Do you know how ads are selected and served in ad blocks? Read more in the help section.
We regularly receive questions from users about which factors affect click price, which do not, and how these factors are connected. To dispel some of the confusion (and myths) around this topic, we decided to explain the key terms from the display rules and the interface in two handy cheat sheets.
Under the hood
The CTR forecast, quality coefficient, and placement score are calculated using Matrixnet machine learning technology, meaning that you can immediately see how changes you make to your campaign affect its performance without waiting for new statistics to compile. Machine learning algorithms quickly recalibrate using the most recent data and offer a relevant price. Matrixnet’s role in pricing also means that the factors mentioned cannot be found in the interface, so any explanations of how Yandex.Direct algorithms work are based on generalizations and simplifications.
Machine learning is a universal term describing the use of algorithms to process large amounts of data. In an ideal world, we would be able to give a simple answer to questions about how algorithms work, e.g. “your quality coefficient is 42.” The truth is that algorithms are much more complex — they can process data available only at the very moment an impression takes place, like the characteristics of a specific user about to view an ad. If we were to average the quality factors from the campaign and show them in the interface, the resulting numbers would be of little use given how far they would be from the ones actually used in calculations.
When optimizing ad campaigns, it is vital that quality assessments use clear and understandable measures that can be quantified and shown to the machine as an example — clicks, conversions, ROI, installations. That way, the algorithms can handle the rest by themselves.
Imagine that someone gives you a huge list of facts and leaves you with the task of finding the patterns within them. This is exactly how machine-learning algorithms operate: they discover connections that we ourselves would never guess existed and apply them when determining the best ad to display. There are hundreds of factors that can influence each individual impression, from the weather and the time of day to the presence and order of certain words in the ad. Their combinations vary for different users, as do their sensitivity and value ranges. While this may sound exceedingly complicated, the significance of machine learning is that you don’t need to track, weigh, and calculate these factors manually — the algorithms themselves can tell what is important in the data provided.
You don’t need to know all six steps involved or look at the conditional parameters to create a successful ad campaign. What’s most important is that you understand the main idea.
The Main Idea: Prices are affected by the attractiveness of the ad, its relevance in relation to user search queries, and its placement history. In other words, users should like the ad and not be disappointed after clicking on it, and it is in the interests of sites to host high-quality ads. Just a few rules are enough to keep things running smoothly.
Historical and forecast CTR
The typical, simplified version of how CTR is calculated involves the ratio of clicks on the ad to the total number of impressions, and is usually expressed as a percentage. This ratio is what you see in the statistics for your ads’ impressions and clicks.
It would be unreasonable, however, to compare a CTR of 20% for 10,000 impressions and 2,000 clicks with the same CTR for 10 impressions and 2 clicks. It would likewise be unfair to assign a CTR of 0% to ads that have just begun displaying — after all, a new advertiser can create great ads that do an excellent job of responding to users’ needs. With a CTR of 0%, the cost per click would be end up being excessively high.
This is why Yandex.Direct uses forecast CTR to calculate prices for advertisers with different traffic volumes. Furthermore, forecast CTR also helps to decide the block and position in which ads will appear. Determining CTR is a job for Yandex’s Matrixnet machine learning system.
These calculations largely rely on past click and impression statistics for a given keyword, but also consider hundreds of other parameters that influence click probability. These include the presence of keywords from the search query in the ad text, nested queries, user behavior in the selected display regions, seasonality (an important factor for many topics), as well as indicators that can affect the level of trust users have for your domain. Extensions like vCards, sitelinks, and callouts also play a role. Even just filling out the vCard has a positive effect on CPC.
! There is no point where Yandex.Direct begins considering only historical CTR — why use historical statistics when you can immediately make a new forecast based on recent changes and dynamically adjust the recommended bid?
Quality coefficient: assessing user satisfaction
CTR, whether statistical or forecast, is an indicator of an ad’s attractiveness. What CTR does not do, however, is help us understand to what degree the landing page matches the ad text and original search query.
If we were to simply arrange all ads according to clickability, the best positions would most likely go to “attractive” ads that actually mislead users. This is where the quality coefficient comes in — it accounts for the relevance of the ad, landing pages, and domain in comparison to the user’s query.
Let’s use some simple examples to understand relevance. The fullest and most relevant results for the query “Lancôme lipstick” will offer the official Lancôme store, followed by other sites offering this brand, and in last place, lipstick produced by Lancôme’s competitors.
Branded queries are relatively simple to understand, but what about unbranded ones? Let’s use “Japanese stainless steel kitchen knives” as an example query. Potential competitors could be advertisers with campaigns for keywords like “kitchen knives,” “Japanese knives,” and “stainless steel knives.” None of these sites are fully relevant for the user who entered the original search query. The quality coefficient will be highest for the sites that best reflect the user’s interest.
The quality coefficient is partly based on the measures used to rank search results. Yandex.Direct uses the quality coefficient to give the positions and reasonable prices to the most effective ads so that users will be willing to trust the ads they see.
! Even after you improve your keyword list and raise your ads’ CTR, you can lower your CPC even further by developing the connections between your keywords, ad texts, and landing pages — and by clarifying the landing pages themselves.
At the same time, there is no reason to worry about fluctuations in the quality coefficient or to exaggerate their effect on your campaign. If users click on your ads and don’t leave your site, then everything is fine. When selecting and ranking sites, Yandex.Direct relies on the combination of your bid, forecast CTR, and the quality coefficient. Making changes to one parameter cannot make the results significantly worse.
Placement score
A broad assessment of your domain's placement in Yandex.Direct affects the minimum CPC without being linked to a specific account (if an ad has no link, statistics are collected for the vCard’s phone number).
The placement score includes the combined CTR for all the domain’s campaigns, their duration, and other indicators. Algorithms select ads with greater accuracy and objectivity than any human could. If the campaign’s “domain karma” is really bad, Yandex.Direct will signal this by raising the minimum CPC for search.
Simply switching the domain without any other changes is only a temporary fix. The entry threshold can decrease to as low as 0.01 Yandex units as the ad’s effectiveness increases (the amount of time necessary to see changes is determined by how often your ads are displayed — the more frequent the impressions, the faster you’ll notice improvements).
* Why don’t I see the factor’s influence in my experiments? This depends in large part on the quality of the experiment. The most common sticking points are when tested ads’ accumulated statistics and settings differ significantly or, conversely, when no data was available — themes without competition, landing pages never advertised before, or rare search queries. In other words, Yandex.Direct lacks the information necessary to understand which one of a group of nearly identical ads is more effective. The better the algorithms can analyze a specific theme, the more accurate individual forecasts will be. As your ads are displayed and you continue to optimize your campaign, the responsiveness and relevance of the feedback you receive will increase.
Lastly, let’s take a look at a few features of Yandex.Direct likely to raise questions.
Keyword effectiveness and the account quality indicator
Productivity and the account quality index are often applied in the price setting process, but they do not in fact influence prices. Both indicators give an idea of how the campaign is going and can be especially helpful for advertisers just starting to use Yandex.Direct.
Productivity shows how well a keyword is processed in relation to a specific ad. Low productivity typically indicates the usual mistakes that new advertisers make, e.g. overly narrow keywords and absence of the keywords in ad texts or titles. The account quality index, on the other hand, helps advertisers understand how they can increase the output of their campaigns — add images or sitelinks, and generally keep campaigns “fresh.”
In the next few articles, we will describe possible uses of neural networks in search and advertising.