Growing Search's AI answer: a projected +2.4% to its revenue, more education queries answered

Yandex · Search · Neuro-Alice AI answer · Product Design Lead · 2025–present

Summary

Context

Neuro-Alice, the AI answer at the top of Yandex Search (~150M MAU). I design it as a product.

Problem

Neuro needed more revenue, showed up on less than half of education queries, and its answers needed a next step: a place, their sources, a full conversation.

What I did

Turned the third follow-up suggestion into a search, which opens a results page with Search's ads
Rebuilt homework answers as condition → solution → answer, with four learning features on top
Designed a geo card for place answers in the Alice AI chat (awaiting results)
Designed the quick answer's next step: it unfolds on Search and carries on in the Alice AI chat with its sources (sources awaiting results; this version not launched)

Result

+2.4% of Neuro revenue, projected to full traffic. Neuro coverage of education queries +0.6 pp. All four learning features are live.

Situation

Neuro-Alice is the AI answer at the top of Yandex Search (~150M MAU), and I design it as a product. My Neuro Overview, living documentation of every Neuro answer type, its core scenarios and flows, is the team's single source of truth. The answer had to earn more, show up on more education queries, and take people further than one answer.

Goal: More revenue, more education queries answered in a way students can learn from, and a next step after the answer: a place, its sources, a full conversation.

Action

The third suggestion becomes a search

Why. Neuro needed another way to earn. Its answer ended in follow-up suggestions. A suggestion is already something people tap, and a search opens a results page that carries Search's ads. So the third suggestion became a search, and the others stayed as they were.

Animation of a Neuro answer in Yandex mobile Search to the query headphones for flights: three follow-ups grow under the answer, two questions with a chat icon and a search, buy noise-cancelling headphones, with a magnifier; the search is tapped and its results page opens beside the answer, two ads on top and the first organic result under them

Result. In a one-week A/B test (Dec 5–11, 2025, 2% sample), revenue from the results pages the search chip opened, scaled to full traffic, came to +2.4% of Neuro's total revenue.

Homework answers rebuilt as condition → solution → answer

Why. Neuro showed up on less than half of education queries (47.7%). A homework answer only helps if the student can follow it from the task to the result.

I designed this format, the solution restorer.

Animation of a Neuro homework answer in Yandex mobile Search to how many integers are between 3/4 and 13/4: it builds block by block, Condition, Solution turning the fractions into 0.75 and 3.25, the Answer hidden under a spoiler of drifting particles, then three follow-up questions and a Refine your question field

Result. In an A/B test (Nov 21–30, 2025, 1% sample), Neuro's coverage of education queries rose +0.6 pp (47.7% → 48.3%). The test compared several versions of the restorer; the best one raised no-click wins, searches the answer settled on its own: +6.15% per answer block and +5.79% per results page, on the queries the test targeted.

Learning features for homework answers

Why. A homework answer can do more than hand over a result to copy: it can help the student learn.

With the PM, I designed four features for it.

Animation: under the solved equation x² = 5, a Try learning mode card with a beta badge appears: Explains the solution step by step — Learning mode: offers a step-by-step explanation · Hidden answer: under a particle spoiler until tapped · Graphs: the answer draws the figure · Textbook: restores the task's condition

Animation: the answer to how many integers lie between 0.75 and 3.25 is covered with a spoiler of drifting, twinkling particles; a tap scatters them and reveals 3: the integers 1, 2 and 3 — Learning mode: offers a step-by-step explanation · Hidden answer: under a particle spoiler until tapped · Graphs: the answer draws the figure · Textbook: restores the task's condition

Result. All four are live.

A geo card for places in the Alice AI chat

Why. A question about a place is also a question about where it is and what to do next: book a table, get there, call.

I designed the geo card that follows Alice AI's answer about a place in the chat. A carousel opens with a map tile, the place's pin and logo, then its photos. Under it: the name, category and price chips, the rating with the number of ratings, the address and hours, a Book button and round buttons for the site, the route and a call. Call opens the place's phone numbers: a sheet on mobile, a tooltip on desktop.

Animation of the Alice AI chat on a phone: to Recommend a restaurant? Alice answers with a short description and a geo card for a restaurant, a map tile with its pin and the place's photos in a carousel, the name, category and price chips, rating, address and hours, a Book button and site, route and call buttons; the carousel swipes from the map to the photos, the call button opens a sheet with two phone numbers, and Close takes it away

Status. In an experiment; results aren't in yet.

The quick answer's next step: into the chat, sources and all

Why. The quick answer from Alice AI is live above the links, and it suits a simple fact. A harder question needs a full answer you can keep talking to without leaving Search: the step Google took from AI Overviews to AI Mode. And the move into the chat shouldn't lose the answer's sources.

I designed that next step as one continuous move. "Show more" unfolds the full answer in place: sections, sources inline and in a Sites panel, follow-up suggestions and an "Ask anything" composer. A follow-up carries on in the Alice AI chat without a new page: the old answer slides away, the tab switches from Search to Alice AI, the question becomes the first message, and the Sites panel stays where it is and loads the new answer's sources. The Search tab takes you back.

Animation of desktop Yandex Search for how to choose a car: the quick answer from Alice AI with a gallery of cars and a Sites panel fades above a divider with a red Show more; Show more unfolds the full answer in place, with sections, source chips, three follow-up questions and an Ask anything composer; the follow-up Which engine is more reliable? is clicked and the page becomes the Alice AI chat without a reload: the old answer slides up and away, the tab pill slides from search to Alice AI, the question lands as a bubble, a chat rail slides in, and the Sites panel stays in place, shimmers while loading and fills with the new answer's sources as the answer streams in; then the search tab takes it back to the results page

Status. The quick answer itself is live. Sources following the question into the chat: results aren't in yet. This version of the flow is designed and hasn't launched.

Result

+2.4% of Neuro revenue, projected to full traffic
+0.6 pp Neuro coverage of education queries (47.7% → 48.3%)
+6.15% searches the answer settled on its own, per answer block, on targeted queries (best version)
4 live learning features for homework questions

What I'd do differently

Measure what the search chip costs as well as what it earns: whether people who get it ask fewer follow-up questions.
Run the revenue test longer and on more traffic. One week on 2% gives a direction, not a settled number.

Assistant

Ask me about Tim

Tamirlan Askar

Lead Product Designer · 7+ years