How Seeda.io Uses AI to Fix Messy Marketing Data and Redefine ROI

(00:26) I don't want to say longest distance because we have had guests from all around the world here. We also have our AMG APAC series. Michael wound up joining us on this stage at an uncomfortably early or late hour for him, as he chooses. Seeda has been doing fascinating work in marketing mix modeling. I've been talking with Michael and the team for a good chunk of the year, if not longer, and learning quite a bit from them. I'm always excited to hear more about

(01:02) how we really apply AI to this area, which we haven't discussed much lately. Michael, welcome and good morning or good night. Hey David, good morning, and thank you for having me. It will be great to hear more about what you're up to, what you're excited about these days, and dive in.

(01:29) You're welcome to present or share as much as you like, or just keep this conversational. Glad to have you here. Hey David. To introduce myself, I'm Michael Kingston, founder and CEO of [Seeda.io](http://Seeda.io). We specialize in marketing mix modeling, and in particular bringing marketing mix modeling to businesses of all sizes.

(01:55) The biggest surprise people have when I meet them is that MMM is now cost effective and more readily available to businesses of all sizes. It's no longer reserved for the biggest companies in the world. We're bringing it to mid-market businesses and enterprise businesses globally.

(02:20) When you say "of all sizes," I’m used to seeing IAB reports and talking to research firms where you have to spend a ton before you can even talk to them about any kind of media or marketing mix modeling.

(02:49) You seem to be saying that's not the case anymore. No. Seeda is five years old. I set up Seeda five years ago, and in that short period I've seen the cost of the technology that fuels MMM come way down. Our mission at Seeda is to bring the technology to as many relevant folks as possible.

(03:24) The cost has come down to the point where we're bringing it to businesses that spend as little as $50,000 a month on their marketing budget. We have fees for different types of businesses. Our entry-level fee at the moment is $5,000 per month. When I started five years ago, we couldn't have delivered a meaningful MMM solution for $5,000 per month.

(03:58) But the underlying technologies, such as machine learning, have become very cost-effective and very powerful. We're trying to bring it to those businesses. In terms of revenue, it used to be reserved for the billion-dollar-plus range.

(04:18) Now, our smallest business does about $10 million in annual revenue, is growing rapidly, and expanding its marketing mix. We're able to come in at the right time in that business's growth journey and provide advanced marketing measurement, which ordinarily wouldn't have been available to that business just a few years ago.

(04:49) There are two acronyms of MM that I’ve heard: media mix modeling and marketing mix modeling. When you're saying $50K marketing budgets, is that all working media? Are you looking at things that are not paid media as well? Do owned email channels or other things come into play? What are you looking at?

(05:26) We define it broadly. Predominantly I hear folks calling MMM "marketing mix modeling," while the traditional definition is "media mix modeling." At Seeda we define it as marketing mix modeling and we look at the full marketing mix.

(05:53) It's not just media. It could be influencers, physical events, email or CRM. We look at the full marketing mix. It really becomes very important in larger businesses, but we define it as a broader marketing mix, which means marketing mix modeling.

(06:24) What's involved in modeling? How does the process work? Keeping it high level, I can get into more detail. There are various techniques to marketing mix modeling. We believe the most accurate possible model gives the best results.

(06:53) For those who know nothing about MMM, it's an AI algorithm that looks at all of your marketing spend. The underlying technology we use is a Bayesian machine learning algorithm.

(07:15) Our process: we've spent five years building a 30-day onboarding process. There are two really important parts to MMM. First, there's the data that you're going to model. It's the old saying: garbage in, garbage out. If you feed garbage data into any algorithm, no matter how good that algorithm is, you're going to get garbage results.

(07:47) During the 30 days that we onboard new customers, we spend at least two weeks making sure that the data we feed in is not garbage. We've developed a lot of automations for that. Once we have accessed the data, reviewed it, cleaned it, and transformed it into more accurate data, that's our first step.

(08:10) Second, for every single customer, we build a custom model. We start from scratch with every single customer. We don't sell a self-service tool where you're getting an out-of-the-box, preconfigured model.

(08:34) We customize each model for each individual customer dataset. Every brand has unique data. To get the best results, each brand needs a fully customized MMM model. Once we add clean data to a custom model trained by an expert—we use a lot of automation in the background—we then, after the 30-day onboarding,

(09:07) have a model that's ready for testing. We do rigorous testing before we use the model to predict any marketing mix results. Once it's tested, you have a custom-built machine learning algorithm for your business that you can use to ask questions and, as a marketer, get more accurate, fuller-picture answers. It's been working out very well for our customers so far.

