Make Educated (and Data-Driven) Bidding Decisions Based on Time Attribution

September 19, 2016
Johannes Källgren, VP Insights

For our second entry in our attribution series we want to guide SEM experts to make educated (and data-driven) bidding decisions based on the important attribution aspect of time.

The question of when a conversion is reported has a big impact on how useful the data is for decision making and how easy the data is to manipulate. The time aspect is also one of the easiest ways advertisers get fooled by the data and draw faulty conclusions from cases such as "infinite ROI" or alarmingly low CPAs.

The two main ways marketers look at SEM performance is using either click-time or transaction-time data.

Click time attribution means that every conversion is attributed back to the date and time the click(s) occurred. For more information on which click gets what value, see part 1 of our attribution series, Is Your SEM Attribution Model Lying to Your Business?

Transaction time attribution means that every conversion is attributed to the date and time the conversion occurred.

To give a simple example, we will assume an advertiser is using a last-click model. A user makes a search on Google and clicks your ad on August 1st at 1pm. They then return to your site by typing in the URL on August 10th at 9pm (this is called direct traffic and is most often not counted as a click) . When returning, the user buys a subscription of your service. Using click time, the conversion would be placed on the August 1st at 1pm, and using transaction time the conversion would be placed on August 10th at 9pm.

So what does this mean? Does it matter on what date and hour the conversion ended up?

When it comes to bidding optimization and understanding the value of investments in marketing, the answer to this is a strong "Yes, it matters".

To explain why it is meaningful, let me pull out the marketing magnifying glass and investigate how this can play out in reality with two peculiar cases.

The case of the spooky paused campaign

One of the best ways to illustrate how transaction time attribution gives some very odd results is what happens when you pause a campaign. Say that in the above example, the campaign your ad was part of was paused on August 5th. Transaction time attribution would still attribute the conversion to August 10th in this case, and you would therefore see conversions and revenue get generated through a campaign you invested $0 in. Sounds like a great deal, let’s keep doing that! Only problem is that the conversions you get from your $0 investment campaign will slowly fade out over time as you move outside the cookie window you use (usually 30 days), and you soon see 0 conversions generated in this campaign.

Now of course you would likely not believe that a campaign with $0 investment generated actual conversions, but that it was rather an odd measurement error. But this is just the easiest way to notice this impact. Imagine instead that you decide to cut your budget in half for a campaign and evaluate results using Transaction time data. Following the above example, you would see inflated numbers during the initial period after cutting your budget. This would not be as apparent as the $0 investment case, but rather hide in your numbers in the form of making it look like "you didn't lose so many conversions after all". If you use this information to decide that the budget cut was a great idea, you would soon see conversions decline over time as the high budget period starts moving outside the cookie window and all of a sudden performance is bad and you wonder "what went wrong"?

This is the basis of why it is important to use click-time attribution when making marketing budget decisions, and this boils down to every level of decisions made on how to balance investment and outcome, all the way down to granular keyword bidding.

This also illustrates how easy it is to make mistakes in budget decisions when using transaction time attribution – say that you cut your budget in-half, and a lot of your users take a couple of days to convert. This would mean that a lot of the "spooky paused" conversions would end up in the period after you cut the budget in half, meaning that it would look like you did not take a very big revenue loss and might decide that it was a great decision to cut the budget. But over the next 1-2 weeks you see results decline more and more, and you start worrying where you made a mistake – it worked great right after the budget cut, what has changed since? In this case it is the conversion lag that is at play, making conversions seemingly disappear due to your usage of transaction time reporting. Conversion lag is a big area worthy of its own post, so we will do just that, and share some more insights on how you can work with conversion lag in a later post..stay tuned.

The case of the amazing performance, that was found lifeless on day 30

Say that you have just decided to try out a new bidding tool. They are promising you that you will achieve amazing results by implementing their software, and you will see results improve every day their solution is active. All you need to do is implement their tracking and then you will be able to see how the revenue improves every day when they bid on your account. You decide to go ahead, and just as they had described you see your revenue increase more and more every day, which seems amazing! But oddly, after 30 days of a stable upwards trend, revenue suddenly flat-lines, and from that day stays around the level reached by day 30.

So why did performance stop increasing after 30 days? Well, for this example the answer is that performance was never increasing. What is the trick behind this illusion then?

Well, to create this particular magic trick you need three main components:

  • A recently implemented tracking that is used for evaluation
  • Transaction Time Attribution
  • A 30-day cookie window

The trick utilizes conversion lag to create false results. For this case let's assume the business has a conversion lag where 70% of users convert on the same day they click on the ad. What this means is that if the new solution starts bidding the same day the tracking is implemented, Day 1 will see only 70% of conversion as 30% happened before tracking was implemented. By Day 2 you see both the 70% of users who bought same day as well as the 5% lag from users who bought on the second day. This example goes on to Day 30, when you are finally seeing around 100% of user conversions, having passed the full 30-day cookie window. That means that Day 30 will seem to have 43% more conversions than Day 1 of testing did (100% / 70% - 1 = 43%). Worth noting here, is that this example was with a 30-day cookie window, which revealed the flat-line quite early, but if you had a 90-day window instead, not only would the results seem to be growing over a full 3 months, but they would also grow more as conversion lag increases with increased cookie windows.

Now that you know what is behind the curtain you won't fall for this trick and can make more informed decisions when evaluating solutions.

As a solution, if you had instead been using click-time data in this case, results would have been correct, as when Day 1 only sees 70% of conversions at first, users that were bought on this day but end up converting on later days will back-fill into this day over time and by Day 30 you would see Day 1 at 100% conversions. The drawback of the click-time approach is that you would need to wait until Day 60 to be able to evaluate the first 30 Days, to make sure all these days were at "100% of their conversion potential" by being outside the cookie window.

So why would you ever use Transaction Time data?

The above cases illustrate how important it can be to focus on click-time data, but there are some very good reasons to why Transaction time data is used and it all depends on context. Here are a few of them:

  • In any business, one of the big questions that gets asked repeatedly is "how much did we sell today?" This answer is obtained through looking at the conversions that actually came in during the day, so the answer lies in transaction time data.
  • For the question "how many conversions did we get for the money we spent yesterday?", click-time needs a full 30 days (or more depending on your cookie window) to give an accurate answer. Your boss would probably not be too happy if your answer is "let me get back to you on that in 30 days".
  • Depending on the amount of Conversion lag, you may be able to have a very strong indicator after only a few days though when evaluating performance from a click-time approach.

But to get accurate data on how well you did yesterday in terms of conversions per marketing spend, you would still want to use click-time as this will give a more accurate answer than transaction time. So how can you do this? Good news – you can estimate it! In our next part of this series on Attribution we will discuss conversion lag in further detail and give you some tools so you can yourself estimate conversion lag and thereby get click-time numbers that are good reflections of what performance will look like, without having to wait the full 30 days before you can make a decision.

As [24]7.ai Predictive Bid Optimization focuses on how much a Keyword bid should be set at for any given time, we and others interested in maximizing their Paid Search performance will find click-time data much more interesting and valuable. Click-time data informs the advertiser when the best time to buy traffic is, which is exactly what we want to know!

To consult with one of our experts on how you can optimize your SEM programs to drive better conversions, click here.

Johannes Källgren, VP Insights
Johannes Källgren, VP Insights

Add new comment