Site visibility in Alice AI: inside Yandex Webmaster's new report

Yandex just shipped something Google Search Console still doesn’t have: a native AI visibility report. It’s called ‘Site visibility in Alice AI’ - Alice AI being Yandex’s generative answer layer, its analog of Google AI Overviews. This is the report the whole GEO conversation has been missing. Until now, every claim of ‘neural nets cite our site’ rested on manual checks - type a query, read the answer, log it in a spreadsheet. Now the search engine shows it itself, weekly and across the whole mass of queries.

I opened the report for a cleaning company site I run marketing for. Below - what’s inside, what numbers came out, and what follows from them if you have a site of your own.

What the report is and where it lives

Webmaster → ‘Performance’ → ‘Site visibility in Alice AI’. Works for a verified site; data updates weekly and is built on queries where the site has impressions in Yandex search results.

Three blocks inside:

• A verdict badge. Ours: ‘Your site is top-5 by mentions - for queries where it ranks high in Search’.
• Share of Voice over time. The share of queries where the site made it into the source list of an Alice answer.
• Examples. Concrete queries and source lineups, with a ‘View in answer’ button and an XLS export. There’s a toggle for ‘queries with my site / without’, and a separate ‘Sites’ tab - who else lives in the sources for your topic.

The numbers on a live site

Period April 2 - June 28, 2026. By the weeks that have data:

From 48% to 90% in a month. Here it’s important to be precise. I didn’t press any magic ‘optimize for Alice’ button - there isn’t one. But ‘it grew by itself’ would be a lie too. I had been working on machine citability long before this report existed: Schema markup, llms.txt, factual service pages with prices, FAQ blocks, IndexNow on every deploy. That month was just the regular cycle - service pages getting their content tightened, the blog getting new posts, technical debt getting cleaned up. The report simply showed the result of all that work in numbers for the first time.

Some weeks in the chart are empty - the tool is new, data arrives in chunks. That’s normal, but you can’t lean on a single data point.

Which queries get cited

The examples show 20 queries where the site sits in an answer’s sources. The lineup is telling (all example queries are Russian; I’m translating them here):

• B2B clusters: ‘industrial enterprise cleaning’, ‘manufacturing facility cleaning’, ‘high-rise cleaning and washing of industrial workshops’, ‘industrial rope-access cleaning’, ‘night cleaning of stores’, ‘underground parking cleaning’, ‘office cleaning for enterprises’.
• Foodservice: ‘restaurant kitchen cleaning’, ‘professional restaurant cleaning’, ‘deep cleaning of restaurants’.
• Narrow services: ‘office ceiling cleaning’, ‘cleaning of ceilings and light fixtures’, ‘washing in an underground parking garage’, ‘everyday cleaning’.
• Generic and brand: ‘cleaning’, ‘cleaning company’, the company name in various spellings - including a mangled one with a typo. Alice fuses the typo with the brand and still puts the site as the first source.

And the main observation. The pages that make it into the sources are the very same programmatic service pages that rank in the regular top - with the price right in the title: ‘Industrial cleaning in St. Petersburg from 130 RUB/m²’, ‘Parking and garage cleaning from 30 RUB/m²’, ‘High-rise work from 150 RUB/m²’. Not a single ‘special page for neural nets’ was needed, because those don’t exist. How these 194 pages got into the index - I broke that down separately.

Who else Alice cites: the ‘Sites’ tab

The ‘Sites’ tab shows the top 10 domains Alice pulls into sources for your topic. In our ten - seven cleaning company sites and three aggregators: Avito, Yandex Services and Zoon (the big Russian classifieds and services marketplaces).

That’s an easy detail to miss: in AI answers you compete not only with your industry peers but with aggregator platforms. They have mass content across every service at once, and Alice gladly takes them as sources. You can outrank an aggregator in regular search on a specific query - with a deep page built for the intent. The same logic applies to answer sources.

Where we’re absent: 22 gap queries

The ‘queries without my site’ toggle is the most practical part of the report. We have 22 queries there, and they fall into two clean clusters (RU queries, translated):

• Deliberately not our profile: ‘three-room apartment cleaning’, ‘country house and cottage cleaning’, ‘how much does home cleaning cost’ - B2C cleaning we never built pages for. Here the gap is not a problem but a consequence of positioning.
• Adjacent to our services: ‘industrial premises cleaning’, ‘warehouse cleaning’, ‘post-construction cleaning’, ‘cleaning after fire’, ‘commercial cleaning’. That is a ready-made topic list: for each one you can see who Alice cites instead of us and which page we are missing.

The report effectively sorts your semantics into ‘unavailable by strategy’ and ‘missed by neglect’. Doing that by hand would take a day.

Where Alice pulls its sources from

Yandex says it in plain text in the report’s help docs: the probability of landing in the source list is higher if the page ranks high in Search for the given or adjacent queries and contains quality, useful, expert content.

Translation: Alice’s generative answer is built on top of the regular results. No rankings - no citations. So every course selling ‘GEO optimization for Alice in 2 weeks’ can be safely skipped: first the search top, then the mentions. The order of sources in the answer is random, by the way - it’s the Webmaster interface that puts you first, not Alice.

The same mechanics, incidentally, apply to Copilot and the Bing index - I compared it on numbers: 40 citations for the site that’s in the index, 0 for the site that isn’t.

Honest caveats

So the 90% figure doesn’t create excess heroism:

1. The calculation base is queries where the site is already visible in the results. SoV 90% does not mean ‘Alice recommends us in 90% of the niche’s queries’. It means: where we already rank high, Alice almost always takes us into the sources. Tautological, but that’s exactly what’s being measured.
2. Answers are unstable. Yandex itself warns: for the same queries there may be no answer tomorrow, or the sources will be different. The report is about a trend, not a guarantee.
3. You can’t count traffic from it. The report shows mentions, not clicks. How many people read Alice’s answer and clicked nowhere - that’s visible nowhere.

Checklist: what to do on your site in 15 minutes

1. Open Webmaster → ‘Performance’ → ‘Site visibility in Alice AI’.
2. Look at the badge and the SoV. If there’s no data - the site is most likely weakly visible in the results for its topic; the problem isn’t Alice.
3. Export the queries to XLS. Mark which of your pages are getting cited - those are your workhorses, reinforce them.
4. Switch to ‘queries without my site’ and the ‘Sites’ tab. That’s a ready-made map of competitors in AI answers: who Alice cites instead of you and which topics have holes. Split the holes into ‘not my profile’ and ‘my profile, no page’ - the second list is your content plan.
5. Set a monthly reminder. The data is weekly; there’s no point twitching more often.

A neural net already stands between you and your client - I wrote about that in the breakdown of a client’s chat with an AI. Now there’s at least a native instrument showing how you look in that layer. The instrument is free. Not looking is strange.

If you want to understand why Alice cites your competitors and not you - I’ll break it down in a one-off audit. Quick questions - in Telegram @dimik90.