Since the first online advertising content appeared, the online advertising market has revolutionised. Today advertisers can perfectly match content to their potential customers. And they can do it with the lowest possible cost. That’s all thanks to machine learning algorithms. How do they work, and how to achieve the best results from advertising campaigns? In this article we will introduce you to the Real-Time Bidding topic. Then, we’ll describe the machine learning models that are revolutionising today’s online advertising market and can do the same for you!
Online advertising — history
The beginnings of online advertising — the traditional model
During the beginning of the Internet era, buying online advertising was a whole process. It began with a relationship between publishers (selling advertising space) and advertisers (buying this space to broadcast their ads).
This was an as laborious process as placing ads in traditional media. The advertiser had to decide: in which magazine, on which page, and at what time the ad will appear. All those steps required a lot of work. Firstly, advertisers had to choose the best place and time for each advert. Secondly, they needed to set it up. Finally, they were analysing and comparing the effects of the placement between different publishers. Moreover, with this approach, the same adverts were displayed to all users, regardless of their gender, age or interests.
As the number of publishers grew rapidly, the problems with communication and choosing the best offers appeared. Advertisers had to negotiate with thousands of publishers at once to reach the largest possible audience of potential customers. Publishers also had to contact thousands of advertisers to see if anyone was interested in their inventory (advertising space). The system limited the opportunity for publishers with small reach and minor and very niche audiences to make money. At that time, some agencies acted as intermediaries between advertisers and publishers.
Over time, people noticed that more important than the place of publication of an advertisement should be the audience it reaches. Therefore, they created ad networks that categorised a publisher’s unsold inventory to make access and inclusion in media campaigns easier.
Ad networks divided advertising inventory into two categories: premium and residual. Premium ads were sold through direct relationships between buyers and sellers, while the second category included ads that remained unsold. This solution allowed smaller advertisers to gather into groups and sell their advertising space even to prominent players who probably would not choose to contact them directly.
Supply Side Platforms and Demand Side Platforms
Ad networks created an environment in which agencies had more than one channel for acquiring advertising inventory, but they also made the ad buying process more complicated for publishers.
So, they created Supply Side Platforms (SSPs) to solve this problem, enabling publishers to sell advertising space on Ad Exchanges and manage who gained access to inventory. SSPs work as an intermediary between the seller and the ad networks, helping publishers maximise the revenue generated through ad networks. In addition, they give publishers more control over advertising inventory and can influence how it is delivered to the ad network.
At the same time, Demand Side Platforms (DSPs) appeared. They help agencies and advertisers win advertising space by allowing them to manage media buying through a single platform.
Over time, DSP and SSP technologies have evolved to create a Real-Time Bidding (RTB) infrastructure that integrates these two platforms.
What is Real-Time Bidding?
Real-Time Bidding (RTB) is a method created in 2009 to automate the process of buying and selling online advertising in an auction model. In short, it’s an extensive ecosystem for efficiently acquiring adverts and displaying them to the right people.
This model is focused on displaying a specific ad in a particular location. The advertisers bid in real-time for their ads to be displayed. When they win the bid, their ad creation is immediately displayed on the publisher’s website.
RTB is a solution that increases the scalability of digital advertising and provides better audience targeting. RTB benefits publishers, who can sell their inventory at the highest possible price, and advertisers, who can optimise their actions to increase the reach of their display ads to their potential customers.
How does Real-Time Bidding work?
When a user visits a website, a relevant bid request is immediately sent to the Ad Exchange. The Ad Exchange or another component of the system (DSP or SSP) organises a bidding process. Advertisers bid in real-time for their advert to be displayed on the website that the user is currently visiting. Finally, the advertiser who outbids the others with the highest amount – wins the auction.
The decision to match advertising to a particular user is based on some key information about that user, which is collected at all stages of the process by all players involved. The entire process, including sending requests, bidding and displaying an ad, takes just 100 milliseconds.
The Real-Time Bidding process:
- Ad Exchange, using SSD, sends information about page content and users to the advertiser.
- Advertisers bid on these displays, and the highest bidder wins.
- SSDs automate the bidding process and make it easier to target ads to the right users.
Types of auctions in the Real-Time Bidding model
- First price offer — a model in which the buyer (advertiser) pays the same price he bid for a given ad impression;
- Second price offer — a model in which the advertiser gives the highest amount he can pay, but if it turns out that other bids were much lower, he will pay less.
Machine learning models improving the effectiveness of Real-Time Bidding ads
To make ads as effective as possible, Machine Learning companies create advanced models that can improve bidding in the RTB model, impacting the effectiveness of the displayed ads. Below you will find the top 5 models that can have a significant impact on online ads optimisation.
1. Product recommendation models
Detailed user data, including a person’s tracked activities (e.g. age, demographics or interests), is collected in a huge database called Data Management Platform. Then, based on predictive analytics and a user’s past behaviour, machine learning models are able to predict which products a particular user will have the deepest interest in at the moment.
Let’s look at an example: Based on the search history, we know that our user is a middle-aged man who built a house this year. Thanks to using product recommendation models, the ads displayed to him will be related to house decor and renovation such as wall paint.
2. Click-through rate prediction models
Based on a user’s tracking history, machine learning-based models predict the likelihood that a person will click on a displayed ad.
For example: Our user is a young woman who has recently bought a dog. Suppose her tracking history will show that she is much more likely to click on ads for dog toys than this season’s fashionable dresses. In that case, machine learning models will display her pet-related ads more often than clothing promotions.
3. Models for purchase prediction
We have to remember that the advertisers pay for the effect of their ads according to the chosen pricing model (e.g. CPC, CPO, COS/ROAS, CPS). However, if the advertisement brings low returns, i.e. a small number of users take the expected action, it will not be profitable for the advertiser.
Purchase prediction models can be a solution and improve the ROAS (Return on Ad Spend) by predicting who will be interested in buying the advertised product.
4. Look-a-like models
Machine learning models divide user data into groups with a similar behavioural profile to improve target advertising even more. This way, they can predict how a particular user will behave based on other people’s behaviour with similar profiles.
5. Models predicting the probability of winning bid auctions
Advertisers usually pay a specific price per click or other user action. The problem is, if we offer a too high price, it may be not enough to cover all impressions. On the other hand, if we offer too small a prize, there is a risk that others will outbid. Therefore, our ad will never be displayed.
What is the solution? We won’t surprise you because the answer is machine learning models. They can predict the so-called rent scape, i.e. how much other advertisers are willing to pay for a given customer. This solution significantly optimises advertising campaigns and minimises the risk of overpaying.
Real-Time Bidding in YOUR advertising strategy
Would you like to maximise the effectiveness of your ads using advanced machine learning models?
➡️ Contact our experts. ⬅️
We have a complete set of tools that will revolutionise your existing or planned advertising campaigns.