(09:47) Can I ask a question? The garbage-in-garbage-out makes perfect sense; I've lived that. What about if the data is pretty good but it's all in silos and doesn't talk to each other? One big challenge is unifying the data.

(10:11) Are you a middleware layer that sits on top and can find it wherever it resides, or does it need to be centralized first? We tried to be the company that only worked with good data, and I quickly found out that very few marketers have good data. We call it MMD—messy marketing data—because most marketers are embarrassed about their data. They haven’t had the time or budget to sort it out.

(10:49) We provide the full stack. We find your data. It could be in all sorts of places. We unify it—pull it into what we call a big data warehouse.

(11:11) We ingest it, review it, find problems, stitch it together, and use data transformation to combine data from various sources.

(11:30) We've built a rigorous ingestion and data cleaning process to help you sort out messy marketing data. We test it before we model it. You can never get a dataset 100% accurate, but we get our datasets very accurate. We make sure you and your team approve that everything looks good before we build the custom MMM for you.

(12:07) Great. Thank you. I’ve spent my life cleaning messy marketing data. Most of us have. Thanks. Are we heading toward MMM for all, where there's a good-enough out-of-the-box version that anyone can tap into, and then when you want to make this work you need someone like you to create a custom model, albeit more efficiently than in the past?

(12:55) We've tried to build a self-service tool and have spent years trying. Over the next three to five years, I feel the market is moving in that self-service direction.

(13:22) The more complex your dataset, the quicker you'll figure out whether a self-service tool adds value if it hasn't been configured as custom as you need. It's not one-size-fits-all.

(13:45) Figure out what sort of brand you are and what solution you need. There are open-source tools available for free. You just need the skills to set them up. They're a good start. For brands above $10 million in annual revenue and beyond,

(14:05) we advocate doing a more rigorous job because, as I said, the more work you put in, the higher the accuracy and the better the predictive power the algorithm will give you.

(14:26) MMM is based on Bayesian machine learning, which is predictive. Once it's built and tested, you can start asking it questions and it gives you predictive answers.

(14:51) Over the past decade, the last 10 years of marketing measurement have been chaotic, predominantly because of messy marketing data and privacy changes. Many marketers are frustrated and skeptical about numbers from current providers because the landscape has changed so much. It's refreshing to bring a powerful predictive algorithm to marketers so they can trust the numbers.

(15:58) What's your point of view on letting the media agency do the analytics and the marketing mix modeling versus the client doing it themselves? I have a bias, but I'd like to hear your opinion.

(16:24) Many in the industry call it marking your own homework. Our view is to collaborate with whoever we're working with. We don't have a strong view either way. Most of the time the brand brings us in and we work collaboratively and transparently with the agency or multiple agencies, the brand, and our team.

(16:57) Other times agencies bring us in, and they work transparently with us as well. After dozens of projects, I'm not biased one way or the other. I haven't seen a badly performing agency bring us in.

(17:35) The top 10% of agencies are truly in it for their clients. Whether they bring us in and results show the media channels they manage are not performing as well as others the brand runs in-house, we’ve been lucky to work with high-integrity agencies. MMM has been around for decades, but we’re at the early stage of mass adoption of Bayesian machine learning MMM. As we work with a broader set of agencies, my view might evolve.

(19:02) What's your bias? I call it letting the elephant watch the peanuts. The objectivity isn't always as good as you'd like. Agreed. Is it possible to see a demo?

(19:30) Before the demo, I give everyone a conceptual diagram to frame what you’re about to see. Can everyone see my screen? Yes. David, can you see a series of blue moving dots? Great.

(19:55) This shows the data flow we believe you need to set up to deliver highly accurate MMM. On the left side, you've got all the different types of data—it's not just digital media. It's all types of media, all sales data, all CRM or email marketing data, SMS, the digital media like TikTok and Facebook, and offline channels, which are a big part of the mix for larger brands. First, we connect all that data.

(20:46) We spend a big chunk of our time bringing your data together, putting it into a single cloud data warehouse. We've written data processing automations to clean it. Only then, once the data is processed and cleaned, do we feed it into our marketing mix model, where we build a custom model for each customer.

(21:13) Once that's done and tested rigorously—we have a very rigorous testing regime—before we start talking about how we'll use the MMM, we run multi-week tests on the algorithm to make sure it's accurate and useful.

(21:38) Once testing is done, we break our insights and value into six categories: budget optimization, our industry factor concept (IF factor), predictive analytics, saturation curves for every marketing channel,

(22:05) seasonality, which is often overlooked but very important, and sales forecasting, which CFOs love. Conceptually, that's what's under the hood of what I'm about to demo.

