The Future Commerce podcast has some thought provoking interviews. A recent discussion was on the use of AI techniques and testing to try to automatically determine the best products to put forward first, instead of always showing the same category images. That is, adjust the experience for different users and compare their results to work out the best option – but don’t change the navigation links mid-session. That can be even more confusing.
So I thought I would share a few thoughts here.
The Category Tree
Sites organize products into a tree of categories, to allow for simpler navigation. But if you show product images as representatives of categories, should you show the same image to all users? Or should you perform effectively A/B testing of a range of images from the category to find the best converting image.
Doing any such A/B testing assumes a reasonable amount of traffic to your site, in order to make the collected observations statistically meaningful.
This can overlap with personalization. If I know the customer is female, should I be displaying different category images? What if they fall into different age bands?
In a nutshell, should the site experience change as the site learns more about the user and their preferences. My answer is in some regions of the screen this is fine (even good), others this results in confusing behavior).
Another aspect of system design is that humans like understanding how decisions were reached (in this case, recommendations).
In my past life, I spent a fair bit of time in ranking algorithms for search. The application I was looking at was where people were looking for hidden treasures, the opposite of normal web searches where there are probably 20 good answers for any question. If there are not many good answers, when bad answers are being returned the user wants to understand why. Without this, use trust in the system decreases.
So personally, I am not convinced that changing navigational images (e.g. images per category) is a good thing during within a single user’s session. Why not? Because when a user returns to the same navigation point the screen will be different. That can result in the user getting disoriented. Users expect cross sells, up sells, etc to change. They do not expect navigation to change mid-session.
Finding a Balance
This does not mean that the general approach of trying to find good content based on preferences and past history is a bad one, it just means do it in the right contexts.
For example, cross sells, up-sells, merchandising, etc – people expect ads to change on a page (including the merchant’s own ads). Experimenting to find better images there is fine. Just don’t change the navigation sign posts – don’t move the sign posts around while people are walking around the swamp. People don’t like constantly changing barriers.
Modelling user preferences is not trivial. If a user clicked on the search result for an iPhone, for how long should iPhones be recommended? As soon as I purchase my iPhone (possibly from a different store), my next purchase is more likely to be an accessary, such as an iPhone cover. If the personalization engine harps on old data for too long, this can become annoying. Stop showing me the iPhone I just purchased! A challenge here is the web site will not be aware of all
To overcome this, a simple strategy is to decay data over time. For example, a month later don’t bother showing iPhones to the customer any more. More recent activities should be given preference with older references still in as backup support.
Wrapping it Up
I think personalization, A/B testing, AI etc all have a role to play. These techniques can be used to learn how to put better content in the front of users. But apply these techniques to areas of the page where content is not expected to change. So I believe Navigation should not change per consumer, or at least not rapidaly. Merchandising panels, side bar regions, and more are however fair game for automatic tuning.
Well, if you change to the winner in production the change will affect loser variations users of your A/B test.
To avoid this, what you can do is don’t stop a/b test, wait for season to end and then do the real change. Meanwhile, let the test run with 0% trafic on looser variations (still get seen by recurrent visitors).