As a marketer, you know that Google Analytics 4 (GA4) can’t really help you determine what drives your sales. Its default model, last-click attribution, gives all the credit to the last source and ignores the rest of the customer journey, including the channels where your customers found you in the first place. And if you tried their data-driven model, you probably noticed that the results aren’t that different from last-click. That’s because its third-party pixel tracking is often blocked, and most of the time it’s unable to unify multiple devices used by the same person into a single customer journey.
What about your marketing platforms? Can Google Ads, Meta Ads, and your email platform determine how much revenue they drive? Not at all. They’re all unaware of how your customers interact with other platforms (e.g. Meta can’t track when a customer clicks on a Google ad), so they all claim credit for the same sales. That’s why adding all of them up never matches your actual revenue. Plus, they all rely on third-party cookies too, which about 40% of all devices block.
Knowing all this, you decided it’s time to get serious about attribution. After all, advertising is a significant portion of your marketing budget, so you can’t afford to shoot in the dark. But when you started looking for the right attribution solution to help you figure out where your sales come from, you got overwhelmed very quickly. So many options… So much confusion!
In this paper, I’ll help you understand the two main attribution technologies in the market, their pros and cons, and which one may be right for you. I’ll also suggest some great vendors you should consider. While we’re proud of Data Speaks, our attribution platform, it’s not always the right solution, so my intention is to give you an unbiased overview of all the options available so you can decide which is right for you.
The Two Main Approaches to Marketing Attribution
Although there are hundreds of attribution tools available, they all fall into one of two categories: multi-touch attribution (MTA) and media mix modeling (MMM).
MTA and MMM have very different ways to determine what channels and campaigns drive sales.
MTA, also known as touch-based attribution, tries to track every user interaction with a pixel. When a sale or conversion takes place, MTA distributes the credit among all the touch points in that conversion. For example, if a user clicked on an Instagram ad, then a Google ad and then clicked on an email, that sale would be split three ways among Instagram, Google and Klaviyo.
MMM, also known as impact-based attribution, doesn’t track individual users and doesn’t rely on any pixels. Instead, it measures how the change of an input variable (e.g. Instagram Ads budget) impacts the output (e.g. ecommerce sales).
Among the most popular MTA solutions are Triple Whale, Hyros and Rockerbox. Some of the most popular MMM solutions are Recast, Measured and Data Speaks.
Multi-Touch Attribution (MTA)
Marketers like MTA because it’s easy to understand. If a customer interacts with three channels, each gets one third of the credit. It makes sense, right?
One of the biggest problems with MTA is that not all channels have the same impact. Let’s take our example (Instagram > Google > Email > Purchase) and assume we get 100 sales a day that follow that exact same path. We want to increase sales so we decide to increase our Instagram budget by $1,000 one week and our Google budget by the same amount the following week. We may notice that increasing our Instagram budget increases our sales by $7,000, while increasing our Google Ads budget increases our sales by only $2,000. Just because two channels touched a conversion doesn’t mean they both had the same impact.
Another issue with MTA is that we’re trying to track individual users that don’t want to be tracked. Nearly 40% of third-party pixels and 28% of first-party tracking pixels can’t fire due to ad blockers and privacy settings set by devices, browsers and operating systems. Not to mention, most people own multiple devices and device-level data can’t always be unified into user-level data. Although having access to a perfectly-assembled customer journey is a beautiful idea, it’s rarely possible in practice.
You’ve likely heard the phrase “garbage in, garbage out.” These data gaps aren’t just inconvenient, but when attribution models work with incomplete data, they often reach the wrong conclusions, leading marketers into poor decisions.
Given all the data loss and the fact that MTA gives all involved channels participation trophies rather than measuring their actual impact on sales, this is a highly inaccurate method to measure revenue or return on ad spend. If an MTA platform tells you a channel has a ROAS of 5, it doesn’t actually mean that you can invest $1,000 in it and get $5,000 in sales. You may get $10,000 or $1,000
However, MTA isn’t useless. It does something that MMM can’t do. MTA’s ability to track (some) user activity and assemble (some) device-level activity into a single customer journey (when all devices are able to identify the user), makes it possible to give marketing and sales teams unique insights into individual customers and prospects. For example, a salesperson would be able to see that the prospect she’s about to call has read a specific case study and checked out the pricing page. Although MTA can’t track everyone this way and most “journeys” only show the activity from a single device, this can be very useful information.
One of the limitations of working with individual customer journeys is how time-consuming it is to analyze them one at a time. For this reason, MTA is usually a good fit for B2B companies or those that have only a few high-ticket transactions (e.g. boats.)
