Attribution after the death of cookies: 3 working models in Russia 2026

In January 2025, I opened a client report and saw that 31% of “direct traffic” in Y. Metrica were actually transitions from mobile instant messengers without UTM. Previously it was considered “direct”. Now - thanks to changes in referrer processing - these are “direct sessions”. Attribution, which worked for five years, stopped working. And this is just the beginning of the story about cookies in the Russian Federation.

When people say “the death of cookies,” they usually mean Chrome and 3rd-party. But in Russia it's a different story. In the Russian Federation, it was not Chrome that fell off - but ITP in Safari, ATT in iOS and a package of tightening in Ya. Metrica, plus part of the traffic now goes through the VK and TG mini-applications, where the usual UTM does not reach the end system. This article is not a general theory. These are three working attribution models that I am using on 5 projects right now, and real before/after metrics on each.

Attribution in 2026 is not a question of “which tool to install.” The question is “which model to apply, because there is one technology, but three interpretations.”

What went wrong in RF analytics from 2024 to 2026

A dry chronology of what happened to tracking over two years. I checked on the projects ID Store, MM AI Trading and three e-com clients - intersections in this list.

What has NOT fallen off: 1st-party cookies in Ya.Metrica on its own domain, server-side tag manager, and basic UTM markup for desktop traffic. This set is enough for a working model, but not enough for a “as it was in 2020” model.

Model 1. Last-click + post-view: for a clear funnel with one entry channel

When to use: mono-channel performance, short transaction cycle, one visible-touchpoint before conversion. Gaming with direct purchase to landing page, e-com with one retargeting chain. On NEMIFIST case I had exactly this case.

Attribution scheme:

• Last-click assigns the conversion to the source of the last touch
• Post-view adds atribution to the banner display within a 24-hour window if the user converted directly
• There is no doubling - last-click always has priority; post-view is only taken into account if last-click = "direct"

Technically: Ya.Metrika + custom parameters in the tag plus server log of impressions via Metrica Pro (for post-view). This is a working, cheap model - 5K ₽/month. per project, provided that you have one landing page and a clear funnel.

Where it merges: everything that is multi-touch. If a user came from VK Ads, left, returned a week later through search, and after another 3 days bought from a direct visit - last-click will give a “direct” conversion. This not true, and this model cannot be applied.

Model 2. MMM-light: media-mix modeling without 1M ₽ on DataLens

Media-mix modeling is a statistical method that does not need user-level cookies at all. You take aggregated data “channel budget × week” and “revenue × week” for 12+ weeks, run regression, and get elasticities. Previously, this was the prerogative of large advertisers with a DS team. Since 2024, this is a calculator in Excel and one prompt for Claude.

Minimum data set for MMM-light:

1. Budget by channel - weekly spending in rubles (Yandex Direct, VK Ads, TG Ads, influencers, SMM organics as 0)
2. Dependent variable - weekly revenue or number of transactions
3. Control variables - seasonality (Fourier decomposition), holidays, days before payday

Next - linear regression with adstock transformation for each channel (takes into account the “delayed effect” of advertising). I use this as an Excel model + reconciliation through Claude using the prompt, which is described in the article 7 prompts for a marketer, in the section about unit-economics-segmentation.

What does it give at the output:

• Share of each channel in revenue (not in clicks)
• Saturation point for each channel - from what budget the channel stops bringing additional transactions
• Cross-channel effects: “VK Ads warms up the audience for Yandex Direct” - this can be seen in the coefficients

Real figure: on an e-com project for 2024, MMM-light showed that 23% of the budget for VK Ads brought 41% of conversions. After rebalancing in Q3, revenue increased by 18% with the same total budget. This exactly the same story, which last-click did not show - because last-click gave a large share to “organic” and “direct traffic”.

Where it merges: on short horizons (less than 12 weeks) and on unstable budgets (when allocation changes every week). Regression requires sustainability.

Model 3. Incrementality: A/B test of the channel against itself

The third model is the most expensive in terms of time and the most honest. The idea: you turn off a channel in one geography or segment for 3-4 weeks and see what happened to conversions. This is not a theory - this is an experiment.

Incrementality test structure:

1. Divide the country/audience into two groups: A - regular mix, B - without one channel
2. You launch an experiment for 3–4 weeks
3. Compare conversions and revenue in both groups
4. The difference is the incremental contribution of the switched off channel

In Y.Metrica, for such a test you can use geo-features or YA Audience segments, and in VK Ads - regional division. Compare groups based on incremental conversions and revenue, checking their comparability in advance. Attribution figures for an individual project should only be published together with the experimental design and source data.

Where it merges: businesses that cannot afford to “turn off advertising for 4 weeks” are the majority of e-com and SaaS. And the amount of data in one of the groups is insufficient - a minimum of 5,000 transactions are needed in each branch.

Which model to use when - table

What I stopped doing in 2026

A list of tools and approaches that have left my work over the past 18 months. It’s not out of fashion, it’s out of practice that in the Russian Federation they either stopped working or started lying.

• Google Analytics 4 for Russian projects. Unstable reference due to blocking, plus GA4 is basically worse than Y. Metrica on Russian-domain traffic. I keep it for a project in the UAE, but I removed it for projects in the Russian Federation.
• End-to-end analytics based on UTM in TG bots. 25% of UTM is lost in iOS in-app browser. Replaced deep-link with its own parameter in URL-payload.
• Last-click for multi-channel funnels. “All 5 channels are working, what should we turn off?” via last-click is a lottery, not a solution.
• Attribution through CRM-events without reconciliation with Metrica. CRM records the “paid” event, but not the source. Without double reconciliation, it turns out that accounting and marketing are in different universes.

The stack on which I keep analytics in 2026

For those who want specifics on the tools, here’s what I currently have on projects:

• Ya.Metrica Pro + custom parameters via GTM alternative (Matomo Tag Manager). 1st-party cookies on own domain - the basis
• Server-side tracking via Cloudflare Worker for critical events - 8% accuracy is returned to the overall report
• n8n + Google Sheets to aggregate weekly data of all channels into one table (details in my article about n8n agents)
• Excel/Numbers + Claude for MMM models - so far without a separate tool like Robyn from Meta, on the sizes of Russian e-com its overengineering
• Y.Audiences + VK Ads regional split for incrementality tests once a quarter

Cost of a stack on a project with a performance budget of 2M ₽/month: ~30K ₽ of tools + 10–15 hours of my time per week. This is several times cheaper than boxed “end-to-end analytics” with a subscription of 80K ₽/month - and many times more accurate, because the parameters are selected for a specific funnel.

If you look at your report now and see 40%+ “direct traffic”, this is not a business indicator. This is a symptom that attribution is broken. Would you like me to analyze your stack in 30 minutes? discovery-colle — without selling the retainer blindly. During the call, I will show you which of the three models is suitable for your case and how much effort it will take to move.

Related materials: ROAS calculator, CPA calculator, ROAS benchmarks for e-com, a complete guide to performance marketing.