Optimizing Conversions on Low Traffic Websites

A common lament among web developers is how to optimize a website with low traffic when techniques like split-testing (or A/B testing) and multivariate testing require large amounts of traffic to reach any meaningful conclusions.

The Traffic and Confidence Conundrum

Why do we need large amounts of traffic in order for A/B testing to work? It all about confidence. You have two versions of copy for your landing page and you want to check which one is better at converting customers to users. The question you're really asking is : How many times should I show each of these two versions to visitors before I can assert with confidence that version A is better than  version B?

Due to the nature of statistics, you can never be 100% sure that version A is better than version B no matter how many users come through. You can be 99% sure, 99.9999% sure but never a full 100%. Web developers usually settle for something like 95%.

Let's take an example.

• version A shown 100 times and it converts 1 user
• version B shown 100 times and it converts 2 users (Double!)

Is version B better? Not so fast. You'll need another 1400 visitors or so before you can reach a conclusion (90% confidence)

It gets worse as the percentages of conversions are closer together (version A - 5% and version B - 6%)

• version A shown 100 times and it converts 5 users
• version B shown 100 times and it converts 6 users

You'll need 5438 more visitors to figure out which is better.

But, wait, there's more!

Cut Your Sales In Half... You know, for Science.

Why do we split-test in the first place? To sell more and sell better. The very nature of split-testing is that you're showing "bad" copy to half your users and losing potential sales and conversions. For low-traffic sites, you can end up losing a lot of customers in search for the optimal version.

Let's say every user you convert is worth a $100 and you decide to test two versions of your web page copy.

• version A shown 1500 times and it converts 50 users
• version B shown 1500 times and it converts 5 users

That's some copy in version A! version A is the clear winner (99.9% confidence levels). Version B is terrible, converting only 5 users. What if you had funneled all 3000 users through the version A page? 100 users would have purchased $100 worth of product each. You just lost $4500 ( 50 users minus the 5 version B managed to convert). You took 1500 hard-earned pageviews and flushed them down the toilet.

That's not even the half of it. You're going to lose money like this every time you set up a test to see if a pretender can unseat your current killer web copy.

The Holy Grail: Optimize and Test Concurrently.

Let me ask again: why do we split-test? To sell more and sell better. In split-testing, it takes you 3000 users to learn which version is better and that leads to wasted opportunities.

What if you could optimize your web-pages on the fly as these 3000 visitors flow through your site so that you can quickly learn which version is better and maximize the conversions without losing too many users to bad copy? With machine learning techniques, you can! It's called a multi-armed bandit problem.

Let's look at your split-testing problem like this:

Each web page version is like a slot machine. One of these machines gives you money once every 10 spins (a 10% conversion rate), the other one gives you money once every 50 spins ( a 5% conversion rate). You have no idea which machine is which. You get 200 spins total (200 visitors). Your mission is to maximize the amount of money you win by choosing the right machine.

It's a well-known problem with solutions that are reasonably optimal. What they all share is the methodology of testing and optimizing concurrently to maximize revenue.

The simplest one goes like this:

If you have two versions, for any visitor entering your website,

1. show them the best-performing version (so far) 90% of the time
2. show them the other version 10% of the time
3. re-evaluate their conversion rates to decide which is the current leader

What if you picked version A early on as the best and showed it 90% of the time while version B is the actual best? That's the beauty of this method. At some point, if version B is indeed the better one, its conversion rate will beat that of version A (it'll convert its 10% better) and the algorithm will switch over and start using version B as the 90% version.

There are more advantages. You can use this method to add more variations of web copy on the fly without affecting your conversion rates the way split-testing does.


If you have 10 versions, for any visitor entering your website,

1. show them the best-performing version (so far) 90% of the time
2. show them a completely random choice from the other 9 10% of the time
3. re-evaluate their conversion rates to decide which is the current leader

Even Better, Do This Automatically.

What I just described just is one of the basic algorithms I've built into SubIntent to optimize your conversion rates. The best part about this is that you don't need to do anything. You set this up and it's optimizing your pages 24x7.

You set up the variations on SubIntent, add some javascript to your web-pages and, voila, conversion rate optimization. SubIntent is also capable of showing users the best calls-to-action based on the geographical location or their referrer site that brought them to you. If you're interested in using SubIntent for your web application, you can learn more about it and sign up here.

Do you have a web app?
SubIntent can optimize your landing pages for improved conversion rates.

Using machine learning techniques, SubIntent will optimize your landing pages for maximum conversions and click-throughs. See how.