While still on the topic of brand vs. non-brand search, here’s an interesting post by Danny Sullivan earlier this month on the “search funnel” worth reading.

The “search funnel idea” suggest that that high-converting branded (alternatively, highly specific term searches) searches are often recently preceded by non-brand searches (alternatively, very broad search terms) from the same user, and so these brand searches (broad) should be considered as incremental. Sounds good — the only problem is that it often isn’t true in practice.

We did a study on the funnel effect last year for SES NY 2006, and found the effect to be small.

Here are some salient snippets from the Travelocity post, coming from Travelocity’s chief marketing officer Jeffrey Glueck (emphasis mine):

Someone might have done a generic travel search, then later searched for a specific travel brand and made a booking. The branded search gets the credit for the sale, but a non-branded term helped. [However], I think many assume non-branded terms generate more than a 4 percent assist rate when it comes to helping branded terms to convert.

Another snippet, again emphasis mine:

We studied whether people are starting on a nonbrand term like Hawaii vacation deals and, through multiple searches on a search engine, narrowing down to their favorite brands through multiple paid clicks and then, in some magical way, always going to the brand name and clicking on the paid link before buying. That is the funnel theory. If that were the majority of purchases, then the portfolio theory would make sense. We’d be OK to lose money on Hawaii vacation deals and make money on branded search terms… [We learned that] you should do the research on your own brand and nonbrand click behavior and don’t let brand profits fool you into overbidding on nonbrand terms.

Yep. While results may vary retailer by retailer, in general, we strongly agree with Mr. Glueck. The “assist” effect is small.

Here’s our funnel effect study from SES 2006: “Clickstreams, Complexity, and Contribution: Modeling Searcher Behavior Using Markov Models”. (pdf slides)

For this study, we used a Markov chain to model the steps in a user click-stream.

Using Danny’s terminology, our analysis showed the “assist” effect to be quite small — just as Travelocity reported.

We’re in the process of updating this funnel analysis, and will release a more substantial report later this year.

(Strange coincidence — we too used the “Hawaii vacation” search for our example for the SES talk, as I had just returned from visiting cousins living in Maui . A hui hou!)