Hi Gary –
First off, and I am being dead serious, is Kelly outsourcing her blogging efforts? Because she is killing me so far in this endeavor, there has to be a team of writers at her disposal or something. All I have is the infinite monkey theorem at work (hence the delay in my retort).
To answer your first question, YES! Cardinality is key. If you only have 3 possible offers then it’s not hard to use your run of the mill testing platform and do a hack job to pipe in nightly model scores as one of the selection criteria in making a decision of what to offer. Example – say I’ve created some mythical customer value score based on previous behavior in my CRM system. I could use that in conjunction with current behavioral triggers to determine what content to throw at you on this visit or even on the landing page when you enter my lovely “Hazen’s Ambassador of Good Times” Horse Racing Website (that was a joke just for you).
I probably should re-state what I said in my last post in a different way (and when I say different way, that is code for retraction of what I said). Real-time decisioning probably isn’t over-rated but real-time scoring is. I think that is a subtle difference that might be lost on people. Because often I believe people think we need to re-run a mathematical model every single time we want to make a decision and I don’t necessarily agree with that. I believe that is overkill. I am sure some folks think I am crazy for that but that is what the comment section is for.
I still think a majority of organizations can get away with doing batch scoring to determine that “Gary” is ripe for offer A next time we come across “Gary” and creating the queue to work off of because you have a limited set of things you’re going to offer “Gary”. I’m an advocate of doing lookups until you’ve maxed that out. If you are crushing your personalization goals and looking to go to 11, then fair enough, buy the amp that can do real-time scoring and rock out to that kind of wizardry. But as with everything, whether it’s running a marathon or doing personalization start small and build up.
And as I said last time, most organizations aren’t even doing real scoring - batch or otherwise - to utilize in their decisioning or personalization. Most are still at the “He’s bought shoes, he probably needs socks!!!” And then wonder why you aren’t buying a bunch of socks. Seriously…”Gary”, buy these socks they are fabulous!!! As an aside, I joke about the ‘If-then’ rules stuff a bit, but they definitely have their place if done right and are fueled by insights found in the data. And they should be the first way of doing personalization, use your data to create your If-then or segmentation and test the bejeebus (technical term) out of it.
Now back to that queue that I mentioned above, because that’s where the fun begins. If you have only a few offers, not a big deal. But if you are a media site or you’ve taken your business to 11 or an online retailer (like one named after a river in South America) that has tons to offer, it gets back to what you were saying with regards to cardinality. Because even if we do batch scoring and you qualify for offers or content a couple of things start to come into play:
- Which one is best? Or maybe it is multiple?
- What current behavior (i.e. triggers) on the site would cause you to re-think your offer/content choice for the customer/visitor?
- How long do these things stay in the queue before we go to the next in line?
- Do we offer the same stuff in multiple channels (if this company operates that way)? Does that set off the rules to recalibrate again if customer responds?
- How often do I run and query batch models?
- Where would this stuff sit technically (in the cloud, in my walls, at my house, under my desk, etc.)?
These are just a smattering of things that crop up that your testing platform will have problems with. To handle these questions, I think you are looking at building a full-blown custom solution to fit your specific business needs or taking a combination of components from different vendors (some of which I might have worked at in the past).
OK…You had asked about personalization opportunities and how you go about deciding where to identify it on a website. I am going to go all lawyerly (is that a word?) on you and say it depends. And when I say it depends I mean the severity of personalization. The fact that I know you are “Gary” and you live in San Fran and you like the Giants (Lincecum is back baby!) and love analytics and may have some cursory knowledge of horse racing (wink), using all that would be creepy in a personalized message on my website, right? However, using the fact that I know you are visiting my site from San Fran and have been here before and consumed various content and originally came from organic search tells me something about how I should message you in the future, which is probably acceptable. There is a balance. Some sites differ though…banks when you authenticate, you expect them to know you. Other sites, customers get creeped out.
So if I am doing this from scratch, I play it out like this:
1) I take my CRM data and try to mash it up with my website data. Sounds easy but I am glossing over that it might not be easy for some verticals and especially in some countries based on local laws. I get it. But there are creative ways to find a primary key – DMA, zip, devices, time series, campaigns, etc. If I have nothing with my CRM or backend, I just use my web analytics data for what it is. The reason for using CRM or the backend should be obvious but I want to make sure everyone understands I want to know the complete picture of the customer not just what is happening now on the web.
2) I look at the data to make inferences for tests…the first being of the ‘If-then’ variety of personalization. I conduct these tests based on behavioral triggers instead of trying to feed in what I know about you from the backend. I am not trying to put people on Mars, but rather trying to get down the street in one piece. So say in my data analysis I find a high correlation to people in San Fran and using organic search becoming valuable customers…maybe I treat them differently? Maybe I test that theory out first?
3) As with anything it becomes a cycle…I measure whether my personalized offers make a difference over doing nothing to the control group of customers/visitors. It’s an iterative process.
4) At some point you’ll want to start bringing in more and more data points and see if it makes a difference…and test it. So say you can connect a customer based on authentication or using a cookie to tie them to a known customer to determine they are a higher roller and make them a different offer, test it. Maybe you tie in your 3rd data sources from the DMPs of the world and use their behavioral/attitudinal segments as well to enrich your selection criteria and do more micro-targeting/personalization.
5) The point I am making - and I made it above previously somewhere - is start small and then add stuff in. I guess one thing to consider though is if you don’t plan for the future, meaning adding other data points like CRM, DMPs, analytical scoring models, etc than when it is time to use them you might be screwed. So maybe…dream big, start small? Is that a Cat poster?
Here is the rub though…at least I think it might be. There is a point of diminishing returns with personalization efforts. Someone is probably going to tell me I am an idiot (again comments section, please) but doing true 1:1 personalization ultimately becomes sort of pointless and inefficient. I do subscribe to the micro-targeting and small segments. What is that magic number? I have no clue; I wasn’t very good at math.
And finally…on the creative barrier. I am going call you out on this. You must buy this, it is real (I am doing an Obi-Wan hand wave here in Raleigh as I type this). In most companies that I’ve dealt with it isn’t necessarily external agencies that are creating the content, it is internal as well. While agencies might have aided with the origination of imagery it comes down to strapped merchandising and webdev teams to put stuff on the glass. Some of the problems aren’t technological to be honest. They are process or organizational. When creating a ‘campaign’ organizations have been trained to design a creative or two…not dozens. This is a shift not just in technology to enable it but in processes and organizational behavior. Maybe Kelly and I should start a series called ‘Where Personalization Went Wrong’ and interview companies that tried to do this so we can all learn from them because I believe a lot of the problems are humans not machines. Aren’t we always to blame?