Blog

After my spleen vent on the myth of hyper-personalization, guess who I heard from? AI personalization and predictive marketing vendor ZineOne.

When we kicked off our NRF 2021 retail interview, I asked the ZineOne team: why did they reach out to me, given my unhinged critique? As ZineOne’s external PR lead Austin Edgington told me:

It’s good to challenge it – there’s people out there throwing it around and going willy-nilly. Maybe not all personalization is created equal.

Parsing the AI personalization hype

Collegial debates for the (customer) win. So how does ZineOne fit into the picture? Manish Malhotra, ZineOne Co-Founder and Head of Data Science kicked things off:

My primary focus is to think about our data science efforts. But really, it’s the intersection of AI, ML and product that is of greatest interest to me.

Can we talk about AI possibilities without adding to the hype? Malhotra added:

I think there’s a challenge for the industry, for us as vendors, to be able to put data science, AI and ML in use – so that it can actually make a difference.

Your AI can never be better than your data.

To claim that we can do it for any company’s data set is, of course, a challenge  – and something that we need to really own up to.

Asking the typical enterprise to build AI solutions from scratch isn’t realistic. Vendors do need to productize – and they need to help customers with the underlying data issues also. But what about “predictive engagement”?

That’s one heck of a buzz phrase, featured prominently on the ZineOne web site. What does that mean? Malhotra told me they’ve been pressing towards this via several years of data science pursuits. The idea behind ZineOne’s predictive engagement? Center it on clickstream data – and serve up relevant info for a consumer in that moment. The timeframe is intentionally limited: the last five minutes and last five clicks, but really the next five minutes, and the next five clicks. Is the data limited to clickstream data? Not necessarily – but it does make sense as a focal point. Malhotra:

We’re really thinking about making in-session predictions, based on data that is available up until that point. A predominant part of that data set to us is the clickstream data. We are happy to take the CRM data; we are happy to take data from across channels, and so forth. But most of the time, when we get started, we start off with the clickstream data as the primary source.

How is ZineOne’s personalization different than a recommendation engine?

This sounds very similar to a recommendation engine, where consumers are prompted to buy related products when they are in shopping mode. Is that where ZineOne fits in? Malhotra:

The short answer is no. When we think about in-session predictions, our primary use cases are to understand consumer behavior from the standpoint of certain outcomes. So the outcome could be: are they going to buy or not?

Another outcome could be: are they indicating signs of dissatisfaction? The act of making product or content recommendations is obviously a related problem, but it could be a second dimension of it – the primary dimension being understanding consumers, in terms of certain business outcomes.

Understanding consumer behavior? That’s much more interesting – and ambitious – than only making buying recommendations. But upstart companies still need use cases. The over-used phrase “low hanging fruit” comes into play here. And yes, Malhotra’s team has found such a use case:

We call it the early purchase prediction, the epap model. Early in the session, we want to predict whether somebody is going to buy or not; we’ve seen a lot of success with this.

How early? Within one click? Two clicks? Turns out five clicks is the most revealing: 

The first couple of clicks are probably not as interesting, because there’s a lot of bounce and stuff going on. By the time you get to about the tenth click, it’s a more evolved, maturing session. So five clicks turns out to be an interesting number, because it’s almost the null hypothesis.

So we’ve been able to build out these models, for e-commerce and retail. We’ve worked with large department stores, and we’ve been able to establish this on their site data, regardless of the channel. And perhaps regardless of even the product categories that are involved.

Getting to personalization ROI

Does ZineOne differentiate between a serious buyer and someone who’s perhaps window shopping – and not serious about buying at that time?

We think about it in at least three segments to identify. You want to identify the people who are going to buy; you want to identify people who are definitely not going to buy, and then you want to identify persuadable people who can be influenced to buy, who are otherwise on the fence.

Okay – so what does a customer do with this in real-time? Where is the ROI impact? Example: you might show “persuadable” consumers a discount offer in real-time. A customer deemed “likely to buy” might not see that discount. This personalization use case maximizes purchases and margins. As Malhotra explained, you fine-tune to avoid cannibalizing your budget with unnecessary discounts.

