In 2022, programmatic advertising is rising to new heights in its value for marketing strategies. Emarketer estimates an increase in ad spending to almost $97 billion, compared to about $82 billion in 2021. With 54% of marketing budgets devoted to programmatically purchased ads, both publishers and advertisers need super-efficient strategies to make every dollar count.
For publishers, choosing an advertiser, and consequently, the revenue the publishers earn, are influenced by several criteria, like the price of impressions, ad relevance, and response time. The good news is that machine learning algorithms can help publishers make the best choices for real-time bidding in digital marketing and help them maximize revenue.
Let’s review the mechanics of real-time bidding from a publisher’s perspective and find out how machine learning can improve the process and increase revenue.
Real-time bidding and how ML enhances it
The point of real-time bidding (RTB) is selling the publisher’s ad slots (impressions) via real-time auctions. The bidding takes place in a third-party “playground,” or ad exchange. An ad exchange serves as an intermediary between publishers and advertisers, simplifying the interaction between the two parties engaged in the bidding process. However, publishers and advertisers do not interact with it directly; the publisher’s side has supply-side platforms (SSPs) and the advertiser uses demand-side platforms (DSPs).
When offering an ad slot, publishers set a floor price — a minimum price for each impression sold. Floor price optimization is vital to maximizing revenue. Bids below the floor price are immediately discarded. With a too high floor price, the impression remains unsold, and with a too low price, the publisher loses out on profit.
Another choice that impacts revenue is which SSP publishers choose to sell their inventory to an ad exchange because different SSPs charge different commission fees per impression. This choice affects the floor price as well. Next, when choosing an ad exchange, publishers have to decide whether to use a waterfall strategy or header bidding. The latter is more commonly used, but it has some drawbacks, including the possibility of missed opportunities. For example, not all demand partners offer a header bidding integration scenario. So, while you reach a number of them in header bidding, others can be reached through an ad exchange only. Thus, opting for header bidding only cuts publishers off from some ad exchange participants, which means missed potential for the publishers. To solve that, we have to create a hybrid system that combines header bidding with the waterfall strategy and ad exchange.
Header bidding offers publishers an alternative to Google’s monopoly that rules the digital ad revenue market. Add in Facebook’s Audience Network, with its walled garden, and the share of dominant players becomes so significant that you can’t avoid them without hurting your revenue. So why not combine header bidding and waterfall strategy to get the best of both worlds. A system that includes both elements provides more value — publishers can see the highest bids offered by the header bidding partners (HBP), and then send them to the primary ad server to compete against ad exchanges and SSPs with a waterfall strategy. When a publisher receives responses from ad exchange and SSPs, they can compare them to the highest HBP bid and choose the winner.
How does machine learning relate to all the mechanics of real-time bidding in digital marketing?
The decisions publishers have to make during this process rely on volumes of data and many variables. The floor price is just one of them, and others include ad timeout management, refresh time setup, the location, and format of an ad, and setup options between browser and server. RTB is a highly dynamic system. You can’t determine response time, demand distribution, or other variables in advance. Luckily, highly dynamic environments are well-suited to applying supervised and reinforcement learning (RL).
Applying reinforcement learning to real-time bidding in digital marketing
In a simplified RL scheme, we have a decision-maker who has to learn how to interact with the environment to maximize the long-term expected reward. A decision-maker takes action and gets time-delayed labels or rewards, and in this way, over iterations, learns how to operate in uncertain conditions.
That is to say, in reinforcement learning, we have to take actions in the observed environment to understand which actions work and which don’t. RTB has suitable schematics. Your task as an advertiser, for instance, is to reach the minimum winning bid for a particular slot, but you wouldn’t know whether or not you won with the lowestpossible bid. The publisher’s side also has several unknown variables that can only be found by taking action.
How does RL help in practice? With intelligent algorithms, RL helps publishers with:
Automatic selection of demand partners
Ad timeout management
Floor price optimization
Learning from historical data, intelligent algorithms fine-tune timeout settings, track KPIs, and analyze demand partners and ad requests, and they do all this in real-time. Let’s explore a few RTB-related applications of ML.
Want to know more about tech trends?
Sign up to be the first who receive our expert articles
Applying ML algorithms to solve publishers’ problems
From the publishers’ perspective, the key purpose of using machine learning is to optimize ad revenue. Algorithms deal with multiple variables in RTB to help address decision-making problems. Here are a few examples.
Floo price optimization in hybrid systems
We mentioned hybrid systems earlier, which are a relatively new approach to maximizing revenue by combining header bidding with waterfall strategy and ad exchanges. Publishers face some challenges with hybrid systems. They often have difficulty
predicting responses from header bidding partners based on historical data,
predicting the results of ad exchange auctions, and
predicting headier bidding partners’ highest bid and using it as a floor price for an ad exchange auction.
However, with the help of supervised learning, we can predict the highest header bidding bid using a least squares error regression method. The predicted highest bid gets sent to another prediction model to determine whether the ad exchange has a bidder that can outbid the HB price. If there is such a bidder, the system increases the floor price up to the point where it’s still outbid in the ad exchange. The HB bid can still be increased to the point of maximum floor-price optimization, where it remains outbid in the ad exchange auction.
Allocating multiple advertisements to an ad slot
Websites can have many ad slots that can render one or more advertisements, and publishers seek to fill the available space efficiently. Banners and media all have different sizes and bring different profits, and publishers strive to fill websites with the most suitable and most profitable ads without leaving any free space.
This creates a problem for optimizing real-time decision making. To solve this problem, we model the inputs (the ad slots) as a Knapsack problem, and select items from a given set to maximize their total value in the knapsack. However, we have a dynamic environment where the supply side might add or remove ads from the list of available candidates. To solve that problem, we can apply a neural network with function approximation and reinforcement learning.
Floor price optimization is part of a more generalized dynamic pricing problem, and solving it is crucial for sellers in different industries. Mathematical modeling can’t solve dynamic pricing since we have unknown demand distributions. And we have additional variables in our environment — available buyers, their willingness to buy, the quality of the products, and so on.
We can use deep reinforcement learning to adjust the optimal price for an item based on historical data gathered in the dynamic environment.
It’s worth briefly mentioning the advertiser’s side of the deal. One of the most important decision-making problems on the advertiser’s side is bid selection. Advertisers have to set optimal bidding prices to increase KPIs. The dynamic environment, budget allocation, and multiplicity of agents influence the auction environment and can complicate things. Reinforcement learning can meet the majority of constraints in bidding strategies, allowing advertisers to target ads to users’ interests and account for other bidders’ influence on the market.
ML addresses RTB challenges
In programmatic advertising, publishers have to consider criteria like the choice of strategy, floor-price optimization, response time, and ad relevance if they want to increase their revenue. RTB is one of the go-to programmatic approaches, but it has its own constraints. A highly dynamic environment, which means a significant volume of data, many black boxes, and variables make decision-making challenging for publishers and advertisers.
Luckily, when it comes to digital marketing, real-time bidding is an excellent opportunity to apply deep-learning techniques. ML-based solutions help publishers reduce the guesswork and increase their revenue with optimized RTB.
And the best part is that Postindustria has the necessary expertise to develop ML that will benefit your business. Contact us and find out how we can help you maximize your ad revenue.