This is is my second post I wrote based on a series of questions a data scientist asked me some weeks back on Marketing Mix Modelling.
I was really surprised by the interest generated and the new connections I had on LinkedIn.
So this post must start with a great and genuine: Thank you.
Be careful each business challenge is different and also the right answer.
If I use Google and Facebook clicks as media_data, each with different revenue models (CPC for Google and CPM for Facebook), would that pose challenges in modelling?
It depends.
In my experience the goal is always to understand and relate the impact of the media impressions on sales.
Let’s take this example.
Our customers see an amazing ad on YouTube on their smart tv and then the same product pops up on IG ads and they click on it.
If you use the clicks in the model you can overweight IG efficacy, if you use impressions you can reduce this bias.
If you put clicks you risk an overweight of Google Search contribution, because usually it is the channel with the higher CTR.
What is interesting is to create a correlation map between Weekly Impressions / Weekly Sales for each channel and a correlation map between Weekly Clicks / Weekly Sales, you might get different findings.
In general it is good to have comparable media variables, while CPC and CPM are very different and they can skew your results.
I’m considering using clicks for both Google and Facebook instead of impressions because, in my view, impressions on sponsored Google websites might not offer significant value. What are your thoughts on this?
I would not encourage this choice.
If Google websites don’t offer significant value you already have the answer, they should be removed from the marketing plan.
There is nothing bad to say “I am not able to find any relevant impact from this channel”
This is something also discussed in a YouTube Video from the Company Solutions 8
In short, when you promote your campaign on Google Adwords Network Websites or Facebook Audience Network (now they changed the naming) you are displaying the ads in low value websites.
Apart from that it is also important to understand the budget and impression allocated.
Because if the weekly budget and impressions are small compared with sales values it is reasonable that the model will not detect any significant effect.
Why? Because it will be confused with random fluctuations.
This is a well-known problem in A/B testing and that’s why it is critical to identify the sample size and the sampled standard deviation.
While planning to use media costs as media_data for the MMM model, I also want to include organic data. What approach would you recommend?
Organic data could be modelled like weighted distribution for the Fast-Moving Consumer Goods (FMCG). Then you will check how this assumption is reasonable, based on your specific model.
On this topic I really agree with Byron Sharp on looking at Organic and Paid Search as shelf-space.
You can find more here: Why paid search is like shelf-space
I would not suggest using media cost as an input data in your model, I noticed that this is quite common especially among Robyn users. My point of view on that is clear, consumers are influenced by looking at the advertisement, represented by impressions, not by the budget, and for that reason I prefer to use impressions as the independent variables.
Additionally, I’m considering incorporating AOV, CTR, conversion rate, and ROAS. What are your thoughts on this?
I’m not sure why you should put Average Order Value (AOV) in the model, maybe I am missing some piece of the puzzle.
AOV is another way to look at sales, so if it is changing I expect that something changed or in the media strategy or a new SKU was introduced or a new subscription service was activated.
This doesn’t mean you should remove it from your analysis, if you see a trend in the Average Order Value you should discuss with the team.
Trend can also become your dependent variable in a marketing mix model.
I am not a huge supporter of CTR, but try to make the model as much comparable as possible, if you use CTR don’t use Impressions and vice versa.
Return On Advertising Spend (ROAS), absolutely not.
ROAS should be what you get after defining the sales contribution from each media channel and dividing by the cost of it.
Tik Tok, Google ADS, Meta ADS, all give a KPI called ROAS, but how they relate the information from the other platforms is unknown for us (and probability also for them).
Their ROAS doesn’t take into account other platforms so even if you get an Improvement on your R2 I would not put it in the model.
Lastly, do you have plans to write about creating a marketing mix model with Robyn (R) in the near future?
Yes, it’s quite probable, but not earlier than May-June
If you liked the article please feel free to share.
If you want to get in touch with me please fill the form below or add me on LinkedIn, I am happy to connect.
