Personalization, Segmentation or Targeting

Following all my recent posts on targeting (or personalization), I was discussing website content changes with a colleague. I was explaining how we could test some form of interactive, real-time changes on our site.  His comments were that this wasn't real 1-to-1 personalization and what I was actually doing was just segmentation and content retargeting.  This started me thinking, and so I'd like to share my thoughts on 1-to-1 targeting is possible, easy and worth the effort.  Or should we be satisfied with segmentation and retargeting?

1-to-1 targeting requires the ability to show any content to any user. It probably needs a hige repository of content that can be accessed to show content that isn't shown to other users, but which is deemed optimal for a particular user.

1. How do you decide which type of user this particular user should be classed as?  

2.  How do you determine which content to show this particular user (or type of user)?

3.  When the targeting doesn't give great results, how can you tell if the problem is with 1. or 2.?

And, as a follow-up question, why is "targeted" content drawn from a library held in higher esteem than retargeting existing content? Is it better because it's so difficult to set up?

Content retargeting - moving existing content on the page - does not require new content, but "isn't real 1-to-1 targeting."  This is true, but I would argue that the difference - mathematically at least - is negligible.  The huge library of targeted content isn't going to be able to match the potential combinations of content that can be achieved just by flipping page content around to promote a particular group of products.

In previous examples on targeting, I've looked at having four product categories that can be targeted.

How many combinations are there for the four products A, B, C, D?

4 * 3 * 2 * 1 = 24

There are four options for the first placement, leaving three for the second placement, two options for the third and only one left for the final place.

This is a relatively simple example - most websites have more than just four products or product categories in their catalogue (even Apple, with its limited product range, has more than four).

Let's jump up to six products:

6 * 5 * 4 * 3 * 2 * 1 = 720.

At this point, retargeting is going to start scaling far more easily than 1-to-1 personalization. 

Admittedly, it's highly unlikely that all 720 combinations are going to be used and shown with equal probability - we will probably see maybe 6-10 combinations that are shown most often, as users visit just one or two product categories and identify themselves as menswear, casual clothes, or womenswear customers.  The remaining three or four categories aren't relevant to these customers, and so we don't retarget hat content.  I mean: if a user is visiting menswear and men's shoes, then they aren't going to be interested in womenswear and casual clothing, so the sequence of those categories is going to be irrelevant and unchanged.

So, we can group users into one of 720 "segments", not based on how we segment them, but how they segment themselves.  This leads to a pseudo-bespoke browsing experience (it isn't 1-to-1, but the numbers are high enough for it to be indistinguishable) that doesn't require the overhead of a huge library of product content waiting to be accessed.

When does the difference between true personalization and segmented retargeting become indistinguishable?  Are we chasing true 1-to-1 personalization when it isn't even beneficial to the customers' experience?

I would say that it's when the number of combinations of retargeted content becomes so large that users are seeing a targeted experience each time they come to the page.  Or, when the number of combinations is greater than the number of users who visit the page.  Personalization is usually perceived - and presented - as the holy grail of Web experience, but in my view it's unnecessary, unattainable and frequently unlikely to actually get off the drawing board. Why not try something that could give actual results, provide improved customer experience and could be set up this side of Christmas?

Other Web Analytics and Testing Articles I've Written

How not to segment test data (even if your stakeholders ask you to)

Web Analytics - Gathering Requirements from Stakeholders

Analysis is Easy, Interpretation Less So
Telling a Story with Web Analytics Data
Reporting, Analysing, Testing and Forecasting
Pages with Zero Traffic

Why Personalization Programs Struggle

So why aren’t we living in a world of perfect personalization? We've been hearing for a while that it'll be the next big thing, so why isn't it happening?

Because it’s hard.  There's just too much to consider, especially if you're after the ultimate goal of 1-to-1 personalization.

In my experience, there are three areas where personalization strategies come completely unstuck.  The first is in the data capture, the second is the classification and design of ‘personas’, and the third is in the visual design.

1. Data capture:  what data can you access?

Search keywords?
PPC campaign information?
Marketing campaign engagement?
Browsing history?
Purchase history?
Can you get geographic or demographic information?
Surely you can’t form a 1x1 relationship between each individual user and their experience? 
Previous purchaser?  And are you going to try and sell them another one of what they just bought?
Traffic source:  search/display/social?
What products are they looking at?
What have they added to basket?

2. Classification:  how are you going to decide how to aggregate and categorise all this data?  

Is it a new user?  Return user?

And the biggest crunch:  how are you going to then transfer these classifications to your Content Management System, or to your Targeting engine, so that it knows which category to place User #12345 into.  And that’s just where the fun begins.

