Back in 2008 I thought fractional attribution was a complete solution but after the last four years working with brands, my opinion has changed.

I have discovered that while this method is not entirely without merit, it is disappointingly limited and doesn’t do much to help a marketer re-allocate spend meaningfully.

In 2008 Microsoft said 'the company can now provide a scientifically based standard showing how well different media affect an eventual conversion', this was engagement mapping, a classic example of fractional attribution and what have we learnt from it? 

My take on it is that user interface sliders and shiny bubble graphs are sexy but technology has to do more than look good, it must synthesise, it can’t just echo back what you tell it.

It’s important to try and learn about the successes and failures of your marketing spend, this is how we build successful programmes. Attribution analysis should enable you to examine the empirical value of your media and reallocate your spend accordingly.

Here we are, four years on from engagement mapping and most technology suppliers have still not developed any meaningful alternative to attributing returns to media investment on a last-in basis. 

In my opinion, at the centre of this problem we are plagued by two facts:

Correlation is not the same as causation

A can be correlated to B but that doesn’t necessarily mean A caused B. 

Take for example the tight relationship between shark attacks and ice-cream sales, however we all realise that we can’t save lives by banning ice-cream sales because the one is not causing the other even though the two appear to be co-related.

We all want to know which elements of our media investments are causing sales so we can re-allocate spend into those elements and cause more sales. 

Simples! Except it’s not that easy!

Return On Advertising Spend (ROAS) can only really be considered at the level of the entire budget. 

The moment we consider a subset of the total media spend such as investments in digital sub-channels like search, social and display, we need to address the attribution of sales revenue correlated to these sub-channel media investments. 

Suddenly we are now talking about degrees of assumed causality, how much of your reattributed spend are you going to allocate to each sub-channel? 

This is where our industry has willingly ignored the true challenge/opportunity of attribution. What has happened is that to simplify matters, the industry has favoured a dumbed-down approach in which the degree of assumed causality is fractioned out according to the marketer’s arbitrary reckoning.

Typically we see equal share across these digital sub-channels, a model sometimes referred to as “flat attribution”. What’s even worse is that some agencies/vendors would have the capacity to up-weight elements of the media plan that make them more margin.

So what is the true opportunity here? We are talking ourselves out of the last click standard, so what could replace it?

We may be comfortable with the idea that a physical action, like a click, is a more qualified engagement compared to a display impression but strictly speaking we can no more claim causation from a correlated click as we can from a correlated impression.

The good news is that identifying media effectiveness is not like proving innocence or guilt in a court of law. 

It’s OK to let some clicks go that were actually responsible for sales and lock up some clicks in media plans that were caught up in circumstantial evidence and which had nothing to do with developing purchase intent - nothing unjust is going to happen, we’re looking for marginal improvements in performance not proof cases.

What I recommend is working with proven statistical models. The statistical approach identifies highly correlated media events, not just all correlated events and most importantly it synthesises a degree of correlation.

By isolating single variables such as creative, placement or retargeting context while controlling all other variables (in so far as possible) it can yield statistically significant indicators of incremental media value. 

This provides the confidence in the assumption that there was causation effect (remember, you can’t get away from the fact that this will always be an assumption). 

Here are three key takeaways for considering an approach to your attribution strategy:

  • Fractional re-attribution can provide insights but don’t be afraid to question a vendor that offers this as their only solution. How much use is a report that echoes back what you told it?
  • Statistical modelling takes into account media that didn’t correlate to sales meaning you also take into account media investments that didn’t result in sales.  Best not to ignore the media that didn’t convert as it’s likely the bulk of your media spend.
  • Statistical modelling can provide an actual number that quantifies the contributing effect of elements in your media plan so you can confidently reallocate spend. Proper statistical analysis ensures these actual numbers are only drawn from statistically significant sample sizes.

This is mathematical marketing: let’s use methods of which our GCSE maths teachers would approve.