(22:34) This is our MMM hub. We've built a full suite of marketing measurement: a metrics hub, an MMM hub, our IF factor feature, and a business optimization hub. First, the MMM hub. We want the full marketing team on the same page. We present in real time. For example, this brand does about eight channels: some above-the-line traditional marketing and a lot of digital.

(23:24) You'll see us bringing all that data. You'd be amazed how many marketers work with incomplete and disjointed datasets.

(23:43) First, we bring all that data out of siloed systems and present it in a unified and accurate way. Typically we go back two years to give a holistic view.

(24:06) We bring in each channel and how it has contributed from an impressions perspective over time. This gives the first accurate step of what's happened historically. Then the first time you see MMM outputs, we call this our layer graph: a two-year view.

(24:27) All your data has passed through our algorithm, and the MMM assigns a contribution for each channel over about two and a half years. It's liberating to show this to a marketing team for the first time. They’ve seen biased and disjointed reports and never had a full picture. This is the first time they've seen the output of an algorithm. You can go down to a weekly view and get very granular. The MMM we build gives a week-on-week view that you can roll up any way you like.

(25:36) From a performance perspective, our product lets you finally see the ROI of each channel simply. We can go down to the channel level and the campaign level if needed. For a monthly period, here’s the ROI. Many realize their base contributes a lot to marketing. "Base" is what happens if you turned off all marketing today. Most brands have non-marketing impacts on sales—brand effect or market effect. MMM measures that brand or other market factor inputs.

(26:47) Next is our response curve. Once the data is processed and pushed through the MMM model, it breaks down how each channel is performing, and we try to narrow down the sweet spot of your budget for each channel,

(27:12) interpreting the point of diminishing returns. Every week of the year can be different depending on seasonality. For example, a YouTube view might be approaching saturation but not there; search is often saturated. Each week of the year, we show where a channel is on this curve. We try to avoid oversaturation and find the sweet spot. Seasonality is a massive factor. Channels change all the time. We measure weekly. We're an always-on measurement solution.

(27:52) These saturation curves change week on week. You can check them regularly. MMM also measures ad stock and lag analysis. We break down each channel's ad stock and lag to understand sweet spots for those metrics.

(28:13) Seasonality is a big factor. The algorithm measures your seasonality across every channel and gives you an overview of the general trend for your business.

(28:32) The cherry on top of MMM is that it’s predictive. Question: what's the assumption about attribution? First touch, last touch, multi-touch? What are we tracking against here?

(28:57) Great question. MMM is not an attribution model. MMM brings all your data out of those channels wherever it sits—Google, Facebook—stitches it together, and uses a statistical model. It uses an aggregated view.

(29:34) Attribution, by definition, is user-level attribution: trying to track people around the internet and assign value to individual users. MMM doesn't track people. That's become very inaccurate over the past decade.

(30:24) Instead of tracking people and assigning value to an individual user, we aggregate the data to a higher level and use statistics. For example, we spent $10,000 on a Facebook campaign. The algorithm looks at your data holistically and finds signals in that data to assign a contribution to that spend.

(31:22) It could be that the algorithm finds a probability that the $10,000 delivered zero because it can't find a signal. Or it finds a signal that correlates to $100,000 in sales. It assigns a contribution accordingly. That's what you're seeing in the outputs—an aggregated view. Attribution is often binary at the user level. MMM gives a probabilistic, statistical view of contribution across your marketing mix. Does that help answer the question? Yes. It's a predictive model.

(32:39) Yes. Someone else jumped in—Okaro here. Welcome. Any other questions or comments on what we've seen from Michael so far? All good. Back to the demo.

(33:13) Why are we doing this sophisticated data analysis? To get better clarity on our data holistically. As a marketer, I want to know what budget I should spend on each channel to maximize my goal, usually sales or new customers.

(33:44) We've built a budget optimizer. You can use the algorithm's predictive power and configure it depending on your business. Some want to consolidate their budget and squeeze every dollar from the existing or reduced budget. Others are growing fast and want the optimum mix and increase in budget before hitting diminishing returns.

(34:26) We provide a channel view. For a brand with $5.2 million in revenue off a $1.3 million spend, this is the existing mix based on the past two years of data. We then give scenarios to help decisions.

(34:52) In one example, keeping the budget the same but optimizing mix to about $1.4 million in spend yields a 13% increase in revenue. Increasing budget by 15% yields a 19% increase with a different mix. Another scenario models a 30% increase. Once you've done the hard work to set this up, you can move beyond intuition and use a data science–backed approach to measurement and optimization.

