At the YaC 2020 conference, Yandex presented its new algorithm – YATI. The abbreviation stands for Yet Another Transformer with Improvements, which translated into Russian means “Another transformer with improvements.”
Yandex presents this innovation as “the most significant change in search over the past 10 years.” In our article, we will analyze what the new algorithm is, what its task is in search, and how it will affect the ranking of sites in Yandex.
Yati algorithm: what’s new and what is the transformer for?
YATI is designed to analyze the texts of search queries and sites that are shown for these queries, and it should do this more efficiently than its predecessors – Palekh and Korolev. You can dive into the history of Yandex algorithms in more detail in our article.
In this article, we will only briefly go through them in order to understand what has changed with the advent of the transformer.
First there was Palekh
Palekh was created to compare the texts of users’ search queries and page titles, the training took place on positive and negative examples that were taken from previously accumulated statistics.
Since the algorithms of search engines cannot read texts, the determination of the correspondence between requests and headers was carried out by comparing numbers.
How was the determination of compliance? To explain in simple words: each page corresponded to a group of 2 numbers, where each number meant a specific coordinate on one of the 2 axes, and each page corresponded to a point on the coordinate plane.
The translation of the search query took place in the same way – it was placed in the same coordinate space along with the page, and the closer they were to each other, the more the page matched the query.
It is this way of processing requests and pages that is called the semantic vector. The advantage of this method was that it could match the request to the page, even if they did not have a single common word.
This algorithm was relatively heavy, so it was applied only at the very late stages of the ranking, to about 150 best pages from the filtered set.
And then Korolev came
It worked according to the same architecture as Palekh, however, with its appearance, the semantic vector of pages began to be calculated in advance, offline, which made it possible to apply the algorithm to a larger number of pages.
It works like this:
the algorithm in offline mode calculates a vector of pages and saves it to the index base,
the user enters a query in the search,
the algorithm translates the request into a vector,
multiplies with ready-made page vectors, calculating relevance.
If the page vectors were not calculated in advance, then it would not be possible to apply neural networks to a larger number of pages without sacrificing the time it takes to display the results at the user’s request.
In addition to comparing the vectors of the request and the pages, Korolev began to compare the vectors of the newly entered requests with other requests for which the best answer was known in advance. If the vectors were close enough, then the results should be similar.
Why did you need transformers then?
The previous algorithms, despite the fact that they improved the ranking process, still coped with this imperfectly. The main shortcomings include:
lack of full account of the word order,
the whole point of the page was described by one vector with a limited size.
In the transformer, each element of the text is translated into a separate vector, while maintaining the position in the text.
This transformer is trained according to the principle of transfer learning, that is, at first one problem is solved, within which information is accumulated that allows the algorithm to solve this specific problem, after which the same information is used again, but for solving other problems. Initially, the transformer is “fed” a lot of simple and not always reliable information received from Yandex.Toloka users, and it is pre-trained on this information. After this stage, the transformer receives more expert information, but in smaller volumes – from specially trained Yandex assessors.
One of the additional differences of the transformer is the prediction of the user’s click – this is an additional metric that will be taken into account when ranking.
YATI and its impact on search engine promotion
Based on the fact that this algorithm is aimed at a deeper analysis of the text, understanding its essence, we can reasonably conclude that the semantic load of text content will have a more significant role in ranking. This means that expert texts that provide a complete and high-quality answer to a user’s request will be shown more and more often in the SERPs.
By the way, it is these texts that we have long begun to implement on the projects of our clients and have written more than once about how to improve the content on our website. I recommend reading this article about useful content.
YATI and BERT: what’s the difference?
Last year, Google announced a connection to the ranking algorithms of its transformer – BERT. This neural network solves the problem of analyzing search queries and their context, and not a separate analysis of key queries. That is, BERT analyzes the entire offer. We have already talked about this in more detail with illustrative examples here.
Based on the descriptions of YATI and BERT, it becomes obvious that both are transformers, and are aimed at better understanding the meaning of user requests. However, in this case, YATI looks more advantageous, since in addition to analyzing the text of user requests, it also analyzes the texts of documents, learns to predict clicks. Therefore, we can conclude that YATI is a more significant update than BERT.
How to prepare your site for YATI ranking?
Does the appearance of Yandex YATI mean that the old optimization methods will no longer work?
No, in no case does the appearance of this algorithm mean that, for example, the same optimization of the H headings and the title and description tags does not play a role. It is necessary to understand that the new Yandex algorithm does not cancel the ranking factors that were deduced earlier, YATI only supplements them with a better analysis of texts.
And this means that technical optimization, attraction of natural links and improvement of behavioral factors (on the search and on the site) cannot be abandoned. As before, effective results will be brought only by working on the site as a whole.
General recommendations for improving texts:
Structure text with headings using keys.
Work out the title and description to improve the CTR of the snippet.
Use LSI phrases. Learn more about LSI copywriting “
Collect the most complete semantics for pages.
Summing up the results of the parsing, the new YATI algorithm will definitely lead to changes in Yandex search results, however, since the system is learning, it will still take time, so even if your site uses SEO texts, there is still time to refine them to modern search standards systems. To check how relevant your texts are, leave a request on this page.