Cleaning placements in the Yandex ad network: build a whitelist, not an endless blacklist

Ad networks are built to let almost everyone in by default. Thousands of placements, new ones connected automatically, formats of wildly different quality. If you do not intervene, the budget spreads across random sites and you pay for impressions that will never become an inquiry.

Control in the network is not a one-off blacklist cleaning. It is continuous work: who to let in, who to exclude in advance, and when not to launch a format at all because there is not enough audience for it. Let’s take that in order.

Why the network hits sometimes and misses often

First, about what targeting in the network rests on at all. It works on behavioural data about the user, and that is not an abstraction.

Turn off activity tracking permissions in an app, and the recommendation feed slides within months toward almost random content unrelated to real interests. Turn them back on, and relevance returns not at once but cumulatively: first the most popular things, then the finer, more stable interests.

The conclusion for advertising is simple. Precision in the network is a function of the data the platform has managed to collect. Where data is thin or noisy, the system shoots blind, and that is exactly where waste accumulates.

So placement control is not a fight against the algorithm, it is help for it: you narrow the field to where your audience actually spends time, instead of waiting for the network to guess. And do not count on the network cleaning itself. It earns from impressions being served, so its incentive to cut weak placements is much smaller than yours. Narrowing is your responsibility, not the platform’s.

A whitelist instead of endless blacklist cleaning

You can clean a placement blacklist forever: new domains appear where blocked ones were. The inverse logic works better, a whitelist, where only verified domains are allowed in.

You cannot assemble one by hand across thousands of placements, but you can do it semi-automatically.

From every domain, pull the metadata of the home page: title, description, first-level heading. Clean the text of punctuation, lowercase it, lemmatise it. Across the whole array, build a frequency dictionary of words and label the top of that list by hand: core industry words, trusted news portals, entertainment and information resources, adjacent topics. Run the labelled dictionaries back over the domains and sort the sites into categories. After sorting, direct industry competitors rise to the top, and there you re-read the metadata with your own eyes so nothing wrong gets in.

The finished lists then go two ways.

As a whitelist of allowed placements in the network settings.

As the basis for interest audiences. Each list of domains, whether competitors, news portals or entertainment resources, becomes a separate custom audience. Those can then be intersected with each other and with standard segments using AND and OR, narrowing or widening reach for the task at hand.

One quality note: the result depends heavily on correct lemmatisation and on the final manual check. Russian morphology makes naive word matching unreliable, and with uninformative metadata the automation will happily file a domain in the wrong category.

Exclude the obviously dirty pre-emptively

A whitelist is about who to let in. There is a mirror task: throw out in advance the placements you already know you should not appear on.

A classic example from network practice. When a large social network became a partner of Google’s display network, ads started running next to content that would not pass the usual partner-site moderation. The fix is simple and pre-emptive: add both domains of the platform, the historical one and the current one, to the exclusion list at account or campaign level, after which impressions there stop across every campaign using that list.

The meaning is wider than one case. Brand safety in the network is a blacklist you maintain ahead of time, not a reaction after the fact, when the money is already gone and the reputational contact has already happened. And such exclusions are worth reviewing periodically: partnership arrangements between platforms change.

For a foreign brand there is an extra layer here. You may not recognise which Russian sites are respectable and which are content farms with a familiar-looking name. That judgement needs someone who reads the language daily, which is an argument for doing the labelling with a Russian speaker rather than by domain authority metrics alone.

Audiences and retargeting: size them before you launch

A separate zone of waste is retargeting and complex audiences launched without checking volume. It is easy to build an elegant construction here that physically cannot gather impressions.

Before launching retargeting on an audience with a short retention window, calculate whether site traffic can fill it. In analytics, create the widest possible audience with no extra segments, set a short retention period, and look at the volume estimate for a week.

Such an audience refreshes itself: people leave it when the period expires, and without a constant inflow of new visitors it does not grow, it holds steady or falls. The technical minimum to launch is small, but steady delivery with actual clicks starts in the tens of thousands of users.

The logic is simple. With the modest click-through typical of networks and a site conversion rate of a few percent, getting even single-digit inquiries requires many times more impressions than clicks, which means a large base audience. For that much to accumulate with a short retention window, the site needs on the order of hundreds of thousands of unique visitors a month, uniques, not sessions. If a week brings a few thousand, there is nothing to launch: the audience has nowhere to come from. And if it is clear in advance that the site will not reach that volume, waiting a week to confirm is pointless.

The second trap is the restrictions on audience types, which people trip over after the construction is already built. Combined audiences, made by intersecting several custom ones, work only in classic network campaigns. In unified performance campaigns, custom audiences are allowed but not combined ones built on top of them, so for that campaign type you prepare enough variants in advance: you cannot combine them on the fly inside the system. On search, neither custom nor combined audiences apply at all, neither for observation nor for targeting.

The pattern is simple: the more complex the audience construction, the fewer campaign types support it. Hence the rule: plan the audience architecture starting from the choice of campaign type, or the construction you assembled turns out to be unusable where you need it.

What to do with the network this week

1. Check tracking settings and data quality. Where the network has nothing to aim with, waste accumulates, and placements have to be narrowed by hand.
2. Build a placement whitelist by labelling domains from their metadata, instead of cleaning the blacklist forever.
3. Start a pre-emptive blacklist of known-dirty placements and maintain it ahead of losses, not after them.
4. Size the retargeting audience before launch: work out whether site traffic can fill it.
5. Pick the campaign type before building audiences, so combined constructions do not turn out to be inapplicable.

Networks do not forgive “switch it on and forget it”. They are built so that by default they work toward serving impressions, not toward your inquiry. The difference between waste and a working channel is a few lists that somebody maintains by hand: who we let in, who we exclude, and whether there are enough people for the format to make sense at all.

If you want to see which placements your network budget goes to and how much of it is junk, that is one of the first things I look at in the free review.

Frequently asked questions

Sources

• Yandex Direct help, the Yandex Advertising Network: yandex.com/support/direct/en/campaigns
• Yandex Audience, building and combining segments: yandex.com/support2/audience/en
• Russian version of this article: РСЯ без слива: чистка площадок и контроль