One of the most interesting topics talked about at SES London was the problem of conversion attribution. The easiest way to describe it is with an example. Suppose you sell widgets over the internet and that you use AdWords rank well for some organic terms and are also running a radio advert. A customer’s path to conversion might look something like this:
- They hear your radio advert
- They search on Google and click your PPC advert
- At a later date they arrive at your site organically
- They buy something from you
If you are using AdWords conversion tracking to monitor your campaigns all the credit is given to the PPC advert. If you use Google Analytics, the credit will be given to the organic keyword. Neither will give any credit to the radio ad slot which prompted the initial search. Fuor.net have an informative white paper about the problem of conversion attribution.
Tracking offline advertising is a lot harder then online (particularly PPC) so the problem is simplified a bit if we ignore it. Of course this is not a good assumption to make; another key point from SES London was that “Search does not exist in a vacuum;” people can be prompted to search for things for a variety of reasons, including TV, radio and magazine adverts.
Ignoring the offline channel still leaves us with the problem of splitting the conversion between organic and PPC. There are 4 possible options:
1. Credit the last referrer
This ignores the fact that the buyer first found the website through a PPC advert. If that advert had not been there then the conversion might not have happened
2. Credit the first referrer
Similar to above, this ignores the fact that without the 2nd visit there wouldn’t have been a conversion. For some queries the last referrer may be a brand keyword which indicates that the searcher had already decided to buy from you and was just trying to navigate to your site. In this case there may be some justification for only crediting the early referrers.
3. Credit both referrers
Crediting both referrers leads to the SEO people saying “we made 500 sales last month” and the PPC people saying “we made 750 sales last month” when fewer than 1250 sales have actually been made.
4. Partially Credit both referrers
As far as I am aware, this is the best method that anyone has thought of so far. Unfortunately, it leads to another difficult question: “how much credit should I give to each referrer?” I think it is more complicated then crediting all referrers equally.
Credit Crunch? Splitting Credit between referrers.
This is a really hard problem and anyone who comes up with a perfect solution will make a lot of money. How will you know if you’ve got a perfect solution? I think a perfect solution must have the following properties:
- If each keyword used is credited with x% of the conversion then adding up all the x’s should not be more than 100%. It would be acceptable to have the total at less than 100% if you think that offline marketing played a part in the conversion.
- A solution must distinguish between different buying cycles: Imagine one day I search for “football boots” then the next week I search for “tennis rackets” and make a purchase (of a tennis racket). Then my first search should not receive as much credit for my purchase. I’m not sure if it should receive no credit at all because maybe appearing for “football boots” as well as “tennis rackets” may have increased my confidence in the site.
- In the case of an existing customer, their terms used in their previous purchases should be partially credited with each new purchase; this builds lifetime customer value calculations right into the system.
- It must be workable. This should go without saying, but turning an idea into a workable system could actually be the most difficult step.
Avanish Kaushik thinks that time on site may be a good way to split up the conversion. I can see the advantages of this method because if a user visits your site and purchases quickly then they have probably already decided to buy from you meaning that the final referrer is not so important. Attributing the conversion in proportion to the time spent on site satisfies property 1 from above and I can see how it could also satisfy property 4 but more work needs to be done to make it compatible with 2 and 3.
My preferred solution would be to consider all conversion paths and compare the effectiveness of each step. For example, by finding all conversion paths that differ by only 1 stage the effect of not engaging with the user at a particular stage can be measured. This will fail property 4 unless you have a lot of conversions and a relatively small number of conversion paths; for most people the number of users taking each conversion path will be small because there are so many possible paths.
The conversion attribution problem is not a new one and as I’ve said, a perfect solution does not exist. I will discuss some existing methods that partially solve or improve our understanding of the problem next week.