..The Fuzzy Math of Home Values

But for figures that carry such weight, critics say, the estimates can be far rougher than most consumers realize. Indeed, if the websites were dart throwers, they'd seldom hit the bull's-eye, and they'd sometimes miss the board entirely: Valuations that are 20, 30 or even 50 percent higher or lower than a property's eventual sale price are not uncommon. The estimates frequently change, too, for reasons that aren't always easy for homeowners to discern. According to the companies themselves, some quotes have swung by hundreds of thousands of dollars in as little as a month as new data gets plugged into the algorithms the sites rely on. (Those algorithms also change, as happened this summer when Zillow made adjustments that affected all of the 100 million homes in its database.)

Indeed, appraisers and real estate consultants say that those models veer off target with alarming frequency. Typically, data for valuation models come from two sources: records from tax assessors and listing data for recent sales. Middleman companies -- the dominant ones are CoreLogic and Lender Processing Services -- gather this data from more than 3,000 U.S. counties and license them out to the Web sites and other model-builders. Collection is itself a challenge, because not every county tracks properties the same way.

Public records are hardly the only problem. Automated models aren't designed to account for the unique details that often make or break a deal -- something their designers readily acknowledge. AVMs usually can't capture data that determines the condition of a property, such as whether there's been a ton of wear and tear. Is a home right next to the railroad tracks or a golf course or a landfill? AVMs can't always answer those questions, say industry pros, though GPS technology is improving things on that score. Models also can't decipher the motivations of a buyer or seller, says Leslie Sellers, a past president of The Appraisal Institute. A couple who's going through a nasty divorce, for example, may have taken the first offer that came along just to unload the property. For all these reasons, says Lee Kennedy, managing director of AVMetrics, a firm that audits and tests industrial-grade AVMs, the models that banks use often add a "confidence score" to their value estimates, with a low score signaling that it's best to send in a human appraiser.

Consumers, however, don't get to see a confidence score; instead, they get disclaimers, some of which are eye-opening. Zillow surfers who read the "About Zestimates" page find out that the site's overall median error rate -- the amount the estimates vary from the actual fair value -- is 8.5 percent, and that about one-fourth of the estimates wind up being at least 20 percent off the properties' eventual sale price. In some places, the numbers are far more dramatic: Gibson County, home of the West Tennessee Strawberry Festival, has a 57 percent error rate; in Hamilton County, Ohio, where the Cincinnati Bengals play, it's 82 percent. Site users are always one click away from this data, but agents say few homebuyers read it (on Zillow's homepage, the font for the "About Zestimates" link is slightly smaller than the main home-data type -- and quite a bit fainter).

The sites argue that, over time, edits and corrections will help them perfect their numbers -- and many of the corrections will come from their customers. On Homes.com, for example, anyone who knows certain specifics, like a homeowner's surname and the year the home was last purchased, can edit the details to reflect, say, a sprawling two-bedroom addition. Zillow also allows site visitors to modify its property details, and in four years, it has accepted revisions on 25 million homes -- perhaps the strongest testament to how seriously consumers take the estimates. Today, Zestimates are helpful enough, says the site, to give consumers an accurate sense of any home's value. In the meantime, says Humphries, the company's economist, "We're always tweaking the algorithm or building a new one."

In the end, some critics say, the sites' business models may pose a bigger problem for consumers than their algorithms. Even their flaws help to sustain the buzz around the estimates, drawing curious visitors. The online firms earn significant revenues from advertising, and the more traffic they get, the greater that ad revenue is. Zillow says 57 percent of its revenue comes from display ads from the likes of home-supply store Lowe's, realty franchisor Century 21 and builder KB Home. Realtor.com's parent company, Move Inc., generates 42 percent of its sales from listings by local agents, while Homes.com says advertising is its fastest growing revenue area. Trulia expects its traffic to grow now that it has launched a beta version of an online estimator, says head of communications Ken Shuman; after all, he adds, "consumers asked for it." As long as they keep asking, say industry insiders, stumbles in reliability aren't especially important. "It's not about being accurate or precise; it's about being sticky," says Kennedy, of AVMetrics. For their part, the sites say stickiness matters to their business plans, but that they take the estimates very seriously; otherwise, as a Zillow spokesperson put it, "we wouldn't have a team of Ph.D.s trying to make them better all the time." They depict the estimates as an ongoing experiment that is likely to achieve a very high degree of accuracy -- someday. (At least for now, one site is deferring to agents in the home-value game: Realtor.com says it removes its estimates from homes once they actually go on the market.)

Read it all