Data-driven attribution: what Google dropped and what Russia still runs on

You open the conversion settings in Google Ads to pick an attribution model, and half the list has disappeared. First click, linear, position-based, time decay: none of them can be chosen for a new conversion action any more. Two options remain, data-driven attribution (DDA) and last click. This is no interface glitch. It is a change of standard, and your reporting has to be rebuilt around it.

For a company that sells in several markets, there is a second layer to this. Google stopped serving ads to users in Russia in 2022, so the Russian part of your paid traffic runs through Yandex Direct, and Direct reads its conversions from Yandex Metrica. Metrica did not follow Google’s path. The result is two reporting systems that divide credit for a sale by different logic, and a team that reads both as if they were the same.

What rule-based models actually did

Ad platforms used to offer several ways to split the value of one conversion between the touchpoints on a customer’s path. First click gave everything to the first interaction. Last click gave everything to the final touch before the conversion. Linear divided value equally among all touches. Position-based put extra weight on the first and last. Time decay raised the weight as the touches got closer to the purchase.

All of these, last click aside, share one trait. They distribute value by a rule fixed in advance instead of by analysing how the channels really performed. You pick a template, and it splits the credit the way its author decided, whatever your own data says.

What data-driven attribution does differently

DDA works on another principle. In place of a fixed rule, it uses an algorithmic model that looks at the historical data in your account and estimates what each touchpoint actually contributed to the conversion. The logic is no longer “the first touch always gets forty percent”. It becomes “in your niche, with your mix of channels, this touch is worth this much”. For paths that cross several channels and take a long time, that is more accurate than any template.

Google rolled the change out in stages. Time decay, position-based, linear and first click stopped being available for new conversion actions, first in Google Analytics 4 and then in Google Ads, and the full retirement for existing actions followed. Last click stayed as the only preserved alternative to DDA. Since the rollout, the share of conversions measured on rule-based models has fallen to fractions of a percent, and DDA became available to most advertisers without special data volume requirements.

Where Russia differs

Here the Western playbook stops describing your whole business.

Yandex Metrica still offers a set of rule-based models in its reports. You can view the same traffic by first click, by last click, or by last significant click, which ignores internal and certain low-value transitions when deciding which source gets the credit. Yandex adds and renames options over time, so check the current list in your own account rather than relying on a screenshot from a year ago.

That has three practical consequences for a foreign company.

The first is that your global dashboard and your Russian dashboard can disagree for structural reasons. If Google Analytics credits a sale by DDA and Metrica credits it by last significant click, the same week will show different channel shares. Neither system is wrong. They answer different questions. Before anyone argues about which channel “really” drove revenue in Russia, write down which model each report uses.

The second is that the model choice in Metrica is still a decision somebody makes, and it quietly shapes conclusions. In Metrica, a person picked a model at some point, possibly years ago, and every report since then reflects that choice.

The third concerns bidding. Yandex Direct optimises toward Metrica goals, so the quality of those goals matters more than the attribution label on top of them. An automated strategy fed with a vague goal such as “visited two pages” will optimise toward vague behaviour whatever the model says.

What this changes in your reporting

The practical shift sounds simple and is heavy in execution. On the Google side, the split of conversion value is now decided by an algorithm, not by a rule you chose. So familiar conclusions like “first click brings little, let’s cut it” can no longer be drawn from a model setting, because that setting has almost vanished.

DDA has its own cost. It is less transparent for manual checking than a rule: you cannot open the formula and see with your own eyes why a channel received exactly that share. The compensation is a change of habit. Trust granular data more than convenient aggregates. If the machine calculates contribution from data, your decisions should come from data too, and not from rounded averages in the top line of a report. The habit applies on the Yandex side too.

Distrust the aggregate, look at the keyword

Here is the principle on an example where the aggregate misleads especially badly. Conclusions such as “a higher position means a higher CTR” or “a cheaper click means a cheaper lead” cannot be drawn from a summary report for the whole account. That report hides what is happening at the level of a single keyword.

