Attribution modeling is used by advertisers to determine the value of different channels on their marketing, which helps them understand which channels provide the most benefit to their marketing campaign, and thus make better decisions about using them in the future.
In the case of online advertising, the task is fairly simple. When it comes to offline advertising, such as TV, the task is much more complicated. Getting actionable metrics and insights is a challenge, due to the media break between TV and the web.
In this article, an approach is presented that enables TV modeling to effectively allocate investments in TV advertising.
Television advertisements lead some multitasking users to take immediate, measurable actions online, such as the search for specific keywords and brands or website activity. These actions generate spikes in the online response volume in the period following the ad airing. Analyzing these spikes by computing the difference between the actual traffic and an estimated baseline can be used as a measure for the impact of the ads. In the lack of validation data, it is difficult to assess the performance of any baseline estimation, which makes it difficult to choose a proper algorithm for this task.
In my master thesis, I used minute-level website activity data to build a synthetic dataset. The latter was built upon a predefined baseline, to which noise and spikes were introduced. This dataset was used to compare the performance of algorithms including the Continuous Wavelet Transform (CWT), the Discrete Wavelet Transform (DWT) with different Thresholding techniques, and a Moving Average approach for the baseline estimation. The results show that DWT with thresholding gives the best results and it was adopted in our Scope TV attribution model.
In the case of online advertising, the task is fairly simple, since various metrics can be accurately tracked with each campaign, including impressions, clicks, conversions, and more. Tracking makes it easy to know how a campaign is performing while it is still running, and therefore allows for optimization and adjustment to improve it. When it comes to offline advertising, such as TV, the task is much more complicated. The challenge with TV ads is that marketers have difficulty getting actionable metrics and insights, due to the media break between TV and the web. However, despite the unmatched ability of digital mediums to deliver personalized messages that are easy to measure, it is expected that TV will continue to be a significant part of an effective marketing strategy, given that it is one of the most efficient ways to create widespread audience awareness. Therefore, it is of high interest to advertisers to find ways of measuring TV advertising impact.
While it is relatively simple to measure TV ad recall and impact on attitudes, usually by means of surveys, it is often difficult to measure its behavioral impact, such as the actual sale or actions that might lead to a sale. The survey responses present several challenges:
Moreover, an ad seen by a large audience is not necessarily successful if it fails to achieve a predetermined quantifiable objective. In contrast, search-engine queries have become a key part of the purchase process for many consumers and could represent a valuable outcome measure as an indicator of potential purchases for advertisers.
Many studies about media multitasking have shown that people are generally online or in the proximity of a second screen device while they are watching TV, which results in a new advertising effect. Nowadays consumers can easily obtain more information on an advertised product by searching for more information online while watching TV. Zigmond & Stipp published the first case study in which they show how TV commercials can trigger search spikes for the advertised brands. In their study, they used indexed search data from Google Insights (which was merged to Google Trends later on), along with ad airing data, and were capable of showing that TV advertisement can lead to a significant increase in online search. They conducted five case studies. In one of them, they studied a new ad for a car in the 2008 Olympics broadcast. The effects on minute-level search queries following the ads were both immediate and drastic, with very sharp spikes in search volumes, as displayed in figure 1.
Figure 1: Search Queries for “Chevy Volt” during Beijing Opening Ceremonies
Search data show clear causal effects within minutes or seconds of ads airing, while purchases due to advertising are typically recorded within days or weeks after exposure to the advertising and are therefore difficult to link directly to the commercial. These online interactions, correlated to ad exposure, can be used as key data points for TV ads evaluation. Estimated uplift generated by these ads could be used for making recommendations for TV budget reallocation across network, daypart, ad length, and spot position in order to optimize ad effectiveness.
An ad displayed on TV may lead some multitasking users to immediately search for the advertised product using their second screen device. Aligning the TV ad schedule and the aggregated search data can be used to compute the incremental searches that can be attributed to TV ad spots. Current TV advertisement is not personalized to conduct randomized experiments, that could be used to compute this uplift. Therefore, the baseline search volume should be estimated and deducted from the actual search volume in order to obtain an estimate of the incremental search generated by TV ads, considering a period from 5 to 10 minutes after the ad airing referred to as the attribution window, in which the impact of the ad is visible. Figure 2 shows an example of spikes generated in online activity, in a period following TV Ads airing.
Figure 2: Online activity uplift (light blue: baseline, blue: online activity)
The process of TV attribution based on online activity can be divided into three main steps, summarized in figure 3 and detailed afterward:
Figure 3: TV attribution process
The first step in the process is data acquisition and preprocessing. Two main inputs are necessary for this step, namely the web traffic and TV spot data. It is better to have high-frequency data, which enables to clearly visualize the impacts. For the TV data, the following information is necessary:
For the online data:
Some preprocessing might be necessary, such as geographical filtering of the traffic, by retaining the traffic of the country or region of interest, which are relevant to the ads being measured. In addition to that, it is also possible to exclude responses that are driven by other sources (e.g. visits from an email campaign, clicks on a banner advert), thus only keeping direct entries and search entries (organic search as well as paid search). TV spots data should be precise and include all the channels where the ads are aired. Finally, it is important to define an attribution window, which will be used to compute the online activity uplift.
The baseline estimation is the most important step in the process since it is used to estimate the counterfactual: what the traffic volume would have been if the ad was not aired on TV. Much of the online activity is not generated by TV advertising, there are likely many other sources of traffic, including digital advertising, email campaigns, radio, natural search and word of mouth, to name a few. In order to fairly determine the impact of TV, we must first eliminate activity not driven by TV, which is represented by the baseline. It should be a continually moving and adjusting curve that produces a separate value for every minute of every day in response to marketing activity, time of day, seasonality and other factors.
Getting attribution right is extremely difficult, even more so for mass media such as TV. Some solutions rely on a Moving Average approach, which is a calculation used to analyze data points by creating a series of averages of different subsets of the full data set. However, this approach leads to a flat baseline, and thus a lack of adaptability to short-term fluctuations of visits, leading to unreliable results. Therefore, more sophisticated algorithms are being used for this task.
In our model, we use the “Discrete Wavelet Transform” (DWT), a powerful signal processing technique that translates a signal into various frequencies that might change over time. DWT uses finite wave functions called “wavelets” that oscillate in a limited time window, to decompose the signal into different signal components with different frequencies. The traffic induced by TV ads is represented by the high-frequency (short-term) spikes. Filtering these spikes from the signal is used to obtain the baseline.
Figure 4 shows the results of the baseline estimation using DWT, applied to the synthetic dataset. DWT has the ability to filter noise and spikes in the data, and from the graph, we could see that the estimated baseline is close to the initial synthetic baseline.
Figure 4: Synthetic website data (gray) synthetic Baseline (blue) Estimated Baseline using DWT (red)
Once the baseline is computed, it is possible to estimate the traffic volume generated by the TV ads, by calculating the difference between the actual traffic and the estimated baseline. In the case where several spots are linked to the same uplift, the latter could be distributed based on the spot’s reach for example. Further analysis could be done at this point to understand which factors, such as channel, time of the day … (etc), lead to the most uplift.
To conclude, even though the online response following TV ads could not substitute the final conversions or sales, they are definitely a good indicator for the interest and engagement of consumers for certain products, and could be used to more effectively allocate investments in TV advertising.