Some retailers in particular have made a great deal of extra profit through offering clever recommendations to their customers so much so that the notion of the recommendation engine came of age in the last three or our years spurring a wave of new offerings form software vendors and plenty of noise in the blogosphere. Before you go running off to buy one or even nurturing plans t build one, you should give due consideration to exactly what you expect it to do for you and from there you will be better equipped to decide if it is the right thing and finally what type of engine you need and how to acquire or build the right thing.
Ins and outs of recommendations and personalization
Perhaps the oldest and best known recommendation engine is the one used by Amazon.com. This is sometimes claimed to be responsible for 35% of sales. If that’s the case then it’s not hard to see why there is a strong interest from the ecommerce community. Every customer who selects s product then receives a number of recommendations to other products she may like . Since the customer s usually there to browse, she can live with the annoyance of being sold to and is statistically reasonably likely to find the suggested product worth looking at even if it were driven by a schoolboy randomizer function. A portion of that 35% would undoubtedly be equally achieved by a placebo tool and I would strongly recommend some experimentation before spending large sums.
The down side of recommendation engines can also be potentially substantial. When Microsoft first experimented with personalization on their website they were a leader in innovation on the internet. I was a regular user at the time and I remember being frustrated by my inability to find something that a colleague was recommending me to. The fact was that when I visited, my cookie told them I was of type A and these widgets were only of interest to type B. I t took them about a year to realize their mistake and loosen off the personalization rules.
In the past year I have had similar experiences with Google search. It is now so focused on commerce that it sends me results it believes I want to see rather than a list of cold hard facts that I want and need. These are only the few occasions when I became aware of the filters. How much of my online activity is tailored to a weird misconception of me created by a mad algorithm. Even I don’t have a great idea of what I’ll like tomorrow and that’s how I like it.
Do you want to risk excluding products from your customers because Mr customer looked at something last year that suggests he would not be interested in X. Imagine buying a Vegan book for your best friend and never again being offered a meat menu. Ugh!
Types of recommendation engine and what they can do
The Amazon, or Netflix type of engine with which we are all familiar is sometimes referred to as content based because it uses knowledge of your stock database (i.e. content) to decide what Ms H might like and make a recommendation.
The simple version is that:
Product (a) has been tagged to be about [1,2 and 3]
Product (b) has been tagged to about [3,4 and 5]
You looked at products A and B from a long list therefore there is a strong likelihood you will like other products about  but possibly also about [1.2.4 and 5].
Of course the algorithms are somewhat more complex, but hopefully you get the gist.
It can also mine the records of previous customers and genuinely say, people who selected product A also Bought product Y and Z. This will have a reasonable potential to be useful also to the customer.
Provided you are searching with intent, this type of implied logic can quickly build a useful picture of what to recommend. If you are just browsing then this interference could be just plain annoying.
The most reliable statistic however when it comes to shopping is that the more things a customer sees the more she is likely to spend, so even the mistakes are not that serious. Remember also that what works for books or movies may not work so well for other products or services.
The key to this type of engine is that it needs little knowledge of you the customer, it takes as inputs knowledge of the content and of what you searched for and how you reacted to the search results.
That is good behaviour, it has no preconceptions and it takes you at face value based on what you do.
Other engines receiving a lot of attention now are referred to as collaborative filtering engines.
These engines use vast amounts of data collected in various ways to form opinions about you and use those opinions to show what they think you will like. Some of the data in uses is controversial third party cookie data that is collected without your explicit permission.
Every action by a customer is a piece of information that potentially says something about that customer and the combination of these actions says a little more.
A simple method id to mine click streams and create segments based on identical click streams. Suppose that a high proportion of customers form segment B purchase product Y and your clickstream data puts you in segment B then guess which recommendation my engine will make.
Other information that may be collected and used against you is your interaction in social media. Who you are connected to says the type of people you like and a profile created from the commonest likes expressed by members of this group can be applied by default to you the moment you are seen to me a member of the group, any accuracy this profile has will then improve as a result of your on-going interactions with the engine via its recommendations plus any likes or other social sharing you, or your associates may express.
What is different about this method is that very little needs to be known about the content or stock in order to make predictions about the customers interests, it all comes from social an other interactions.
In theory at least, such an engine can recommend the white box to you with confidence, not knowing what it contains, simply because your colleagues whose tastes most resemble yours all bought the white box.
In reality a combination of the two methods works better because it uses some knowledge of the customer alongside some knowledge of the content to make a more intelligent match with a better likelihood of success. For the second method you do of course need to spend a considerable time collecting the useful data before you can make a start on creating recommendations, though many commentators grossly overestimate just how much data is required. Analysing 10m transactions (data points) wont necessarily give you a significantly better result than analysing 100,000 and certainly not sufficiently better to cover the extra cost.
A further problem with the big data approach is that data is time bound and therefore data form last year may or may not be valid this year. It may as easily be detracting form the result as adding to it.
A large amount of user data may well be relevant to a surprisingly small segment of heavy users whose needs are very different from most of the people you want to offer recommendations to.
People learn and change and the world changes. In 2014 attitudes are very different indeed to what they were on 2004 or even 2010. Most people I know have changed substantially in the past five years and a great deal of surfing is without doubt serendipitous. Google’s Z Moment of Truth is an interesting approach to discussing this subject.
Currently there are armies of start ups offering to find you the perfect restaurant, or movie or whatever and to my experience they are a long way from way form delivering on the promise even if I wanted to be told what to do.
My personal experimentation with Siri ended in abrupt divorce after just a few days and google’s sad attempts at knowing me have gone via the same route. The owner of Ness claims as his mission to: “become that trusted source for people to find out the next thing they’ll like.”
Isn’t that what advertisers have been doing since the first TV invaded the first living room?
There is no doubt that at a certain level for certain industries, recommendation engines can deliver substantial extra revenue and that is always a vote in favour, but the right one for the job is important.
For others a revenue upside sufficient to justify the cost can be achieved without causing damage to the user experience, but for some industries, some or maybe most recommendation engines will struggle to improve revenue an may well have serious detrimental effect of your user experience and therefore your brand