The catch lies in how impressions are distributed. For most keywords, nearly all impressions concentrate on one or two positions, and not necessarily the top one. It can be the third or fourth if the bid does not allow higher. Test how metrics depend on position using such keywords and the picture comes out distorted. This is classic survivorship bias: the observed “dependency” reflects the small number of observations in most cells, not a cause.

To get a meaningful answer, the sample includes only keywords whose impressions are spread significantly across at least three or four positions, with no single position dominating. You build this in a BI tool at keyword level, broken down by position, date, cost per click, traffic volume and conversions. Only on such a slice can you see whether a real dependency exists and which way causality runs, which metric comes first and which is the consequence.

In Yandex Direct this analysis needs one adjustment. Direct describes where an ad appeared through a traffic volume metric rather than the classic position number Google users are used to, so the “position” axis in your slice has to be built from that. Check it on your own data. The composition of the auction and the distribution of placements differ from account to account, and in Russia you compete against a different set of advertisers than at home, so conclusions borrowed from your Google account or from someone else’s case study do not transfer. This is the same shift DDA embodies: from a universal rule to testing against specific data.

Collect the data first, then attribute

Neither DDA nor manual keyword analysis works if behavioural data has not been collected. And collecting it from scratch on every new project is slow.

A reusable Google Tag Manager container helps here. A large share of the tags and triggers in a typical container is universal for any site: basic clicks on buttons and phone numbers, visibility of standard blocks, form tracking at every stage, handling of the cookie consent banner. Such a set is kept as a template and carried into new projects, with only the site-specific events configured on top. In my experience this saves up to a working day per project compared with setting everything up again.

For a company selling into Russia from abroad, the same container can send events to both Google Analytics and Metrica goals, so the two systems at least count the same actions. Two local details are worth building in. Contact through messengers is common in Russia, so clicks on messenger links deserve their own events next to phone and form events. And rules around personal data and consent follow Russian law, 152-FZ, and not the EU model you may have configured at home. Treat the consent logic in the container as a legal question for Russian counsel, because it decides what data Metrica gets to see at all.

One nuance deserves attention while building such a template. The overall scroll depth percentage says little on its own, because pages differ in length and structure: on one page half the scroll is a couple of sections, on another it is ten times more. So the template includes visibility triggers for specific sections in addition to scroll percentage. Otherwise neither the algorithm nor you will see on which screen a visitor leaves and on which one they take the target action.

Fully configured site analytics means dozens of event types, from reading depth in text blocks to form work and video starts. A large share of them does not depend on the specifics of the site, which is why keeping them as a template pays off. These detailed events, rather than bare visits, give DDA and your own analysis the material that shows the contribution of each touch.

What to do with your analytics in the coming weeks

• Check which model each of your conversion actions uses in Google Ads, and whether forgotten settings on retired models are still lingering somewhere.
• Write down which attribution model your Metrica reports use, and who chose it.
• Rebuild reporting around the fact that an algorithm now distributes value on the Google side: compare periods and channels instead of arguing about the distribution rule.
• Stop drawing conclusions about placement and bids from the whole-account summary. Test dependencies on single keywords whose impressions are spread across several positions or traffic volume levels.
• Assemble a reusable event tracking container that feeds both systems, and include section visibility, messenger clicks and consent stages.
• Test every hypothesis on your own data, not on other people’s cases or universal rules.

A rule is convenient because it requires no thinking: pick a template and the credit is divided for you. DDA takes that convenience away and demands that you look at the data instead. That is irritating right up to the moment when the rule misleads you once again and the data shows things as they are.

If you are not sure your reporting survived the move to the new attribution, or suspect that your Russian and global numbers are telling different stories, book a free review and we will look at your data to find where the picture diverges from intuition. How analytics fits into a paid traffic system for Russia is covered on the Yandex Ads page.

Frequently asked questions

Sources

• Google Ads Help, about attribution models: support.google.com/google-ads/answer/6259715
• Yandex Metrica help, attribution models: yandex.com/support/metrica/general/attribution
• Russian version of this article on belokrylovo.ru: Атрибуция DDA: конец правило-ориентированных моделей