Yandex Direct automated strategies: train them instead of fighting them

An automated bidding strategy is not a lever you push in the right direction. It is a student.

It looks at your conversion data and decides for itself who to show the ad to and at what price. While data is thin, it makes mistakes. When data is sufficient, it often beats manual settings. Most conflicts with automation start where people lean on it like a mechanism instead of feeding it like a learning system.

Here are the situations where advertisers usually break something that was already working.

Why you should not pull a campaign out of a portfolio strategy

A portfolio strategy joins several campaigns into one optimisation zone. The system learns and distributes bids across the whole group at once, not campaign by campaign.

A typical scene. One campaign in the portfolio consistently under-delivers traffic. You want to speed it up, and the hand reaches for moving that campaign into its own strategy to get control.

That is the trap. To take one campaign out of a portfolio, the system has to reassign the strategy and the daily budget to every campaign that was in it. Each of them starts learning from scratch. Including the one that had been running like clockwork.

The outcome is predictable and annoying. The campaign that was under-delivering carries on under-delivering, because the problem was not the strategy and separating it cures nothing. And the second campaign, previously stable and through its full learning phase, which cost both time and budget, sags after being forced to relearn. Especially if the strategy runs without a cost per click ceiling: with no upper limit, bids after a reset can go anywhere.

The rule is short. What was not working will not break much further. What was working and then stopped is something you will have to fix and explain. An urgent “push this one harder” is not worth the risk of zeroing out a neighbour’s learning.

What to feed the algorithm when conversions are thin

An automated strategy learns from conversions. No dense flow of conversions, no learning: the algorithm has nothing to find the link between audience, creative and the probability of an inquiry.

Hence the rule for the heavy campaign types. A fully automated campaign type is built entirely around the automatic strategy and gives no manual control over where impressions go. Launching it before the account produces at least a few dozen conversions a month is pointless: the campaign either gets stuck in learning or comes out of it on thin, noisy statistics, and then its bids are close to random. Grow the conversion base on manageable campaign types first, then connect the automation.

When real inquiries are too few, learning can be fed temporarily with micro-conversions: scroll depth, time on site, a click on a quiz button, revealing a hidden phone number. These are intermediate signals that happen more often than the final inquiry.

And here a quiet overpayment hides.

If the payment settings have several different micro-event types selected at once, money is charged for each triggered event separately, not once per user. The same visitor in one session often scrolls the page, spends the required time and opens several cards. Payment is configured for all of that at once, and for one real lead you pay two, three or four times more than you calculated.

A rough but working correction: divide the target price of each micro-conversion by the number of event types you pay for simultaneously. Paying for two types, divide by two. For three, divide by three. It is not exact maths, it is insurance against paying several times over.

Where manual is more honest than automation

Taming an automated strategy does not mean handing it everything. Some tasks it solves worse than a direct setting, and trying to do them “through account architecture” creates problems of its own.

Raising the bid for one region. The temptation, especially for a newcomer, is to create a separate campaign for that region. But a new campaign is not free: its own history, its own learning, its own budget, extra load on management. Creating a heavy object for the sake of one bid is irrational. Switch the campaign to manual control, add all the regions to the observation list and set an adjustment for the region you need. “Raise the bid on part of the traffic” is a point setting, not a reason to split the account.

Rescuing a badly built look-alike audience. Suppose you built a look-alike and its gender and age distribution does not match the profile of the people who actually submit inquiries. You do not have to throw it away. Leave the audience as it is and apply a minus one hundred percent adjustment to every irrelevant gender and age combination. The ads then run only on the part inside the audience that matches your conversion profile. Geographic mismatch is not fixed with a separate adjustment: campaign geotargeting already handles it, so restrict impressions to the region you need and switch off its expansion.

The difference in thinking matters more than the techniques themselves. You will not out-stubborn an automated strategy with adjustments layered over its decisions. But trimming an irrelevant slice of audience, or setting a regional range, is not a fight. It is normal work alongside the algorithm.

What this means for a company running Yandex from abroad

Two things, both about patience.

First, the learning phase costs money and time, and it is the asset you are building in the first two months. Any request that resets it, moving campaigns between strategies, restructuring on week three, switching the objective because a weekly report looked bad, throws that asset away. Agree in advance how long you leave a launch alone.

Second, thin conversion data is the normal starting state for a market you are entering. That is an argument for starting on search with manual or semi-automatic control, gathering conversions, and only then handing the wheel to automation. Not the other way round, which is what the platform’s own interface will suggest.

What to do with your strategies this week

1. Check the portfolio strategies. Before pulling out an under-delivering campaign, make sure the cause is the strategy and not the keywords, the bids or the landing page. Usually separating it does not solve the problem and does wreck the neighbours’ learning.
2. Assess the conversion base before switching on automation. Fewer than a few dozen conversions a month, and the fully automated campaign type waits; first grow volume on manageable campaigns.
3. When paying for micro-conversions, divide the target price by the number of event types, or you pay several times over for one lead.
4. Fix a regional bid with an adjustment inside the existing campaign, not with a new campaign.
5. Do not throw away a skewed look-alike. Trim the irrelevant gender and age with a minus one hundred percent adjustment and close geography with geotargeting.

Automated strategies punish impatience more harshly than they punish a settings mistake. A mistake is visible and gets fixed. Reset learning is invisible for weeks: the campaign simply “sagged for some reason”, and nobody connects it to one line pulled out of a portfolio a month ago.

If your automated campaigns behave unpredictably and you cannot tell whether you are feeding the algorithm or fighting it, that is exactly what I look at in the free review.

Frequently asked questions

Sources

• Yandex Direct help, automated bidding strategies: yandex.com/support/direct/en/strategies
• Yandex Direct help, bid adjustments: yandex.com/support/direct/en/campaigns
• Russian version of this article: Автостратегии Директа: приручить, а не бороться