When most SEOs think of clicks, they think of click-through rate as a ranking factor and optimizing page titles and meta descriptions. If that’s the extent of your involvement, then you’re leaving money on the table. Let’s look at how click models work and their impact on search engine optimization.
What Are Click Models?
As you work your way through this primer on click models, keep in mind the concept of clicks, attention, and satisfaction. These are three important attributes that click models are attempting to understand. I’ll talk more about them later in this article.
For now, let’s start with a simple definition. Click models are predictive probabilistic data models of search user behavior that aim to predict future user patterns by analyzing historical data. So to create a click model, you examine past data to determine the likelihood of certain patterns occurring in the future. But since no one can predict the future with 100% certainty, these models can only look at the probabilities of certain events occurring.
Basic Click Models for SERPs aim to understand the examination process of each link presented and understand the biases of other elements on the page — things like position, ads, and other features. The goal of modern search engineers is to describe click models in a portable, unified way to then be able to relate them to each other and use them in ensemble approaches.
How do Click Models Work?
Click models help search engines with relevance estimation – retrieving relevant URLS from a giant corpus. Here’s a basic overview from Chao Wang at Baidu, et al., Building a click model: From idea to practice.
In their paper the authors explain how “most existing click models are formulated within the framework of probabilistic graphic model. In these models, a group of variables are usually used to model each search result for a specific query.” These variables include:
- Observable click actions
- User examination
- Result relevance
- User satisfaction (after viewing the results)
Every model creates a structure to represent how the variables are related to each other, based on differing user behavior assumptions. Once the model is created, it’s trained on a large set of click-through logs, after which it can be used to “predict click probabilities for results or to rerank the search result list according to the inferred relevance.”
Two models are presented in this paper. The first is the Vertical-aware Click Model (VCM) which addresses the problem of behavior bias when vertical results are combined with ordinary ones. Those biases include:
- Examine bias for vertical results (especially those with multimedia components)
- Trust bias for result lists with vertical results
- Increased probability of revisiting vertical results
The second model addresses a bias in most existing click models — they believe users follow a linear top-to-bottom approach in examining search results. In reality, many studies reveal a large amount of non-sequential browsing (both examination, and click).
An Ensemble Approach to Click Modeling
So far we’ve looked at relatively simple and singular models for click prediction. Here’s an ensemble approach example from the genius minds of Danial Bakhtiarvand and Saeed Farzi – An Ensemble Click Model for Web Document Ranking.
In this case, they’re using three modules simultaneously to predict users’ clicks on the SERP:
- Probabilistic Graphical Models to determine click behavior
- Time-series Deep Neural Click Model to predict users’ clicks on documents
- SimRank Algorithm to predict similarity in a graph of document-query relationships
The outputs from the modules is fed into a multilayer perceptron (MLP) classifier, which in turn predicts how likely a document will be clicked by a user. The advantage to this approach is that using multiple models is more robust. It negates any bias that could arise using the single-model approach.
Here’s more from Saeed Farzi because he’s awesome. Make sure you check out the query auto-completion and suggestion review if you need a detour.
The Impact of Click Models on SEO
Why does this matter for SEOs? Understanding that click models exist and how they work relates to Information Retrieval and how systems are designed. Click models improve search quality by correctly interpreting user clicks.
Click models help with:
- Understanding and simulating users
- Determining document relevance
- Evaluating search
How does this relate to Google? Here’s one of the best tutorials about how Click Models can be applied to Google Search by Chuklin, Markov, and de Rijke. (Google Switzerland – 2016). Their website has a collection of materials related to studying and using click models, as does their book, Click Models for Web Search.
The CAS Model – Clicks, Attention, and Satisfaction
Now, let’s look at later work on the CAS model — Clicks, Attention, and Satisfaction. Basic click models don’t account for modern SERP layouts and that the user examines the results in a very complex way, not top to bottom.
Early evaluation models had trouble accounting for the cases where a user gained value and experienced information gain directly from the SERP like a currency conversion query, which yields nearly zero clicks and has nearly 100% satisfaction. Other activities like abandonment and mouse movements were also not included in early models.
User perceived rank is different than actual organic rank and those need to be calculated differently as well. If you haven’t already realized your rank tracking isn’t enough, stay tuned.
Attention (examination) models assess time spent on the SERP, mouse patterns, geometry, CSS/SERP Feature structures, real rank (think VED), and perceived rank are needed. Satisfaction models are how we validate behavior against possible satisfaction.
This may be a good time to revisit Google’s QRG, either through our post, Google Search Quality Rater Guidelines: How Google Evaluates Your Site, or by reviewing the 172 pages of Google’s Search Quality Evaluator Guidelines.
The CAS model authors suggest we should “move away from the ten blue links approach & adopt an evaluation metric that uses rich features and relevance signals beyond traditional document relevance.” For those that debate this, by this we mean the ’10 blue links approach,’ it really isn’t debatable.
So what else do Click Models do? User simulation.
Click models decompose click probability into the properties of a document like attention, attractiveness, and satisfaction parameters. They also identify the impact of document rank, surrounding documents and features.
What more do Click Models do? User interaction analysis.
With parameters of a click model, we can QUANTIFY how document placement and structure changes consumption of SERP and introduce recommendations; e.g., “News packs only impact their neighbors when placed at the top.”
Importantly, click models enable the inference of document relevance. If click model-driven relevance is used as a learning feature in an ML-based ranking system, it significantly improves the end result.
This means that it is a powerful signal of document quality.
That’s why you need to care about click models. Want to know more? Buy this book, Click Models for Web Search.
MarketMuse sets the standard for Content Quality. Get that, too.