Yandex publishes how it decides whether an algorithm change is an improvement. It is the most revealing page in the whole documentation, because it names the two metrics every ranking change is tested against.
The stated principles
- Results are generated entirely by machine-learning algorithms, and cannot be reordered manually.
- Relevance is computed from the query, the page content, the history of user interaction with those pages, language, location, relationships between pages and many other factors.
- Indexing completeness is a top priority, applied uniformly per source type. Content is removed from results only for spamdexing, harm to users, or legal violations.
- Presentation format is chosen by the likelihood of meeting the user’s objective — not by the source.
Proxima — the page quality metric
Built from assessor evaluations plus additional signals. It weighs relevance to the query (including expert evaluation by subject specialists), the likelihood of the user’s objective being met on that page and site, the quality, usefulness and uniqueness of the content, the balance between useful and intrusive content, additional signals on content quality and author credibility for complex topics such as health, legal and financial services, and how convenient the content is to consume.
Read as a checklist, that is the most concrete statement of what “good content” means in Yandex — and the intrusiveness clause makes advertising density a ranking input rather than a taste question.
Proficit — the results usability metric
Measures how quickly a user’s objective is met, from the quality and quantity of interactions with individual results and with other elements of the page. Formats with higher predicted Proficit are chosen, and the arrangement of the whole results page is optimised for both metrics together.
How assessors actually work
They evaluate pages and result elements, but their evaluations never enter ranking directly — they are training input for the machine-learning algorithms and a check on whether a proposed change is an improvement. Impartiality is managed through recruitment, training, instructions, tooling, and deliberate overlap between assessors. Changes that pass are then tested in a live experiment against a control group.
Why this is worth reading
Because it answers the question underneath every ranking discussion: what is being optimised. Work that raises Proxima and Proficit is durable; work that does neither is not, whatever it does to a position this month.
How we apply this
We use the Proxima criteria as the content brief on Yandex-facing projects, because it is the search engine’s own definition of quality rather than an inferred one. The line clients react to most is the balance between useful and intrusive content — it makes the pop-up conversation a ranking conversation. And the fact that assessor scores never touch ranking directly ends the recurring myth that a site can be manually penalised or promoted by a person.
Related services