And how do you choose the right data?  I'm personally becoming bored of seeing recommendations based on items I've bought:  "You bought this printer... how about this printer?" and "You recently purchased a new pair of shoes... would you like to buy a pair of shoes?" As an industry we seem to lack the sophistication that says, "You bought this printer - would you like to buy some ink for it?" or "You bought these shoes, would you like to buy this polish, or these laces?"

3. Visual Design

For each category or persona that you identify, you will need to have a corresponding version of your site.  For example, you’ll need to have a banner that promotes a particular product category (a holiday in France, the Caribbean, the Mediterranean, the USA); or you may need to have links to content about men’s shoes; women’s shoes; slippers or sports shoes. 

And your site merchandising team now needs to multiply its efforts for its campaigns. 

Previously, they needed one banner for the pre-Christmas campaign; now, they need to produce four, five or more instead.  This comes as they are approaching their busiest period (because that’s when you’ll get more traffic in and want to maximise its performance) and haven’t got time to generate duplicated content just for one banner.

Fortunately, there are ways of minimizing the headaches that you can encounter when you’re trying to get personalization up and running (or keeping it going).

Why not take the existing content, and show it to users in a different order?  Years ago, there was a mantra (with a meme, probably) going around that told us to 'Remember: There is no fold' but I've never subscribed to that view.  Analytics regularly shows us that most users don't scroll down to see our wonderful content lying just below the edge of their monitor (or their phone screen).  So, if you can identify a customer as someone looking for men's shoes, or women's sports shoes, or a 4x4, or a hatchback, or a plasma TV, then why not show that particular product category first (i.e. above the fold, or at least the first thing below it)?

4. Solutions

The flavour du jour in our house is Airfix modelling - building 1/72 or 1/48 scale vehicles and aircraft, so let's use that as an example, and visit  one of the largest online modelling stores in the UK, Wonderland Models.

Their homepage has a very large leading banner, which rotates like a carousel around five different images: a branding image; radio controlled cars; toy animals and figures; toys and playsets; and plastic model kits.  The opportunity here is to target users (either return visitors, which is easier, or new users, which is trickier) and show them the banner which is most relevant to them. 

The Wonderland Models homepage.  The black line is the fold on my desktop.

How do you select which banner?  By using the data that users are sharing with you - their previous visits, items they've browsed (or added to cart), or what they're looking for in your site search... and so on.  Here, the question of targeted content is simpler - show them the existing banner which closest matches their needs - but the data is trickier.  However, the banners and categories will help you determine the data categorization that you need to - you'll probably find this in your site architecture.

However, here's the bonus:  when you've classified (or segmented) your user, you can use this content again... lower down on the same page.  Most sites duplicate their links, or have multiple links around similar themes, and Wonderland Models is no exception. Here, the secondary categories are Radio Control; Models and Kits; Toys and Collectables; Paints, Tools and Materials; Model Railways and Sale.  These overlap with the banner categories, and with a bit of tweaking, the same data source could be used to drive targeting in both segments. 

As I covered in a previous blog about targeting the sequence of online banners, the win here is that with six categories (and a large part of the web page being targeted), there are thirty different combinations for just the first two slots, with six options for the first position, and five for the second.  This will be useful as the content is long and requires considerable scrolling.

The second and third folds on Wonderland Models.  The black lines show the folds.

Most analytics packages have an integration with CMS’s or targeting platforms.  Adobe Analytics has Target, which is its testing and targeting tool.  It's possible to connect the data from Analytics into Target (and I suspect your Adobe support team would be happy to help) and then use this to make an educated guess on which content to show to your visitors.  At the very least, you could run an A/B test.

5. The Challenge

The main reason personalization programs struggle to get going is (and I hate to use this expression, but here goes) that they aren't agile enough.  At a time when ecommerce is starting to use the product model and forming agile teams, it seems like personalization is often stuck in a waterfall approach.  There's no plan to form a minimum viable product, and try small steps - instead, it's wholesale all-in build-the-monolith, which takes months, then suffers a "funding reprioritization" since the program has nothing to show for its money so far...  this makes it even harder to gain traction (and funding) next time around.

6. The Start

So, don't be afraid to start small.  If you're resequencing the existing content on your home page, and you have three pieces of content, then there are six different ways that the content can be shown.  Without getting into the maths, there's ABC, ACB, BAC, BCA, CAB and CBA.  And you've already created six segments for six personas.  Or at least you've started, and that's what matters.  I've mentioned in a previous article about personalization and sequencing that if you can add in more content into your 'content bank' then the number of variations you can show increases exponentially.  So if you can show the value of resequencing what you already have, then you are in a stronger position to ask for additional content.  Engaging with an already-overloaded merchandising team is going to slow you down and frustrate them, so only work with them when you have something up-and-running to demonstrate.

