How to measure Yandex Search quick AI answers and Alice AI separately

A test condition here means the screen and mode in which a person receives an answer. The quick block under the Yandex Search bar, a normal Alice AI run with automatic model selection, and a run with Expert selected manually belong to related products but start differently. Combining them into one share makes it impossible to see where a brand appeared, where a link disappeared, or where the service did not show the required answer type at all.

The practical unit is one run of one prompt in a named condition. It stores the launch method as well as the final text. This approach does not claim that the three conditions must use different website-selection rules. It simply avoids replacing an unknown internal process with a convenient blended number. The official announcements support the product differences; separate measurement is our editorial inference.

Section sources:[1] Яндекс[2] Яндекс

The editorial team checked two official Yandex announcements and extracted their statements into a two-record evidence set. This is not brand-visibility data and not a user experiment. It is a reproducible source matrix that records the date, named product, directly supported function, and the boundary of what each announcement does not establish.

The matrix lets a later editor repeat the check without adding claims that the announcements never made. The article uses it only for supported product descriptions. The three-condition protocol remains an editorial inference and is not presented as a Yandex ranking rule.

• 11 September 2026 — Alice AI Search Pretrain: the announcement supports the open base model and quick answers from a live fine-tuned version under the Search bar; it does not disclose web-page selection rules.
• 15 September 2026 — Alice AI automatic selection and Expert: the announcement supports automatic or manual launch and work with search, files, services, and tools; it does not disclose source weights or a formula for showing a brand.

Section sources:[1] Яндекс[2] Яндекс

On 11 September 2026, Yandex announced Alice AI Search Pretrain, a base model released openly. The same announcement says a fine-tuned Alice AI Search version has produced quick answers under the Search bar since July. This distinction matters for verification: the open base model and the live fine-tuned version are not the same object. Running the downloaded model does not reproduce the live Yandex Search result.

For a quick answer, record whether the block appeared, what text it showed, which page addresses were visible, and whether a link led to the target material. A missing block, a missing brand, and a missing link are three different records. The announcement describes the model and answer location, but not the pages considered, their order, or the reason a particular source appeared.

Section sources:[1] Яндекс

On 15 September 2026, Yandex said Alice AI assesses a request and selects a suitable model. It may use a fast model for an everyday task and start Expert for a task with many conditions. Yandex describes an automated selection system that considers the request and the expected computing need. The person does not have to choose settings in advance.

Record that the selection remained automatic. If the screen names a mode, preserve that name; if it does not, do not guess. An answer produced after automatic selection is not directly comparable with a run where a person selected Expert in advance. A difference may come from the model, web search, tools, account state, or normal answer variation.

Section sources:[2] Яндекс

Yandex explicitly says Expert can be selected manually, although Alice AI starts it automatically when the system considers it necessary. Expert can search for information, examine several files, call services, and use other tools. Manual selection therefore changes a known launch condition before the answer appears.

If a brand appears only after manual selection, the defensible record is narrow: it appeared in that Expert run. This does not show that the brand is present under automatic selection or in the quick Search block. Likewise, one missing brand does not prove global absence from Alice AI. Each condition needs repeated runs and the same classification rules.

Section sources:[2] Яндекс

Begin with user tasks: finding a company fact, comparing options, choosing a product, or locating an instruction. Freeze the exact wording of each prompt before testing. Run that set separately in the quick Search answer, Alice AI with automatic selection, and Alice AI with Expert selected manually. If one option is unavailable, mark it unavailable rather than recording zero visibility.

Do not change a prompt after the first successful answer or use different wording for different modes inside one comparison. A separate study may test wording variants, but it needs its own version. Repeats reveal variation, yet they do not turn one pattern into a permanent Yandex rule.

• The same prompt wording and language in all three conditions.
• The same region and comparable account state.
• A planned repeat count and one collection period.

Section sources:[1] Яндекс[2] Яндекс

For each run, record date and time, country or region, language, device, sign-in state, exact prompt, launch condition, and repeat number. Preserve the complete answer, visible page addresses, and a screenshot. If the service shows its work or names the mode, save that too. Do not publish personal data, access tokens, or restricted details from uploaded files.

Record errors, access limits, incomplete answers, and a missing answer block separately. Do not automatically count these cases as a missing brand because the test may not have completed. When a link is visible, preserve its exact address rather than only the domain name. A homepage link and a link to the page supporting a specific fact answer different measurement needs.

Section sources:[1] Яндекс[2] Яндекс

For each condition, show the raw count and the total number of successful runs. For example, “brand named in 7 of 12 successful answers” and “exact link shown in 4 of 12” are different results. Do not put technical errors in the total of successful answers, but show them on a separate line so a high share cannot hide weak coverage.

Separate answer-block availability, brand mention, recommendation, visible link, exact target-page link, key-fact accuracy, and confirmed user visit. A quick answer may name a brand without a link; Expert may show a link but state a fact incorrectly; a server may receive a crawler request while the user answer does not show the page. One score cannot substitute for these events.

Section sources:[1] Яндекс[2] Яндекс

If a quick answer shows a brand while automatically selected Alice AI does not, that is a difference between two conditions, not proof of a website defect. Repeat the run and check whether language, region, account state, or wording changed. If the gap persists, examine each answer and its links separately. The public Yandex materials do not reveal the hidden cause behind one observation.

If Expert finds the target fact only after manual selection, that is a useful product observation: the more involved path completed the task in the tested repeats. It does not reveal which page was considered first or which internal signal determined the outcome. State the measured event and period, and do not generalize it to every request, account, and future product version.

Section sources:[1] Яндекс[2] Яндекс

The search specialist checks whether an ordinary search system can access the target page and whether the page supports its facts in clear text. Editors and public-relations specialists keep the same facts consistently supported across public pages. The analyst keeps three separate run sets and does not present a crawler request, brand mention, or link as a visit or sale.

Neither official announcement provides a formula that guarantees a website or brand appearance. Page markup, a newer date, a file aimed at AI systems, or a wording change may be testable hypotheses, but they are not proven levers. Review this guide on 5 November 2026 or earlier if Yandex changes quick answers, automatic selection, or Expert availability.

Section sources:[1] Яндекс[2] Яндекс