Open web measurement

Corina Alonso Updated by Corina Alonso

Open web measurement: Audience measurement of open web campaigns

AudienceProject uses advanced technology and robust methodology to measure campaigns running on the open web. By using deep learning and probabilistic modelling in combination with first-party panel data, AudienceProject provides audience reach and frequency insights for open web campaigns in a cookieless and privacy-safe manner.

Today’s media industry is increasingly defined by walled gardens, different ID universes, fragmented media consumption and privacy. This challenges marketers’ ability to measure and analyse campaigns holistically via cookie-based measurement.

AudienceProject overcomes this challenge by using three different enablers; direct clean room integrations, advanced technology and robust methodology.

When measuring campaigns on walled gardens like Meta and YouTube, we utilise our direct clean room integrations with those platforms. However, for open web campaigns, such integrations are not possible, and thus, we use other means to measure these. More specifically, we use deep learning and probabilistic modelling in combination with first-party panel data to provide audience reach and frequency insights for open web campaigns. 

How it works

Our general measurement strategy is to use and combine all relevant data for the specific measurement purpose. In the case of open web, this means combining information from loglines with information derived from our measurement panel.

Raw log-level data delivers an abundance of potentially useful information - even when completely stripped of personally identifiable information (PII). Based on raw log-level data such as geographical locations, timestamps, user agent data, etc., we initially use our deep learning model to estimate how many devices are reached by a campaign. 

To ensure that we deliver precise device reach estimates, our deep learning model is trained with high-quality and high-volume data from our historical campaign measurements and online behavioural targeting. At the same time, it is constantly validated by taking a critical approach to the outcome of the model, ensuring that it is continuously fine-tuned by incorporating new learnings. 

When we have calculated the device reach, we combine this with our knowledge about the“device universe”  and geographical information as well as demographic information.  derived from our consented and representative measurement panel. This allows us to estimate how many humans are reached by a campaign and thus provides insights into audience reach and frequency. The profiled reach is then modelled based on our consented and representative measurement panel.

As the input of the measurement is aggregated data, the measurement is based on groups, not individuals, ensuring that it is done in a fully privacy-safe manner.

Measurement of audience reach and frequency for open web campaigns:

 

What is deep learning?

Deep learning is a framework of machine learning methods that utilises flexible mathematical models that contain a very large number of free parameters, enabling modelling of a wide range of phenomena. Systems utilising deep learning have successfully solved many challenges, including image classification, fraud detection, language translation, and even art creation (the latter to a debatable degree of success).

This can sound a bit abstract, but let’s try to simplify it with an example from ‘the real life’.

Imagine you are given the task of telling how many people are visiting a public park on a given day. The only limitation is that you are not allowed to count the number of people. This means that you need to look for other traces that can help you estimate how many people are visiting the park.

For example, you can look for how many to-go coffee cups are in the bins, how many footmarks are on the trails and how many marks are in the grass after picnic blankets. Using these traces will help you get an idea about the number of people visiting the park, and with the presence of historical data, you can even validate and readjust the estimation. Maybe you learn that only half of the park visitors usually bring to-go coffee or that people normally do picnics when the weather allows it.

In the same way as you are looking for traces and using your prior knowledge to estimate the number of people, AudienceProject uses deep learning to estimate the number of devices reached by open web campaigns. 

What is probabilistic modelling?

Probabilistic modelling is a type of statistical modelling that incorporates probability distributions to account for uncertainty when drawing conclusions from the data. Combining prior knowledge and stochastic variables, probabilistic modelling can predict the most likely outcome of an indeterministic process. This indeterminism can stem from truly random or unpredictable events or real-world facts that are just unknown in the model.

What is statistical Extrapolation?

In many instances, we do not have access to a direct measurement of the objects we want to know about. Instead, we look at patterns, connections and correlations for similar objects we can observe and extrapolate to the objects we want to know about using the assumption that the patterns we have seen for the objects we extrapolate from are the same as the objects we extrapolate to.    

To validate our measurement methodology, we have compared campaign measurement results from hundreds of campaigns across our markets where our traditional cookie-based measurement methodology and our new measurement methodology based on deep learning and probabilistic modelling have been used.

The graphs below show the reach build-up over time for two of the campaigns where we have made the comparison. The grey line represents the reach build-up for a campaign measured with our traditional cookie-based measurement methodology, whereas the blue line represents the same campaign but measured with our new measurement methodology.

The graphs show that the reach calculations based on our new measurement methodology are very close to those made with our traditional cookie-based measurement methodology.

Reporting and benefits

Reporting

  • Metrics: Reach in target group, frequency, hitrate, on-target percentage and events in target group
  • Reach building event types: Impressions, viewable impressions, clicks and video quartiles
  • Demographics: All demographics (gender, age, income, employment, education, household size, children in household)
  • Reporting period: 84 days

Benefits

  • Get independent measurement of open web campaigns
  • Understand the total, de-duplicated reach generated by campaigns on open web and other channels 
  • Understand the incremental reach generated by open web campaigns

How did we do?

Meta measurement

YouTube measurement

Contact