Data Analysis in Marketing; What Google and the Flu Can Teach Us

I saw an interesting tool the other day from Google, that analyzed raw data on searches related to the flu in order to predict the severity and timing of flu outbreaks. Available at http://www.google.org/flutrends/, this tool is an interesting example of approaching data in a unique way in order to understand a problem.

Whereas an individual person searching for a flu-related topic is clearly not a diagnosed flu victim, the correlation between those searches and an actual outbreak is quite good, and more importantly, the data is available much faster than the medical diagnoses precisely because it is not relying on a properly diagnosed flu victim as the base data point it works from.

As we implement lead scoring algorithms, and other predictive approaches within our marketing automation systems, we face a similar challenge. The best way to look at whether we have scored the leads the right way is to look at whether the leads we passed to sales became qualified opportunities and eventually closed revenue. However, similar to the medical diagnoses in the flu example, this can take a significant amount of time. However, the raw data can show us some very interesting trends and give us immediate insights.

When you have built out an initial algorithm, incorporating the best practices for lead scoring, the simplest thing to do is to pass your entire dataset through the algorithm to see how they would score. This “bottoms-up” look at the data gives a very quick view of the potential results. This technique is best used when looking at the explicit score (the “who” that indicates the right buyer in the right organization) rather than the implicit score (the “how interested” that indicates the level of current buying interest. The reason for this is that the “who” is not likely to change over time, while the “how interested” will obviously vary significantly over time.

With your entire database scored through your new algorithm, the results will tell you some very interesting things.

- Were the final numbers what you expected? If you scored 100,000 contacts on the explicit dimension of lead scoring, and only 0.1% ended up as A leads, was this what you would have anticipated? If you were expecting significantly more, it could easily be a data problem. For example, if your scoring algorithm looks for a key title, such as “Vice President of Marketing”, and you have not cleansed your data, you may miss most, if not all, of the contacts you are looking for. In our own experience, that search returned only 300 results before cleansing, and over 17,000 results after data cleansing to standardize all the other ways of writing “VP. Marketing”, “Vice Pres Mktg”, etc.



- Does the sales team like what they see? If you show your sales team a sample of the leads who scored well in your new algorithm, do they think that these are the right set of leads they should be speaking with? If you show them leads that did not score well, do they agree that these are not leads they would like to speak with? Remember, of course, that you are only looking at explicit information, not buying activity in this example, so it is assumed you would only be passing your sales team these leads when buying activity was detected. Balancing your sales team’s intuition with your objective lead scoring algorithms is as useful here as it is in highlighting flaws in the process through sales “cherry-picking” of leads.


- Does history agree with your hypothesis? When Google looked at their flu data, they compared it carefully with CDC data on actual flu outbreaks to ensure that there were minimal false negatives and false positives. Similarly, your lead scoring results need to match history accurately. If you look at contacts who were scored highly, and who have also been around for a long time, is there a higher number of them who have become customers? It not, what is missing?

Marketing data can be a true gold mine of insight if you use it carefully, much in the same way that search data shows extremely interesting predictive insights when looked at in certain ways.

What insights have you found in your marketing data that surprised you?