Remember - start small, build up your MVP and only bring in stakeholders when you need to.  If you want to travel far, travel together, but if you want to travel quickly, travel light!

Targeted Banners: A study in permutations

"How are our banners performing?"

It's a question I'm being asked increasingly frequently, as we step up our on-site marketing. And banners are banners: they've been around for years; customers are accustomed to them (and possibly tired of them) and the challenge is to make them fresh, useful, relevant and just plain interesting. However, banners are easy, straightforward and simple to execute, measure and analyse.  You didn't think I was going to recommend adding banners without commenting on KPIs for them?

"Which banners are doing the best?"

So the challenge becomes: how do we make sure we have the right banner for the right customers?  How do we drive clicks and - more importantly - increase revenue?

Some form of targeting helps, and this could be keyword, or geo targeting, or behavioural targeting.  But why not try promoting multiple products instead of just one (or just one family of products)?  If you can split a banner slot into two, you can promote twice as many products in the same space.  And, if you can use some form of targeting, then the options for what to show increase significantly. 

Let's take an online toy retailer as an example.  What would you promote on your home page if you were an online toy shop?

There are several categories you might want to feature:

Construction toys

Dolls

Dinosaurs 

Cars

Robots

Board games 

Outdoor toys 

However, your website design only allows space (or 'real estate') for three.  And besides, you've found that having more than three diluted the effectiveness of them - your visitors get "banner blindness".  So which three do you show?  

You could determine which three to show based on various factors:

What does the customer search for?  If they search for "Lego" or "Jurassic Park", then they'll probably appreciate the banners for construction toys and dinosaurs.  Setting up some form of tracking on search usage across the site, and then matching this to the banner categories will enable you to show content that's more likely to be appealing to your visitors.

Better still, what was the inbound keyword that your user searched for?  I know Google doesn't share natural search terms (for some strange reason, the tracking only applies to Google paid search terms), but if you can access the search term that the visitor used when she came to your site, then you can start targeting her from the moment she arrives on your site, and that's a key advantage.

Alternatively, which marketing campaign did she click when she came to your site?  Was it an email that advertised your range of sports toys?

Or which pages has this visitor been viewing? If they're browsing your pages on Barbie, Sindy and similar, then this gives you a better indication of her purchase intent.

What has this customer purchased before? This will take longer for your targeting to initialise as you'll have to wait for your customer's first purchase, and once it's running it will be less dynamic than the other methods, but will be more specific as you'll know that this user has made a purchase in this category before.  The segment will be smaller, but have a higher likelihood to purchase.  Thinking outside the toy store example to other industries, maybe you could target your banners based on items added to basket (cart); videos viewed on site (what were they promoting); PDF downloads and other success events on your site.

So, you can see, there are various options for how you target, and you can then determine how effective each method is, through testing.  The model doesn't have to be perfect, it just has to be testable, and you can be confident that your testing model will go some way towards making your site more relevant (and better converting) for your visitors.  After all, there are so many variations, surely there's a good chance that you'll find a better set of banners than the one you show to all your generic customers.

So, how many permutations are there?

In our example, there are three slots available on the site, and we have seven different banners we can show.  Here are some example images, taken from various online sites.  In practice, these would be more uniform in design and messaging.

We have three slots in total:

The first slot could be filled by one of seven images.

The second slot can be filled by one of the remaining six images.

The third slot can be filled by one of the five images that we haven't used yet.

7 * 6 * 5 = 210 permutations (we use permutations here, because we can't use the same banner in two slots - that would give us combinations, and would be an even larger number).

If we decided we wanted to use only six of the banners - for example if we decide that board games aren't relevant any more - then the calculation would be:

6 (for the first slot) * 5 (for the second slot) * 4 (for the third slot) = 120

Conversely, if we introduced soft toys as an extra variation, so that we had a bank of eight banners altogether, then we'd have:

8 * 7 * 6 = 336

This is an example of "permutations without repetition" - the sequence of the banners is important, and we only show each banner once (we don't repeat them).  I recommend this site for more on the calculations of the number of permutations.  I've also written an article explaining the difference between combinations and permutations (they're not the same thing). The short answer is that the more slots you have, and more banners you have, the more permutations there are (significantly increasing with scale), and the greater the likelihood of showing the best banners to your users.

So, target your banners - you'll be able to dynamically target your content to your users, and start to reduce the guesswork from your marketing.  Even the smallest increases in possible locations or banners will rapidly improve your chances of presenting the ideal banner (if not the ideal permutation) to your users.