Engagement metrics that mislead: four traps in Russian advertising data

A number next to a post or in a report row looks like a fact. But the same number can mean two opposite things, and sometimes it means nothing at all. Views can run several times below real opens. The account average hides the very thing you opened the account to see. One word in different grammatical forms is counted as different words, and the statistics inflate. And a platform metric can break in silence, sitting empty while real clicks keep arriving.

For a company that sells into Russia from abroad, these traps have an extra layer. The reports are read in a second language, often through a translated interface, and the instinct is to trust whatever the dashboard prints. The Russian tooling has its own conventions, and some of the Google habits that your team carries over are exactly what sets the traps off. Here are the four that most often sit underneath a wrong decision.

Views and likes are a poor measure of interest

Several Russian platforms have put view and like counters of their publications on public display. The temptation is obvious: a bigger number, a better piece. In practice the link is weak. A public view counter can be orders of magnitude below the real number of opens. The counter shows units or tens while the article was in fact opened by thousands. As an indicator of popularity the number is useless.

This matters more for foreign companies than it might seem, because the public counter is often the only figure a head office ever sees. A Telegram channel post, an article on a Russian content platform, a card in a marketplace feed: the counter is what gets screenshotted into the monthly deck, and the decision about whether to keep investing in the channel is made from a number that describes nothing.

The more useful signal is the share of people who opened the material and read it to the end. It speaks about the content more honestly than likes. If a large share of those who opened do not finish, the problem is in the structure or in the text, not in “low popularity”. The same principle carries over to any content that drives traffic: look past the top-line view count to whether people reach the end and reach the target action. A like is placed by one person out of many. Reading to the end reflects whether you held attention or not.

If the platform does not expose read-through, your own analytics can. Yandex Metrica shows scroll depth and time on page per URL, and a goal on the last block of an article gives you a completion rate that no public counter will ever contradict. That is the number to put in the deck.

The account average hides the truth

The second trap is subtler. You look at an overall figure for a campaign and draw a conclusion, while inside that average two different groups are hidden that behave in opposite ways. The classic example is comparing ads whose headlines use dynamic substitution, where part of the text is inserted automatically from the query, against ads with static headlines. The average across all ads will not tell you which type works better, because it averages them together.

To see the truth you have to segment the data by hand, and in Yandex Direct this is less obvious technically than it is in Google Ads. The Report Wizard hands you statistics but not ad texts, only their identifiers. The texts live in the campaign management tools. So the two sources are joined by ad ID, each row is labelled “dynamic” or “static”, and only then is the pivot built. And you look past a single average percentage to a breakdown by headline type across several metrics at once: absolute, relative and weighted. Until you have cut the data by a meaningful feature, any “fine on average” may be hiding that one half feeds the result and the other half sinks it.

A separate question is which metrics belong in such a breakdown. Absolute numbers like impressions and clicks speak about volume, not quality. Relative ones like conversion rate and cost per action are closer to the point, but they distort easily on a small base. And then there are the weighted indicators: average impression position and click position, bounce rate, page depth. These are the real engagement metrics, and they are precisely the ones lost when you look at a top-level summary. A segment where inquiries appear to be equal, but bounces are twice as high and page depth is half, behaves differently from its neighbour in the table, and the decision about it has to be different. One average number will never show that.

For a team managing Russia from another country, the practical consequence is that the weekly report from a local agency or contractor has to include the cut, not the average. If the report shows one conversion rate for the account, ask what feature it was split by. If the answer is “none”, the number is not yet evidence.

Word forms inflate the statistics

The third trap lives in word-level analysis of search terms. It seems logical to collect statistics by individual word and see which ones bring conversions and which only spend budget. But if you count words as they appear, without reducing them to a base form, the same word in different cases and word orders is counted as different entities. There are usually far more word forms than lemmas, and the statistics for each are fragmented until they lose meaning.

This is where the Russian market is genuinely harder than an English-speaking one. English nouns have two forms and adjectives have one. A Russian noun runs through six cases in singular and plural, adjectives agree with it in gender, number and case, and word order in a query is free. A phrase that is one entry in an English search terms report becomes a dozen entries in Russian, each holding a sliver of the clicks. Any word-level report that was not lemmatised is describing the grammar of the language, not the behaviour of the buyer.

Hence the distortion. Without lemmatisation you cannot collapse phrases with reordered words and different morphology into a single statistic. You cannot roll figures up to the level of a semantic category, setting aside secondary additions such as a city name or a “buy now” trigger word. The solution is to label the semantics in advance, before the campaign is even uploaded: the lemma, the semantic category of the word, the group of categories. Then the standard account reports let you look past a bloated list of word forms to meaningful cuts by category and by addition. The condition is strict: such reports collect data only from the moment the labels were applied. Accumulated statistics cannot be relabelled retroactively. That is why the labelling goes in as early as possible.

If your keyword research for Russia was done by translating an English list, the labelling step is where you find out whether the translation produced categories or only produced forms. It is also where a local specialist earns their fee, because deciding which additions are secondary requires knowing how Russians actually phrase a purchase.

The platform metric breaks quietly

The fourth trap is the most treacherous, because it gives no sign of breakage. Different systems pass the placement into UTM tags through different parameters. In Yandex Direct it is one dynamic parameter, in Google Ads it is another. They are easy to mix up when a tagging template is copied from one system to the other without replacement.

This is the most common tracking defect I see in accounts run from abroad. The head office has a global UTM standard, built around Google Ads macros. It is handed to the Russian contractor as a policy. The contractor applies it because the policy says so, and the Yandex Direct macro for the placement is never substituted.

On the surface everything keeps working: the link opens, the visit happens, the click is counted. But in analytics the placement parameter turns out empty, or contains something other than what you expected, sometimes the literal text of the foreign macro. Placement reports become incomplete or wrong, and it is hard to notice quickly, because visits are recorded normally. You make decisions about cleaning up placements based on data that does not exist. This is why, before using a tagging template for a particular system, you check it against that system’s current set of dynamic parameters instead of carrying the tagging over by analogy.

The check takes ten minutes. Open Yandex Metrica, filter visits from Yandex Ads, and look at the placement dimension for the network campaigns. If it is empty or shows a curly-bracket string, the template is wrong and every placement decision made since the launch was made blind.

What to do with the metrics

A short route:

• Look at read-through, not views. Judge whether people reach the end and the target action, not how many times a counter ticked.
• Cut averages into segments. Before trusting an overall figure, break it down by a meaningful feature and check whether the halves pull in different directions. Demand the cut from whoever sends you the report.
• Reduce words to lemmas. Label the semantics with meaning tags in advance, so that word-level analysis counts substance rather than Russian grammar.
• Check the placement parameters. Verify the UTM template against the current dynamic parameters of each system, so that placement reports are not empty.
• Ask about the absolute numbers. Behind any attractive percentage, find out what volume it was calculated on.

Data does not lie on its own. What misleads is the way it was shown and read: a top-line figure instead of read-through, an average instead of a cut, a word form instead of a lemma, an empty parameter instead of a placement. A good decision starts with understanding what a metric actually measures, before deciding whether it is the right one.

Which of your numbers can be trusted and which are showing a mirage is something I can go through with you on a free review. How to build traffic analytics for Russia that shows the real picture is laid out on the Yandex Ads page.

Frequently asked questions

Sources

• Russian version of this article on belokrylovo.ru: Метрики вовлечённости, которые вводят в заблуждение