Affiliate Media Buy case study How to deal with very large scale networks

In the last case studies, we offered content relative to new media buyers, in this one we wanted to go a step further and offer advanced content. Yet, if you are new to media buying you can still learn a lot of useful information. Once you get success in small-scale campaigns your wish is to optimize them on a very large scale. But dealing with networks on a large scale is a different challenge, and we’ll cover it in the lines to come.

Large scale : What does it change?

If media buying techniques are universal, large scale networks are more challenging. The first thing that I usually say to a new media buyer in my team is that it’s not difficult to buy good traffic, but it’s more difficult to not buy bad traffic. What does this concretely mean in our case?

On a large scale network you will have a lot of opportunities to buy traffic from multiple placements, and even multiple sources. This, of course, gives you multiple chances to find profitable placements, but also to buy irrelevant traffic. As a matter of fact, except if you are in a Real-Time Bidding (RTB) environment, the placement will be your main mean to target the right users. The ideal method would be to spend a large amount of money and test individually each and every space, for a decent time, and then keep the profitable ones. The problem with that strategy is that you will have to spend tons of money to get insights and you will end up being profitable only in months or even years.

So what do we have to do to adapt? Are large networks only profitables for big companies? Of course, there are strategies to adapt and of course even with limited resources you can be very efficient on them!

Strategy shift

Understanding micro conversions and what they bring

In a previous case study, we saw why conversions are rare and why it is important to make decisions based on statistically significant datas. But we were talking about macro conversions. Let’s deep dive into the conversion world.

First of all, let’s redefine the word conversion. A conversion is a change of state of a user. For example, a conversion can simply be a click on an ad. He is changing his state from a passive user for whom an ad is displayed to a user that is interested in this ad, and the product or service that it embodies. This is typically what we call a micro-conversion.

On the other hand, what really interests us as media buyers is what we call the macro conversion. In other words, the final conversion that the user realizes, and the one we are paid for. It can be either a lead, a sale, or whatever step the advertiser decides to pay for. Please note that generally speaking the macro conversion can be different on the advertiser’s side. For example, an advertiser can pay you for leads, but consider that his macro conversion is a sale.

Now that we have a better understanding of what micro and macro conversions are we can redefine our conversion funnel as a succession of micro conversions that led to a macro conversion.

But what does it bring to us? Each micro-conversion marks the growing interest of a user for the offer that we promote. It means that each time a user realises a micro-conversion we can consider that is more qualified than before, and so that he is potentially profitable. This qualification process along the conversion funnel is the base of the concrete strategy that we develop.

Giving more conversions to fit our data needs : the base of our strategy

All we saw before is not pure theory, we use it really concretely in our campaign strategy. First of all, what we want to do is having less spent and so win time over a usual strategy. For that, we need more efficiency in the placements selection process. And for winning time without making poor decisions we need more conversions.

For this reason, we’ll use the qualification principle and create an intermediate landing page between the ad and the offer itself. The idea is to create a very simple to realize micro conversion that confirms that the user has a real interest in the offer we promote, and not just clicked on the ad because of his curiosity or for whatever reason. As you can imagine our landing page will be a single click landing page, we won’t ask for any personal information.

Using this strategy will allow us to optimize on the click on our landing event, and so will produce faster and cheaper data. We will of course ultimately optimize by the cost of a macro conversion but we now have a solid method to exclude problematic spaces. If their cost per click on the landing page is too high, we will exclude them regardless of their lack of macro conversion.

The concrete setup : traffic source, tracker, offer, and more

Traffic source

Of course, we were needing a traffic source that has two fundamental points for this demonstration. The first being a large scale one ! Second, being a traffic source that puts us on an equality stand with other media buyers. In other words, a traffic source that either requires or is compatible with a landing page. The native was a natural starting point in our selection, and Outbrain appears as the natural leader in this kind of traffic.

Tracker

As we want to get high volumes of traffic, we will need a quick, efficient, and high standard tracker. Well-known in this industry Bemob was a quality choice. He had everything we were needing from ultra fast redirections to S2S communications and templates for lemonads as well as Outbrain. Our choice was fastly made !

Offer

The offer also has to be compliant with native standards (ie a mainstream one) and also compatible with the idea of a pre-lander. By chance, we already had a tested pre-lander which was a love dating comparison website for the United States. Except for creatives that we’ll see later on, we had everything ready from strategy to concrete needs.

