Barney Pell's Weblog

« Barney moderating a panel on Vertical Search: Show Me The Money | Main | Magic Wallpaper »

September 20, 2005

Highlights from tonight's vertical search panel

This evening's event on business models in vertical search (see my previous post ) was fun. I estimate about 150 people attended. Half the audience raised their hands claiming to be working in a vertical search startup.

Since I was the moderator, I was unable to take careful notes (I hope someone in the audience will post their notes to the blogosphere). Below are some points that I did note down or remember.

• Definitions: We essentially agreed that the phrase "vertical search" has come to be used synonymously with "specialized search". Search can be specialized along many dimensions (including content, users, and context) and is a matter of degree (leading to more or less deep specialization).

• The "Vertical" distinction here is independent of the horizontal/vertical application distinction (tools for HR are horizontal as they cut across all companies, while tools for securities trading are vertical as the apply only to that industry segment).

• Vertical search applies to searching all forms of data (varying from structured databases to unstructured text documents).

• The additional constraints and context of vertical search should enable improved search performance (precision, recall, usability, integration, convenience). Hence we should expect to see fragmentation of general search into a set of verticals.

• Moreover, this fragmentation would likely continue (so if Pets is a valuable vertical today, then Dogs is one too, and perhaps there is even a market for Poodles.com). To prove this point, Scott Rafer planted the audience with representatives from Dogster.com (complete with dog costume). The Pet-Web is not easily accessed via a general search engine, and this 3-person company is profitable, with 200,000 unique visitors/month and 125,000 registered users!

• Vertical search businesses are combined media and technology businesses. They can be supported like any other such businesses, including advertising, subscription revenue, and paid content models. Successful companies will focus and define a core audience and develop offerings to serve the informational needs of that audience. It may even be more profitable to lose audience members that are not core, as that provides a more focused audience for your advertisers.

• As an example of the potential benefits of supporting a focused audience with what is otherwise a generic search, I pointed to big.com. This site was created as a result of user testing during the development of Snap.com (a Mayfield company, and one for which I am an advisor). During the testing, it turned out that a segment of very happy users were much older than the typical internet search demographic. When Snap asked these users why they liked the site, they explained that they preferred the big font that Snap happened to be using, as it was as easier to read. With this insight, Snap recently launched Big.com, the first search engine that lets users increase the font size with a single click on an icon. This site will undoubtably have an older demographic than most search sites, and may well get higher advertising rates than the main snap.com by virue of the focused (and economically important) demographic.

  
  

• Compared to pure media companies, who pay to produce content, search businesses have the nice property that they draw on "other people's content". Such content can be continually refreshed at little cost to the search provider, creating an "evergreen" product offering. However, a thin human layer can amplify the value, as Scott Rafer illustrated in the case of the "Feedster 500".

• Past vs. Future searches: Paul pointed out an interesting asymmetry in past-oriented searches (search existing content) vs. future-oriented search (alerts, standing queries, push content). With the former, the user can iterate quickly over many different ways of expressing the information goal until he is satisfied with the results. With future alerts, by contrast, users have a hard time writing a query that captures what they will find interesting, lacking concrete feedback. (As I reflect on it now, I'm unsure why you can't just debug your queries on previous data and then store the best results as standing queries...).

• I wrote down this quote that Scott Rafer attributed to Mark Evans (Blogger founder): "Don't mistake clear vision for short distance". That seems like a great caution for visionary entrepreneurs and VCs alike.

• What makes a vertical big enough to be interesting? (Answer: all marketing and competitive strategy elements apply. But lasting uniqueness is certainly a big part of a good answer).

• Search boxes on the major portals are becoming the starting place for many internet users. When the portal recognizes a query that can be handled by a vertical system, the query is routed to that expert. If all searches go through the portal's start box, then how does this affect the prospects for vertical search providers? What can they do to ensure that users will come directly to their site?
  (Social networks, user-generated content, personalization, and vertical-specific guided navigation all help).

• A few audience questions discussed business to business verticals and the connections to the $250B paid content industry.

• One audience member asked if any technology platforms existed that could enable many small vertical search sites to be created without having to build up a huge technology capability from scratch. I said that this was part of the vision for Whizbang! Labs, and that while there are some providers of pieces at present I think there is still a niche to be a platform provider. Scott Rafer disagreed, saying that you could build a vertical search directory on the cheap using an open source search engine like Lucene. I'd say that is fine if you just need to provide an unstructured search over a bunch of documents of webpages. But if you want to provide structured search from unstructured data, you need to crawl, classify, extract, compile, and normalize the data, which is definitely something not be tried at home without a strong technology team.