Quality Score: what it means and what it does not promise

Quality Score in Google Ads has grown a set of legends. The main one goes like this: raise the score and your click price falls in proportion. It sounds coherent, it sells to a client in one sentence, and it does not survive contact with data.

For a company selling into Russia the legend has a second problem. Google Ads has not shown ads to users in Russia since March 2022, so the platform where the score lives is not the platform you will be buying traffic on. The working tool is Yandex Direct, and Yandex publishes no per-keyword score with named components. Yet the promise travels anyway, because the people building Russian accounts for foreign companies were mostly trained on Google and bring their habits with them.

What the score is made of

Quality Score is a rating that Google attaches to a keyword, and it is assembled from several components. The first and heaviest is expected click-through rate: how likely it is that people click on the ad for this query. It is calculated from a large mass of industry-wide statistics rather than from your account alone, so it reflects how people behave on this type of query more than it reflects your creative. Then come ad relevance to the query, and landing page experience.

An important consequence follows. Part of the score depends on you only weakly. If people on a given query rarely click on ads at all, no text will lift the expected click-through rate: that is a property of the query, not of your work. The score is first of all a diagnosis of where the weak link sits in the chain of query, ad and page. It is not a grade for effort.

Which is why it is worth reading by components rather than by the headline number. Low ad relevance says the text and the keyword grouping have drifted away from demand: phrases that only look similar have been dumped into one group and the ad answers a fraction of them. A weak landing page points at speed, at whether the page delivers what the ad promised, at how it behaves on a phone. A failed expected click-through rate on a commercial query is a reason to doubt the ad itself rather than the query. The overall score averages these signals and therefore hides them. Two keywords with the same score can be ill in different ways and need different treatment.

The same reading in Yandex Direct, without the score

Here the Russian angle stops being a footnote. Yandex Direct does not show you a number of this kind. What Yandex states is that the price and position of an ad depend on the bid, on the forecast click-through rate and on a quality coefficient. The coefficient is not shown, and it is not broken down. The forecast click-through rate is derived from accumulated statistics, which for a fresh account from abroad means there is little history to draw on.

So the three-component reading has to be assembled from other places. Expected click-through rate becomes actual click-through rate by keyword in the Report Wizard, compared across keywords of the same intent. Ad relevance becomes a manual check: open the search queries that triggered the ad and read whether the headline answers them. Landing page quality becomes behaviour in Yandex Metrica: time on page, scroll depth and goal completions for visitors arriving from that keyword.

That reconstruction is tedious, and it is the honest version of the work. The convenience of a single score is exactly what lets people skip the reading and jump to the legend.

One more difference is worth naming. A foreign company’s Russian pages are usually translations, and a translated page can fail the reading twice: the ad uses the wording people actually type in Russian while the page uses the translator’s wording, and the layout was built for another market’s buying habit. Both show up as weak signals, and both are fixed on the page, not in the account.

What the score does not promise

Now the legend. The common claim: the higher the score, the lower the click price, predictably lower, and improving the score is a reliable proportional lever on cost.

Checking this on data does not confirm it. If you take a cut across several different accounts in an analytics dashboard and put the Quality Score of active keywords next to their actual click price, no stable linear relationship appears. Keywords with a high score do not line up obediently into a cheap row, and keywords with a low score do not turn out consistently expensive.

The practical conclusion is hard: do not promise yourself or your client that work on the score will bring the click price down in proportion. The link between them, if it exists at all, is neither linear nor guaranteed. The click price in an auction is pulled by competition, by the position you are buying, by seasonality and by match type. The score is one factor among these, not the main lever. Before you build a promise on this assumption, check the relationship on the specific account.

In Yandex Direct the same caution applies with fewer illusions, because there is no score to point at. The click price for the same keyword moves with who else is bidding this week and with the position you are competing for. A company entering from abroad usually lands in a category where local competitors have years of accumulated statistics behind their forecast click-through rate, and it has none. Improving the ad and the page is still the right work. Expecting a proportional discount from it is the wrong expectation.

The same trap in a different pair of metrics

The mistake “two metrics are obliged to move together” is not specific to the score. People stumble in exactly the same way on the pair of impression volume and impression share lost to rank.

Intuition suggests an inverse relationship: impressions went up, so the ad is winning the auction more often, so the share lost to rank should fall. And the reverse. In practice, in one and the same campaign over different periods, both illogical pictures occurred. Impressions grew several times over and the share lost to rank rose rather than fell. And the mirror image: impressions dropped and losses to rank went down.

The cause is the same as with the click price. Both indicators depend at once on many independent factors: demand volume for the period, competitor activity and bids, ad relevance, seasonal swings on specific keywords. These variables move in any combination, so hard-wiring two metrics together and explaining one through the other is a methodological error. A rise or fall in impressions explains nothing and predicts nothing on its own, without an analysis of causes.

Yandex Direct reports its own version of these figures, and the temptation to read them as a pair is identical. A foreign team watching a new Russian campaign is especially exposed: in the early weeks impressions swing for reasons outside the account, as the forecast click-through rate settles, the negative keyword list is still being built, and seasonality in a Russian category may not match what the team knows from home.

How to test a metric instead of trusting its name

The general conclusion from both cases: a metric cannot be believed on its name. It is checked by comparison with data. Here is what an honest check looks like, on a different example, from work with audiences.

The task: decide whether a collected look-alike audience is good enough to raise bids on it. The name “similar audience” guarantees nothing. The check goes like this. Take the highest-converting segment in web analytics and look at the distribution of shares for each attribute separately: gender, age, region, device type. Then open the preview of the assembled look-alike audience and compare its histograms on the same attributes with the histograms of converting visitors. If the shares match, the audience repeats the profile of your buyers and it can be sent under an upward bid adjustment. If they diverge sharply, with the wrong region dominating, a skew in gender, a shortfall in the target age, then the audience was assembled from a mixed source and it either gets rebuilt or gets rescued with downward adjustments on the off-target segments.

The essence of the method carries over to any metric, including the score. Not “the metric has a good name, so everything is fine”, but “let us put the value of the metric next to actual behaviour on our own data and see whether the link holds”. A histogram, a cut, a pivot table: the tool is secondary. What is primary is that you test the claim instead of taking it on faith.

For a company running Yandex Direct from abroad this matters more: most advice you receive is second-hand, and you have no local intuition to weigh it against. Your own account data is the one check that does not depend on who is talking.

What to do with the score

• Read Quality Score as a diagnosis of the weak link in the chain of query, ad and page, not as a grade for your effort.
• In Yandex Direct, where no score is shown, assemble the reading by hand: click-through rate by keyword, ad text against real queries, page behaviour in Metrica.
• Do not promise a proportional fall in click price from a rising score: on data that link is not linear.
• Check the relationship between quality signals and click price on your own account before you build plans on it.
• Do not explain a change in losses to rank through impression volume, or the reverse, without an analysis of causes: competition, seasonality, bids.
• Test any metric, from the score to the quality of an audience, by comparison with actual behaviour, not by its name.

The score is useful exactly up to the line beyond which an indicator is turned into a promise. As a compass it tells you where to look: where relevance has sagged, where the landing page is weak. As a guarantee of a cheap click it does not work. The difference between “look there” and “it will be like this” is the difference between analysis and guessing.

If on your account it is unclear where a quality signal points at a real problem and where it is just an alarming number, come for a free review. How this fits into the wider system of paid traffic is on the Yandex Ads page.

Frequently asked questions

Sources

• Russian version of this article on belokrylovo.ru: Показатель качества: что Quality Score значит и нет