That means you might not show the same offer sitewide. Even with a category, some consumers might see the offer – and others won’t. Malhotra says this intelligent granularity, to use a phrase I just made up out of my buzzword-infested brain, matters a great deal to margin-conscious retailers (which is pretty much all retailers at this point).

So has this use case paid off? Malhotra says the results have been notable, including one of the largest department retailers:

We’ve seen numbers that are multiple tens of millions of dollars in annual benefit. 

But a bump in revenues isn’t so lovely, if you give away your profit with offer discounts. Malhotra says ZineOne’s goal is to boost customer revenues that are margin-preserving, or margin-aware. 

Anytime you decide to make an offer, you have control on whether you choose to be margin-positive, or margin-neutral, or margin-acceptable. We help you to do that.

I’m always wary of the downside: who have you alienated? Just because you saw an uptick, you need to measure the “annoyance factor” of personalization also. Malhotra told me so far, so good:

We’ve not had anybody call customer support and say, “Hey, you guys showed me an offer, and I have a problem, or you guys bait-and-switched me, or whatever.” So I think we’ve done it in a manner where transparency to the consumer has been good. It’s been hard to game these offers.

My take

We talked about another potential issue: some consumers might feel excluded. Example: I talk to my friend; she got a 20 percent offer discount at checkout, and I didn’t. A well-educated customer support team could solve that, by extending the same offer to anyone who complains, but needs thinking through.

How accurate are these “likely to buy” designations? That’s a complicated question. Malhotra says that the possibility of 90 percent accuracy is there for simple marketing messages. But in other cases, he acknowledged there is a wider margin for error.

In product segments, he says if the traffic for that category is substantial enough, a bump of 20 percent is reasonable. What has surprised Malhotra is how this carries across categories – from clothing to furniture (if you want to try to see ZineOne in action, Men’s Wearhouse is one site that uses their technology). And no, you don’t need to be a logged-in user for ZineOne’s tech to work. Their personalization can work with anonymous, first-time visitors.

I really don’t have a problem with any of this. So hold up, you may be asking: how do I reconcile this with my caustic views on personalization? Well, ZineOne is doing a number of things that tone down my concerns:

• They aren’t trying to boil the freaking ocean by claiming they can help customers understand their consumers in an outlandish “365” or “complete customer data platform” kind of way. They are focused on clickstream data, limiting their ambitions to understanding the scope of a customer’s behavior in a given moment, on one specific channel. We already know personalized recommendations can drive sales. This is not dissimilar, in that ZineOne is not over-reaching beyond what personalization tech is now capable of. Nor are they promising to understand what a customer needs over time – an entirely different problem (think vendors that push the crap out of luggage offers, after you bought the last luggage you’ll need for a long time).
• You can get going quickly without building a new data platform to get a result.
• They aren’t trying to push a financial transaction at all times. They are looking at a range of use cases, including customer dissatisfaction. Attention to configuration details such as margin control indicates there are plenty of complexities to solve. But you limit them but addressing a narrower problem.
• I don’t care what the CX gurus claim; B2 B (complex) purchases are inherently different than B2C. The cost of alienating B2B influencers is much higher, for example. Nudging customers towards purchase is a different game in B2C, where a shopping cart can be abandoned and never returned to. Most B2B purchases require numerous touch points over time.

We are in less controversial terrain with AI for commerce than say, AI for hiring, where enormous ethical questions come into play, as systems in production can literally wreck people’s lives. That said, we still need a vigorous debate on what’s possible with personalized commerce – and what’s fanciful.

You can’t do this type of real-time personalization at scale without serious performance chops. You also need to offer customers ways to engage beyond individual transactions and moments. Then there is the problem of data readiness and compliance – never a negligible issue. Those topics were beyond our interview time limit, but deserve more attention. I’ll look to do a ZineOne customer use case down the line – and dig in further.