Another MTA benefit is its cost. Because MTA is a very simple technology, most solutions are only a few hundred dollars a month.
Media Mix Modeling (MMM)
MMM has been around since the ‘50s when digital tracking wasn’t available and companies needed to measure the effectiveness of traditional media, such as TV, radio and print. When digital tracking became available, marketers loved it.
But that love was short-lived. Users started demanding more privacy, so hardware and software companies delivered. Users got smartphones and tablets, making tracking individual users much harder. More marketing channels emerged: Google, Facebook, Instagram, TikTok, influencers and more.
The promise of tracking everything customers do online, on every channel and every device, never materialized. In this landscape, marketers started noticing that their advertising budgets were growing much faster than their sales. And they started demanding a more sophisticated solution that didn’t require tracking individual users.
That’s when MMM became mainstream again. But although marketers in the ‘50s could wait six months for a team of statisticians and econometricians to produce a model, modern marketers face different challenges and have different expectations. They need models that update in real-time, learn from changes in the market, and show them on Friday the performance of Thursday’s promotion.
MMM isn’t the right solution to understand what case studies a given prospect checked out. But it is the best method to measure impact. Assuming you have a quality model, a ROAS of 5 does indeed mean that a $1,000 investment will deliver $5,000 in sales (+/- 5% because no model is 100% accurate.)
If you’re investing hundreds of thousands or millions of dollars a year in advertising, it’s essential to accurately measure how much revenue each channel and campaign generated in the past, and the expected return on investment for all future investments. This is why it’s called impact-based attribution.
However, MMM isn’t always the right solution. Building accurate models is extremely complex, even for seasoned data scientists and machine learning experts. It’s not an optimization problem (which most senior machine learning engineers can solve) but a causal inference problem. This requires experienced econometricians (who specialize in capturing causal relationships in data), machine learning engineers well versed in the digital marketing world, and a very robust model validation process to ensure that a model doesn’t just fit past data well, but it can predict future performance, even in changing conditions (e.g. what if we increase Facebook Ads by 25% and reduce the email frequency to once a week?)
Although a software developer and a data engineer can build a model in a week using open-source libraries like Robyn (by Facebook) or LightweightMMM (by Google), that model will likely be very inaccurate, and the bad decisions made based on its output will result in a lot of money spent on ads that don’t actually perform.
Why is creating an accurate MMM model so complex? Because so is the world. There are many marketing channels, even more campaigns, several customer segments buying many different products in a world that’s always evolving.
Advertising isn’t linear. The ads you run today will result in sales today, tomorrow, next week and next month. Return on investment isn’t linear either. ROAS is very different when you spend $10,000/day vs $1,000/day. Channels interact with each other. When you spend more on TikTok and Instagram, you increase brand awareness, so more people search for your brand on Google, driving up the demand and cost of your branded Google Ads. Your ads don’t perform the same all the time. They do better when you run a promotion and worse than average after it. If you sell swimsuits, your ads will do better in the summer.
A great model needs to take all that into account. The time shifts between when someone sees your ad and when they act based on it. The diminishing returns as channels get saturated. The interaction effects between demand generation (top of funnel) and demand capture (bottom of funnel) channels. Seasonality, promotions and other external factors (e.g. elections, the economy, and the fluctuating advertising costs as your competitors invest more or less on ads.)
As a result, MMM has been out of reach to most brands. After all, investing $200k/year in an econometrician specialized in causal inference, a machine learning engineer and a few servers only makes sense if you’re investing $2M+/year in advertising.
Even working with vendors like Recast and Measured, which would bring down that cost to about $100k/year, would require a minimum annual advertising budget of $1M.
At Data Speaks, we decided to do things a bit differently. Rather than charging a fixed amount, our pricing is based on your advertising-driven profit. When your bottom line grows, so does ours. Our interests are always aligned.
Conclusion
While MTA is great to track specific user activity, it’s a very inaccurate method to determine the source of your revenue and the ROAS of your marketing channels. MMM is the perfect tool for that job but it’s very difficult, even for seasoned data scientists, to capture the underlying causal effects of real life into a model. A poor model leads to poor choices and wasted budget.
If MTA is right for you and you spend less than $20k/month on online ads, consider Triple Whale, as they offer great features and flexible pricing. If MMM is right for you, Recast and Measured are excellent choices. If you’re not ready to spend $100k+/year and like the idea of performance-based pricing, consider Data Speaks.
Get in Touch
Let’s talk! Schedule a 30-minute demo to see Data Speaks in action and get answers to all your questions.
Book your call or email us at hello@dataspeaks.ai
Chad Mellen CEO, Knack Inc