Robin Davies

Published 27 February, 2012 by Robin Davies

Robin Davies is Country Manager UK at Mediaplex and a contributor to Econsultancy.

1 more post from this author

You might be interested in

Comments (7)

Save or Cancel
dan barker

dan barker, E-Business Consultant at Dan Barker

hi, Robin,

thanks for writing this great post.

I agree with lots of this, including that a central problem is that that 'attribution' has been dumbed down - both in order to sell the concept, and to sell software.

As with anything like this, the data that 'attribution' tools give us are a means to the end, rather than the end themselves. And, as in other cases, hypotheses, testing, observation are the things that give us the actual insights & results.


over 6 years ago

Matthew Phelan

Matthew Phelan, Director and Co-Founder at 4Ps Marketing

Hi Robin,

I would like to echo many of your comments. For me the key with all these techniques is to know their limitations and to understand the data as a human being before making decisions.

My biggest frustration at the moment is technology vendors that pitch their tech as a 100% solution. Yes these tools are great but saying they are a 360 solution for all our problems isn’t helpful to anybody.

Having said that I am excited about the future and I look forward to a more integrated future.


over 6 years ago

Robin Davies

Robin Davies, Country Manager UK, DE & APAC at ValueClick

Hi Dan and Matt, thank you for your contributions. I think you are both spot on for highlighting the role of experienced people in turing data into actionable insights. Whether that's a test-and-learn cycle looking at the contribution/attribution ratios or SQL queries that answer specific marketing hypotheses the output is actionable. I have seen this non-system approach reward brands with some very valuable insights: Check out this short video case study:

Thank you so much for your comments I look forward to meeting you both one day soon.


over 6 years ago


Lucia Mastromauro

Agree! I have seen vendors proposing which is mostly a finger in the air approach to valuing each event.

Just analyzing correlation is certainly a start, but one would want to be down the path of understanding causation, true incrementality levels and model converting and non converting paths in order to develop a sound a cross-channel attribution model.

over 6 years ago

Robin Davies

Robin Davies, Country Manager UK, DE & APAC at ValueClick

Hi Lucia

Thank you so much for your comment - I see you understand this space well and I'm glad you highlight both non-converting paths and the importance of empirical outputs, "incrementality levels". Big brands like yours can't make important media investment decisions based on finger in the air approaches.

over 6 years ago


John Shomaker

I feel your frustration, but this very topic is both the source of innovation around marketing and the basis for its provider fragmentation: it's impossible to conceive of a model that accurately predicts and reallocates spend across a virtually unlimited set of online and offline marketing stimuli. Point: get comfortable with imperfection.

Tactically, not sure I buy the correlation-causation issue, which has arisen lately in blogs. Auto-correlation statistics can help solve the 'shark-ice cream' issue, and good analytics professionals know that correlation is only a first step to validating those variables' ability to predict in-market.

I'm also not a hater on CTR. In a funnel construct, any campaign should have a clear objective function and CTR is "adequate" for promotional advertising targeted against high-intent audiences. Offline action or other behavior is good here too, if the data is available. For branding, the right measurements are still old-school changes in attitudes: aided and unaided awareness, recall, consideration, and intent or even buying velocity. Mixing stimuli with non-obtrusive attitudinal work would be nirvana for brands.

over 6 years ago

Robin Davies

Robin Davies, Country Manager UK, DE & APAC at ValueClick

Hi John, thank you for sharing your thoughts here. You covered a lot of ground in your comments (causal inference, in-market modelling, blending online with offline interactions, DR vs Brand objectives and more) I hope other people will wade in and comment on some of this.

over 6 years ago

Save or Cancel

Enjoying this article?

Get more just like this, delivered to your inbox.

Keep up to date with the latest analysis, inspiration and learning from the Econsultancy blog with our free Digital Pulse newsletter. You will receive a hand-picked digest of the latest and greatest articles, as well as snippets of new market data, best practice guides and trends research.