Structured vs. unstructured

Yellow Pages folks surely do love structure — especially when it comes to data. Here at the latest Kelsey conference, where YP folks abound, the only good datum is a structured datum.

Consider the title of yesterday’s most interesting panel:

Building a Better Database: Acquiring Content in a Dysfunctional Environment

The title is a bit grad school, but “dysfunctional” is a strong word that caught my eye. Here it mostly means “resistant to structure.”

And them’s fightin’ words in the world of Yellow Pages.

By now I’ve gone to a bunch of YP-oriented conferences. All of them featured a discussion about how to gather structured data. But I’m starting to suspect that this isn’t the most important problem to solve — and not just because these conference discussions never go anywhere.

Here’s my thinking:

In what a YPer would call a functional environment, every business location, small or large, would authorize a regularly updated master version of its “attributes” (hours, certifications, parking facilities, etc.), and would post this information in some microformat on its Web site, or supply it directly to each data vendor, or send it to an industry-wide data clearinghouse that’ll probably never exist.

In addition, lots of other data sources — licensing bodies, rating sites, whatever — would distribute structured information that’s already normalized and can be correlated perfectly to these master records.

All this data would then be collated by data vendors such as Localeze and sold to Web companies such as Google or, for that matter, Loladex.

Finally, the Web companies would build applications that use the structured data for searching by consumers (input) and display to consumers (output).

This worldview may be summarized thus:

More structured data in → Better answers out.

Or as Marchex‘s Matthew Berk (who’s a smart guy) said at the panel here: “We think local search is about structured search.”

Berk gave a very good example, which I also use when discussing Loladex: If you’re looking for a doctor, you need to know whether he takes your insurance. That’s true, without a doubt.

But here’s the problem I have:

The majority of information available about any company, and particularly about any small company, will never be structured. It’ll exist only on the general Web, where it must be searched on its own terms — that is, as unstructured text.

To me, this suggests that the most pressing data problem isn’t how to gather more structured data, but how to search unstructured data (on Web pages) and return structured answers.

I live on both sides of this equation, by the way. My wife runs a small cookie bakery, and I’m in charge of distributing her data to online sources.

Because of my background, I’m more informed and motivated than most small business owners. And yet, to be honest, just keeping her Web site up-to-date is a chore. On Yelp right now, I’m sorry to say, her hours are incorrect. I should update it, but I just haven’t.

Accuracy on our own Web site is always my #1 priority, because that’s our official voice. Also it’s where most people land when they search for “Lola Cookies.”

Keeping Yahoo Local accurate is on my list, too, but it’s lower down. Ditto Google and YellowPages.com and the other big sites.

I never think about the data vendors one layer back, like InfoUSA, unless they happen to call the store. (Which InfoUSA does, to its credit.)

Meanwhile, plenty of interesting and searchable information about the bakery exists in other places on the Web, in formats that aren’t even addressed by the concept of “attributes.”

A TV broadcast from the bakery aired live on the local morning news recently, for instance. If you watched the show, you might search for us with a term like “fox 5 cookies virginia.” Where does that fit in the world of structured data?

I raised this general issue at yesterday’s panel. What were the panelists doing about this wealth of unstructured Web data, which right now is the dark matter of the local-search universe?

The answer I got was, basically, “Not much.”

Most panelists said they do only highly targeted crawls, focusing on sites that have structured data that can “extend or validate” their own data, in the words of Localeze’s Jeff Beard. An example might be the site of a professional group such as the American Optometric Association.

No panelist was ready to start indexing the sites of individual businesses, or locally focused blogs, or any other sites that are unstructured but potentially rich in content.

The only (mild) exception was Erron Silverstein of YellowBot, who also said his company limits itself to targeted crawls — but included local media, such as newspapers, among his targets.

A few players are indexing the broader Web and then associating pages with specific businesses (which is the important part). Most notable are Google and Yahoo, who do it for their local search products.

Of course, they’re already indexing the entire Web. It’s less of a stretch for them.

Google and Yahoo also buy structured data from InfoUSA, Localeze and others, so it’s not like such data is obsolete. But they’re getting the same info directly from some businesses, and those updates are likely more timely, more accurate, and more complete.

Meanwhile, their Web indices are opening up a realm of data that traditional vendors like Acxiom — represented by Jon Cohn on yesterday’s panel — simply don’t care to address.

I suspect that, sooner than you’d imagine, Google and Yahoo will be buying structured data not so that users can search it directly, but for two less-flattering reasons:

  1. To help find Web pages they can associate with each business
  2. To fill ever-smaller gaps in the coverage that results from #1

Matthew Berk of Marchex argued that a good local search must be structured to “help someone walk down the decision trail” by using filters to narrow their search progressively:

I need a orthopedist in Boston … in the Back Bay … who accepts United Healthcare.

I think users are more likely to learn that they can go to Google and type “orthopedist back bay united healthcare” — particularly if it produces a good top result the first time they try.

The burden of local search, it seems to me, is to do something that Google can’t match with an unstructured Web search.

In any case, the search portals will ultimately use their indexed Web pages to extract and cross-check structured data directly. Over time — probably just a couple of years — such automated processes will yield data that’s more current and detailed than anything that’s produced by scanning phone books or calling stores.

The resulting search functionality, integrating both structured and unstructured data, will be sold to other companies as a Web service, and data vendors such as InfoUSA will become irrelevant to local search.

Now that would be a dysfunctional environment for many of the Kelsey attendees.

I’m not sure exactly how companies like InfoUSA and Acxiom should tackle the unstructured Web. It’ll demand a new way of thinking, and probably a new way of selling.

But I’m certain that they ignore unstructured data at their peril.