The setup itself

The first thing we wanted to do was setting up the communication between the three important poles of our campaigns : the traffic source, the tracker, and the network. Thanks to bemob's very important template collection it took only 5 minutes to set up a postback feedback on outbrain and link it all the way through lemonads. The outbrain template is greatly detailed in bemob, so that we are able to track every placements, ads or whatever needed. On the other side we are already creating a stock of macro conversion in outbrain that will help us optimize directly through their backoffice. Everything is going well on this side.

So now let’s deep dive in the campaign itself and it’s configuration :

First of all the campaign objective is pretty straightforward : we want conversions. Please note that in this case we use the macro conversion as an objective not the micro one. We could use the micro one if we already knew our average revenue per click on the landing page. But as we are starting a brand new campaign, on a new vertical and a new source, we aren’t able to tell it. So we are using the macro conversion as a direct advertiser would do.

The creative format was set to single (opposite to carousel). We wan’t to strike spirits with a simple message and also be able to clearly differentiate elements from one to another so a single format is from far the best.

The next part of the configuration was the budget itself. We first opted for a target CPA budget, with the same idea of having the less spent possible until getting results. The target CPA was determined by the maximal payout we could possibly have with advertisers. The idea was not to be from scratch profitable but just the most balanced possible, or at least not to lose money.

The other idea was to run an ab test to clearly determine an optimal CPC. The suggested epc was at 0.78 euros for this campaign. We were considering it as far too high for a learning campaign, and rather that getting the absolute best positions we wanted to stay on an average CPC for the native industry. So we set our base CPC to 20 and try a ab test based on a -50% optimization. 0.15 euros is generally the average. That way we were able to have a solid test base without overspending for a single click. The daily budget was originally set to 10 times the target CPA.

We had a standard schedule as the idea was to test every hour if possible.

For the moment we are making a really broad test so we only set our location and didn’t use audience segments or targeting.

Same thing for outbrain categories and placements we let really open. We don’t need for the moment specific platforms, neither wifi connections or high impact placements. We also use the extended network to have an idea of what other networks related to outbrain can offer. (This last tip is a good way to find new converting traffic sources).

Regarding the tracking we use a really large specter of variables that will allow us to follow closely the campaign and on multiple dimensions directly through bemob.

Now we only need one last thing : the creatives. Outbrain has an outstanding tool that allows you to fastly create some variations :

We decided to create some variations, with long and short texts, several pictures and also put in the balance creation with or without the dynamic city parameter. The idea was to get a bunch of really different texts and images to find the best clicking and converting ones. The other idea was to use the well known principles of “native sentences” as a way to leverage the CTR. For that we use several tactics : include surprising elements (“you will never imagine who is first”, the city parameter as well as interrogative sentences. The pictures test was made with the same idea in mind : with variation from people in their 30’s to their 50’s, couples or people alone etc. We end up with more that 40 unique creatives based on the combination of our samples.

The campaign is ready, time to press the launch button

The elephant problem

Generally speaking when you launch a campaign, you have a period of several days where you are letting datas appear. On a very large network you have to take actions really quickly.

As you notice on our very first day we generated 476 visits on our landing page for only 5 clicks. The first round of 50 euros was very rapidly (less than 5 minutes) disappearing. We decided to double it to have another round of data.

The most part of the problem couldn’t be seen in this dashboard, which was only as an overview. The first check was about unique visits :

More that 95% of users that had visited the landing page were unique. That’s the first good point. 73% of users were unique clickers, as we are using a comparison website people tend to click on multiple results that’s also a normal behaviour. But it was less normal that we had 2,42% of visitors that ultimately clicked our landing page (11 out of 453).

So we decided to go deeper into our analysis. The first question was related to the publishers, how do they individually perform ?

If we zoom on the first fifth one we have a really interesting problem (the famous elephant problem !)

The first publisher alone is bringing almost half of the visitors (207 out of 453 uniques), and he is having a fairly low performance regarding clicks : 2,41% of unique users only clicked. What does it mean in terms of cost :

With an amount of 53 euros the unique landing page clicks on this publisher was costing 10 euros. It was way too much for a CPA of 4.75 euros for macro conversions. So this source can only be stopped.