(36:02) It's becoming more cost effective as technology improves. That's our MMM hub.

(36:27) Unless there are questions, I'll show our IF factor section. Do you have a point of view on customer acquisition costs versus ROAS?

(36:58) The neat thing about MMM is we can measure everything. Brands have their own ways of making decisions. Many are CPL- or CAC-focused and use that as a primary metric. We can set the model to measure whatever metric matters to you. ROAS comes into it, and I'm about to show our IF factor. Do you mean which is better? There’s a trade-off.

(38:05) We consider ourselves a data science company. We can measure everything, but when we come in we assign a customer success expert to understand what you're measuring and why. What we’re seeing in the market is what I call the great sugar addiction of the past few decades: measuring ROAS, especially in-platform ROAS. Marketers have become dependent on in-platform metrics.

(39:13) I'm a big believer in not using in-platform metrics like ROAS as a north star at the macro level. You should look at holistic measurement solutions like what I showed earlier. To help marketers, we've developed the IF factor. Instead of looking at the in-platform ROAS number, we pull that attribution data out of the platform,

(40:18) cross-reference it with our holistic model, and give you calibrated numbers we call an IF factor. For example, a brand looked in GA4 and saw Google Search at close to a 4 ROAS in-platform. Cross-referenced with MMM, the incremental factor is closer to 2.

(41:32) You should use in-platform metrics for smaller, day-to-day decisions. For weekly or monthly strategic decisions, don’t use ROAS as your north star. Use a calibrated, incremental view like IF factor. Does that answer your question, Amy? Absolutely. Thank you.

(42:21) We provide IF factor for each digital channel we can. For linear TV, without a data feed we can’t provide it, but for other digital channels we provide a more calibrated number. If the platform is over-reporting by 90%, you'd make very different decisions. We aim to give access to holistic measurement so you can make better choices.

(42:46) That’s 90% of the demo. Lisa, does that give a better view of what to expect? Yes, thank you. Much more real and vivid.

(43:39) Another question: we work on channel strategy and suggest channels to experiment in. Is your platform capable of surfacing channels to explore based on data across your clients? No. For us, MMM needs data before we can give results. You have to be advertising in that channel before we can measure it. Once the algorithm picks up a strong enough signal from a new channel, we can advise.

(44:34) We’re not an advertising agency; we’re a data science company. We can give market-research-style perspectives but won’t advise you to start a new channel. For one QSR brand, we saw radio performing well for peers and suggested it may work, but until they started and we measured it, we couldn’t say. They bought a three-month radio plan with strong creative. After about two months we had a significant read and advised another six months. We need data to measure before giving a view.

(46:14) Follow-up: what's the lower threshold where it breaks even for a brand to pay for your SaaS model? It's a monthly fee. Typically brands under $50,000 a month in marketing spend—between $500K and $1M a year—we don't think we’ll generate enough return to cover our fee. You're better off spending on media to grow the brand.

(47:03) We try to deliver at least a 10x return on our fee. For now, the line is about $50,000 per month and up. We're a sophisticated solution with associated fees. There are open-source algorithms you can start with for free, though you may need to hire someone to set them up.

(47:44) It's not as good as what I demoed, but if you're in that smaller spend range, it gives a better read beyond in-platform metrics. It's a maturity cycle: start open source, then move to more sophisticated solutions as you mature. Traditionally these have cost hundreds of thousands of dollars. Our entry fee is $5,000 per month on a SaaS model. At the right time, it's cost effective for the right brand.

(48:35) Quick question from Kizia: how low can spend be on a new test channel before you get reliable results? Do you need $100K or $1M, or how low?

(49:05) We've seen results in the thousands of dollars. It depends on channel, time of year, and other factors. A $10,000 media buy over multiple months of measurement has been enough for the algorithm to pick up a signal and assign contribution.

(49:35) So, $10,000 and above in spend over time should give some read to inform whether to scale or plateau. Thank you. This is great. Anything else before we wrap? Michael, what's the best way to stay in touch?

(50:24) LinkedIn is great. Or go to my website—I'm on the other end of the chat on the site. Message me there. I'm in your Slack groups too, so mention or DM me there.

(50:43) This is great. We'll share the recording for those who couldn't make it. This is just the start of conversations to come. It's important to see where AI is changing things. With all the messiness of AI, it's good to see areas where data becomes more accessible and decisions more impactful.

(51:21) Thanks for highlighting some of that. You're very welcome, David. Thanks again for having me and for the great questions.

(51:38) Thanks everyone for joining. Appreciate the great questions from Natalie, Keia, Lisa, and everyone who contributed. Great to see you all and see you next week.