AI VK engineers built a separate product recommendation pipeline in the VKontakte Feed: its own candidate selectors with a catalog of approximately 74,000 product posts, Ozon as the first partner, separate logging, and a multi-target ranker that evaluates the probability of an order for a "user — product" pair instead of the usual CTR. To overcome cold start, product domain logging was raised from the standard 2% to 100% via the VK Discovery platform, increasing the sample size by 15 times. Over a five-month A/B test, end-to-end conversion from impression to created order increased by 53%, the number of orders increased 43.2 times, partner clicks increased 55.3 times, and CTR increased by 95.6%.
What happened
The AI VK team published a breakdown of the mechanics and results in a blog post on Habr; the author of the article is Nikolai Karasov, a senior developer in the content services recommendations department. The logic is as follows: product posts are isolated in a separate pipeline within the established Feed recommender, with its own candidates, separate logging, and separate ranking. The multi-target ranker evaluates the probability of an order for a "user — product" pair, and the weight of signals is built across the entire funnel: the maximum weight is given to a created order, then a partner click, and only then likes and viewing time. To solve the data problem, product domain logging via the VK Discovery platform was raised from the standard 2% to 100% of impressions, which increased the sample size by 15 times. As a result of the five-month A/B test, the team reported a +53% end-to-end conversion from impression to order, ×43.2 in the number of created orders, ×55.3 in partner clicks, and +95.6% in CTR.
Context
The root problem that the case solved is the cold start of a new domain with rare target events: there are almost no orders in the Feed initially, which means it is impossible to train the model directly on them. Until orders accumulated, the model was trained on more frequent proxy signals, for example, on the expansion of the text of a product post, which correlates with a subsequent partner click; then the target was migrated up the funnel — this is a classic "proxy → target" scheme. The key lever was data, not the model: rare target events become trainable only with dense logging, so the targeted transfer of domain logging to 100% was more important than the choice of architecture. Essentially, the scientific novelty here is minimal — the architecture of the entire system was not changed, but the methodological value is high: this is a documented playbook for launching a product domain within a content recommender.
Why this matters for the industry
For teams building recommendation and e-commerce systems, this is a proven in production recipe for launching a new domain within a mature recommender: do not rewrite the architecture, but isolate the pipeline, raise the logging of the rare domain to 100% through the existing discovery infrastructure, start with frequent proxy signals, and run a multi-target ranker across the entire purchase funnel. This approach is applicable anywhere the target event is rare — from marketplaces to local services and job postings. The immediate benefit is a methodological checklist: check how much of the rare target domain is actually being logged (VK had 2% against the required volume) and what proxy events are available right now. However, the solution cannot be implemented "out of the box": there is no open API, OSS, or pricing, this is an internal engineering case. In the coming months, it is expected that other platforms with rare targets will repeat this, and the combination of "pipeline isolation + dense logging + target migration" will become a template in cold start breakdowns.
Why this matters for users
Product posts — "shops" — in the VKontakte Feed are now selected by a separate model that optimizes not likes and CTR, but the probability of bringing the reader to an order, so such posts are now selected more accurately: candidates are selected based on the user's product behavior and content interests. The first partner of the pipeline is Ozon, so purchases are made through its assortment. Sellers should take into account the change in logic: ranking now rewards content that actually leads to an order, not one that collects likes, so it makes sense to rework the design and presentation of product posts for the funnel. For those studying recommendation systems, it is useful to open the original breakdown by the VK team on Habr: it describes the mechanics of candidate selection and the full results of the A/B test.
What is still unknown / limitations
Methodologically, one should be cautious with the metrics: ×43.2 orders and ×55.3 clicks are primarily an artifact of launching a new pipeline from scratch with a low base at the start, and not proof of the superiority of the approach itself. An arithmetic self-check gives a volume of product post impressions that increased approximately 28 times (55.3/1.956 ≈ 43.2/1.53), meaning that a significant part of the absolute growth in orders is explained by the scale of impressions. Data for five months and one partner (Ozon) is not enough to judge the sustainability of the effect over the long term; the source does not claim anything about transferring the method to other VK domains. There is no open API, OSS repository, or pricing, and there are no separate external assessments of the system's quality outside the team's own A/B test.
Sources
Author
Look at AI, editorial team