But the most important problem was that this publisher was taking a large amount of the budget. The cost of the second most important publisher was only 6,72 euros. That means that the most important publisher was taking only for him 50% of the budget and has a spread of x9 with the second most important publisher. Cutting this “elephant” will not only reduce our cost but also dispatch more efficiently the budget across multiple publishers that are able or likely to convert. In this example the second publisher had a cost per landing click of 3,36 ie 3 times lower than the first, and under the target CPA.

So what do we got for day 2 now ?

Overall the performance is a bit better : 14 unique clicks and 3 macro conversions. But the elephant problem appeared on two new publishers, so we decided to immediately cut them as well.

Day 3 was really really better: 21 unique clicks and 2 macro conversions and moreover 733 unique visits where we had only 457 on day 1. So we doubled up visits and out clicks.

So what does it take to come there ?

The multi dimensional optimization

During this period of 3 days we had the following elements :

• 326,43 euros spent
• 1 125 742 impressions
• 2073 visits on our landing page
• Average CPC of 0,157 euros
• 46 unique landing page clicks
• 5 conversions

That allows us to compute a lot of useful information. If we are going down our conversion funnel at that state here is the key point indicators :

• Unique landing page clicks max cost : 0,51 euros.
• Formula : Conversions X target CPA / Unique landing page Clicks
• Average visits to unique landing page clicks : 2,21%
• Formula : Unique landing page clicks / Visits on landing page
• Max CPC : 0,01 euros.
• Formula : Conversions X target CPA / Visits on landing page

Does this sound realistic ? with a 0.15 euros average CPC on native campaigns; of course not !

So the idea was to optimize not only by placement but by all means.

The first solution came from outbrain itself, the test stabilizes itself with a 0.13 euros per click, so we decided to lower it to 0.10 euros and see if we still got traction on the campaign. Already 25% won ;).

Alongside this optimisation phase, we drew another regarding creatives.

As you can see, there are clearly 2 baselines and 2 images that are really winning the contest in terms of diffusion. Based on this we created a baseline that was mixing up the two first one : “Don’t want to be single anymore in {city}”. We created a new set of creatives with the mixed one and the other winning variations only.

As it’s less directly measurable, the gain is real. Most networks pay CPM and are paid CPC, that means that the higher your CTR is, the less bid you make. Optimizing CTR is also a way to overcome competition at a lower price. Of course in our computation we also take in account the landing page click as a KPI. If our CTR is really high and the click on landers are really low, that means that we give a deceiving experience to our users, or that we are overpromising on what we deliver. In any case it’s bad for users and it’s cost us a lot of money.

We also had a look on platforms. As you can see, mobile is way more interesting than desktop. Several things can explain that for this vertical, but it’s not what really matters for us. By cutting Desktops and tablets, we are lowering our CPA from 65 euros to around 39 euros. The decision was easy: focus on mobile.

Lastly we had a look on interests, for now with only a few macro conversions nothing was absolutely clear. We decided not to optimize on this parameter for the moment.

From early optimization to long term campaign structure and processes

We now have a better understanding on how the network works by using less than 400 euros, the idea is to make a transition to a long term campaign.

For that we will use the “learning, whitelisting, blacklisting” triangle. The idea is pretty simple : on a large scale networks competition and publishers' moves tend to make it pretty unstable. Some competitors will try to overcome you and some publishers will go to another network, whereas some will come to the network you are working with.

The learning campaign will be there exactly for this reason : with a constant learning campaign you will be sûre to meet every new publisher and try their traffic. But this campaign won’t make you a living.

That’s why there are 2 ways to exit a learning campaign : either being blacklisted or whitelisted. Blacklisted simply means that you don’t want underperforming traffic and are excluding them from the learning campaign. The whitelisting system will also exclude campaigns from the learning campaign but will put them in a dedicated campaign with other performing placement.

Over the time you can elaborate more sophisticated whitelist campaigns, such as average performing one, high performing one, or one dedicated to one platform or another etc. The idea is that you will have specific campaigns for some placements or segments with specific strategies, creatives and bids in each of them.

Conclusion: The 5 things to remember for a good large scale campaign

As a way of conclusion I would like to underline the 5 tips that you have to remember :

1. There is always a way to find more conversions for the very early stage of the campaign (think about “micro conversions”).
2. Take care to not overpay in the learning phase.
3. Quickly react and adapt to the datas, yet stay confident in their statistical significance. Be careful about the “elephant problem”.
4. Think about all the dimensions available for optimization.
5. Switch to the learning / blacklist / whitelist triangle as fast as possible.