation after Share feedback. People need a receipt that the signal landed.

Takeaway

Collect structured feedback on synthesized answers. Tune the chip set to the artifact type, not only generic chat complaints.