Predicting Sales from Online Buzz

The relationship between online buzz and business outcome (sales, market share, subscriptions, etc.) is a topic of increasing interest to many businesses.

The thought is compelling: What if you could listen to the online debate and precisely predict how much you (or your competitors) are going to sell next month or quarter?

Actually, in many cases you may just be able to do that.

Before I move on to a real life example of how online buzz and sales are related I would like to take a step back and explain the rationale for why that relation exists.

The typical way an organisation drives sales is via some sort of market communication; be it advertising, PR or some other form of activity.

In its simplest form the chain works like this: A company runs an advertising campaign. Those who are exposed to the campaign may be impacted in several ways. Some may accept the positive message conveyed while others may just become more aware of the brand and the offering presented. In all circumstances the brand moves up in our attention. We become a little bit more aware of the brand than we were before.

This increased awareness (as well as the message presented) leads to increased conversations about the brand and/or the message. The increased awareness (and the offering) usually also leads to more sales. This, of course is nothing new. If it wasn’t true, the £100 billion advertising, marketing and PR business would be resting on a sham. But advertising actually works – and more often better than traditional models capture.

A practical example may look something like this: A mobile phone operator (let’s call it ‘Mobuzz’) runs a fairly traditional advertising campaign on TV, billboards and online and in doing so exposes hundreds of thousands (maybe millions) to their message.

Even if the campaign is pretty average the very least it achieves is that it gives Mobuzz a little higher mindshare among those who are exposed to it. However it does more than that. It actually impacts those who are not exposed to the campaign as well. It does that via word-of-mouth.

Imagine two friends having a chat. One person, who hasn’t been exposed to Mobuzz’ campaign, tells the other that he is considering changing his phone provider as his contract is coming to an end. The other person may offer some advice or relay some personal experience, but because this person has been exposed to the Mobuzz campaign this person is slightly more likely to mention Mobuzz than if he or she and not been exposed to the campaign. This now raises the awareness of Mobuzz with the person who hasn’t been directly exposed to the campaign and even though the increase in awareness is miniscule this person is now slightly more likely to choose Mobuzz than before this conversation occurred.

If you could listen in on this kind of conversation between friends, you would be able to detect which brands gets mentioned more (and less) in which contexts and thereby start to form an opinion on which brands are gaining more (and less) mindshare. This in turn would be a strong predictor of future sales.

Aside from the fact that the thought of such monitoring is scary (and probably illegal) it is also not necessary if somewhere else we can capture a representative sample of the conversation. Enter the Internet.

Online, individuals, media and other organisations ask questions about brands, discuss them, rave about them and air their grievances. By capturing these conversations and processing them correctly we have, for many – but not all topics, a large representative sample of the conversations.

Now, turning the observed online debate into actual sales predictions is not as straight forward as you may think.

You may think that if there is a sudden increase in the debate about brand A, compared with the competing brand B, this will lead to more sales for A (compared to that of B’s). However, that would be too simplistic.

Very importantly, you have to adjust for the ‘weight’ or ‘influence’ of the voices that speak. If I write a story about Hyundai cars on my blog it doesn’t carry as much weight as if The Times driving section or a blogger with authority on this brand does it.

Therefore, raw buzz monitoring, where all voices are treated with equal importance, is mostly useless to predict anything.

There is a lot of good research to show that the more influence a brand accumulates, the larger share of the market it will gain.

This is intuitive as well: If a greater proportion of the talk is about a particular brand, this brand is going to be at the front of mind of more people. And if those who talk about the brand are voices of influence it is intuitive that it will have greater impact than if said voices have low influence.

In fact, most research I am aware of more than suggests that, over time, relative market share among brands will equate to relative share-of-influence.

Figure 1 (below) shows the percentage change in monthly influence and percentage change in monthly (UK) sales for Nissan Pathfinder for the period of February to July 2007. We can see that the two graphs follow a similar path and they appear to be strongly correlated. In fact, the Pearson product-moment correlation coefficient (a metric often used in PR and Advertising research to measure the relationship between observations and outcome) shows a correlation of 0.99 (the scale is -1 to +1); a very, very strong correlation.

For Pathfinder we can see that the typical lag between earned influence and sales seems to be shorter than one month. This is not always the case but in my experience it is typical for cars that have been on the market for some time and thereby well known to potential consumers.

Figure 2 (below) shows the similar data for Nissan Qashqai. The period here is from March to July 2007 (introduced in February).

At first glance the two variables seem to be less correlated, but in fact the correlation (0.98) is almost as strong as for Pathfinder. There is however, a lag between the earned influence and the sales.

Notice how the change in influence increases sharply from April to May while the change in sales increase slows.Change in influence then falls from May to June, only for the change in sales to follow with a one month lag.

The lag may be caused by the fact that this is a new model. If this is the case then it is likely that the lag will become shorter as more become familiar with the model.

The lag may also be explained by delivery shortages, but regardless of cause, the point is to illustrate that measuring influence is a highly effective way of understanding where the sales are heading.

There is much more to be said about how analysis of online buzz can be used to predict future sales and market share, but I hope the above has raised the awareness of some of the possibilities.