Click here to view this e-mail in your browser
Email WhatsApp Twitter
Business case studies No 260 - 10.05.2019

Can a robot pick up on what makes a TV spot award-winning? And can that robot help advertisers make other ads as effective as the winner? With these questions in mind, the Dutch sales house Ster embarked on an ambitious project two years ago - a new endeavour to automate their ad testing and see how machine learning can be applied to this.

I, Robot

The project culminated in Ster’s AdScan, which was made available to advertisers in November 2018, free of charge. AdScan is an advertising valuation tool that uses machine learning to combine historical panel data, computer vision and smart algorithms to produce an analysis of a spot’s creative value within twenty minutes.

Unlike people - who base their appreciation of an ad mainly on unconscious emotions - a robot can explain in more detail how it came to its assessment. This gives advertisers insights into how a commercial is valued and, more importantly, why.

We set out to make a model that could give a solid indication of how a panel would react to a commercial. Due to the explorative nature of machine learning, we gained a lot of new insights as well. I’ve never heard a panel member say, ‘I thought the commercial was boring because of the lower average scene length.” says Rick Hoving, Project Leader Strategy and Innovation at Ster.

Rise of the Machines

While machine learning is generally used to discover patterns within large amounts of data, AdScan is the combination of computer vision (recognising objects in photos), historical panel results and machine learning algorithms.

Simply put, AdScan recognises for example in a video that the colour red is often used and looks at how the panel has responded to this in the past. The model also takes into account the relationship between red and other factors. The model is initially trained and then blind tested for new data. AdScan, therefore, copies the behaviour of a panel that is watching an advertisement for the first time.

Ster does not have its own panel, but cooperates with research agency SAMR. The database used for AdScan is based on 950 pre-tests of television spots. For every spot, Ster recruits 100 respondents from the SAMR panel.

AdScan values a spot on six statements and compares them to the sector's advertising benchmarks. An advertisement can score "worse", "average" or "better" than the benchmark. In just over 75% of the cases, AdScan's prediction corresponds to the opinion of a panel.

The six questions which AdScan answers are whether an ad is fun to look at, original, cheerful, entertaining, appealing or boring. Moreover, AdScan copies the behaviour of a panel and is able to provide a good indication of how a panel will judge a spot - all within a much shorter time span.

AdScan makes its assessment based on around 90 elements, including objects, colours, music, metadata, emotions and people. For each element, AdScan indicates its importance and provides advice for improvement. With AdScan, advertisements can be improved in a much more targeted way, by providing advice on whether less text should be placed over the video or if a certain colour should be used more.

For example: AdScan recognises written characters on the screen but does not analyse them further than the fact that they are present. The only word that AdScan recognizes is the name of the advertiser If this is spotted, this is registered as a "brand mention".

AdScan also recognises music combinations, such as symphonic rock. Users of AdScan can see how much the music in the advertising resembles a certain genre and its influence. In general, rock, metal and classical music attract attention, make advertisements less boring and more entertaining to watch.

”There’s still a lot to improve upon. We’d like to be able to predict more statements from our original commercial tests database. Currently, we predict 6, but we have about 80 in our database. We’ve been achieving good results with some others, but not at the level yet where we’re confident to use them. Personally, when talking about adding extra elements to our analysis, I’m most looking forward towards detecting humour, but also to automatic logo recognition.” says Hoving.

Towards the future

Is it now advisable for every brand to advertise with the colour red and as little text as possible in the picture? Not quite – considering that the elements need to respond to each other. To come to real advice on how a creation can be better adapted, a specific report is more meaningful than the generic results as mentioned above. They are basically all true, but on their own, they will not turn a moderate scoring into a good scoring advertisement. Machine learning is a useful tool to test creations free of charge and quickly and to demonstrate the effect of subconscious processes that respondents do not quickly cite.

“I’m a strong believer of new technology’s ability to increase the accessibility of services only previously available to the initiated. Why shouldn’t quality research be available to all businesses we work with regardless of budget and size? We’ve seen companies, which have never advertised before, using AdScan and adjusting their commercials, accordingly, improving the effect they are aiming for. But even larger companies use AdScan to test their storyboards to get a general idea of what they’re going for is working.” says Hoving.

But human appreciation changes as cultural beliefs shift, so the panel remains important to feed the algorithms with these changes. AdScan is therefore a welcome addition to Ster’s spectrum of advertising valuation tools.

“The dream is to eventually have a tool that not only analyses videos but also improves upon them based on its results and returns a mock-up for inspiration.” concludes Hoving.




» Website STER - click here


» egta CEO & Top Exec's Summit (TV & radio) 06-07.06.2019, Prague Register here

» egta VOD Masterclass (TV) 24-26.06.2019, Brussels Register here

» egta Marketing & Sales Meeting (TV & radio) 10.10.2019, Brussels Register here

Over 250 case studies are available on egta website. Click here to discover them now!
Privacy policy:
Please note you have the right to manage your mailing preferences at any time by:
> Visiting our Mailings Subscriptions Page or
> Contacting us directly at [email protected].
egta - 22, Rue des Comédiens, boîte 4 - 1000 Brussels - T: + 32 2 290 31 31 -