Market Analysis · September 2026

Why Zillow’s Rent Zestimate Is So Often Wrong

An empty rental living room with hardwood floors, the kind that sits unrented while an owner defends an inaccurate online price estimate — The cost of an over-anchored price is measured in empty rooms. Photo by Max Vakhtbovych on Pexels.

Zillow’s Rent Zestimate is a national statistical model wearing a local disguise. It is genuinely good at describing the average American rental in an average month, and that is exactly why it misses your specific home so often, especially right now. Across the Sunbelt markets we manage in the Carolinas, 2026 has been a year where the number on the model and the number a home actually rents for have drifted further apart than usual. The reasons are worth understanding, because they are structural, not occasional.

Want a number set by people who lease homes here every week? Get a free rental rate estimate →

A median is not a promise

Start with Zillow’s own accuracy disclosure, because the company is fairly honest about this. Zillow publishes a median error rate for the Rent Zestimate, and it reports a wider error band for off-market homes than for homes that are actively listed (see Zillow’s Zestimate methodology page). Median is the word that carries all the weight. A median error in the mid single digits means half of all homes land inside that band and half land outside it, some of them badly. On a $2,000 rental, a model that is “usually within a few percent” is still routinely $120 to $250 off in either direction. And on your individual home that error is not random noise you can average away over a portfolio. It is a single number, and it is the number an owner anchors to and then defends for months.

The estimate is a stack of guesses

A Rent Zestimate is not one prediction. It is a stack of them. Before the model can price your house it has to be roughly right about the national rental trend, then your region, then your metro, then your neighborhood, and only then can it adjust for bedroom count, square footage, and condition. Every layer inherits the error of the layer above it. In a flat year that compounding is tolerable. In a declining market it is not, because the top of the stack is a forward projection. If the macro call is wrong, every local adjustment sitting beneath it is built on a bad foundation.

The Sunbelt is the cleanest example available. Years of heavy apartment construction finally caught up with pandemic-era demand, and Apartment List’s national rent data has tracked year-over-year rent declines across much of the region well into 2026. A model still leaning on the appreciation of 2021 and 2022 is projecting a market that no longer exists. Some years the national call is decent and the Zestimate tracks reasonably. 2026 has not been one of those years; we have almost never seen an accurate top-down prediction hold up. When the top of the stack is wrong, nothing underneath it can rescue the estimate.

The peak the model expects, and the one that never showed

Rental demand is not spread evenly across the calendar. A large share of a market’s annual rent growth arrives in the final eight weeks of summer, in the rush of residents trying to be settled before school starts. We laid out that pattern in detail, with our own leasing data alongside national search data, in The Seasonal Cliff Is Here.

That concentration is a trap for any model trained on history, because the model expects the peak to arrive on schedule. The past two summers, that late peak was a dud. Much of the gain the calendar predicts simply never materialized. An algorithm that has coded in a large, historically reliable August bump keeps pricing as though the bump happened, while the real market quietly skipped it. The estimate is not wrong because the code is sloppy. It is wrong because it trusted a seasonal curve that the last two years refused to follow.

Chart of US median rent from 2017 to present, showing every annual peak lands in August before declining through winter — Every annual rent peak since 2017 lands in late summer, then declines through winter. Source: Apartment List.

Backward-looking in a market that is moving

This is the failure you would think is easy to solve in code, and apparently is not. Algorithms are backward looking by construction. They learn from rents that already closed, which in a declining market are higher than the rents closing today. That would be a manageable lag if rents drifted gently, but they do not. Seasonal swings in rental rates average around 12%, and the market is heavily event driven. Demand gets pulled into a narrow window right before school starts, and once school is in session it drops roughly 10% almost overnight.

A model that sets an early-September asking price from late-July closed rents is not making a small timing error. It is starting about 10% too high on the most sensitive week of the year.

The variable the coders cannot see

There is a human input in this that the people building these models almost never have. They are excellent coders and data scientists, not property managers, so they miss the psychology that makes the math refuse to behave. Landlords are strongly tempted to chase summer rates long after the evidence says summer is over, and that temptation keeps asking prices too high for far too long. If you have watched it play out for twenty years, you price around it. If you have only seen the histogram, you do not know it is there.

We have a saying for this time of year: most homes that are vacant on September 1st are still vacant on December 1st, because owners handle a fast declining market poorly. Instead of an orderly, responsive step down as demand fades, owners tend to hold firm until the holidays start to threaten a long, cold, vacant winter, and only then do they get serious about rate and strategy. That produces a very specific, repeatable shape in real-world prices that a smooth seasonal curve simply does not capture. Because the pre-holiday weeks are actually a solid rental market and January is cold and dreary, the same mispricing shows up year after year, in the same direction, and the models keep getting caught by it.

Spring runs the other way

The same lag works in reverse every spring, and it is just as revealing. When the first major wave of renters hits the market, we routinely set prices 15% to 20% above Zillow’s Rent Zestimate, and we get them. The model is still averaging in the slow winter it just climbed out of while live demand has already turned upward.

The amusing part is that spring mispricing corrects itself far faster than the declining-market version. When owners discover they are underpriced, they raise quickly, because the upside is obvious and pleasant to act on. When they are overpriced in a falling market, they argue with the evidence for months. Same model, same seasonality, wildly different speed of correction, and the entire difference is human.

Zillow already showed us the model can fail

If you want proof that these models struggle exactly where it matters, Zillow handed it to us. In November 2021 the company shut down Zillow Offers, its home buying and selling arm, wrote down hundreds of millions of dollars in inventory, and cut about a quarter of its workforce (CNBC). CEO Rich Barton’s own explanation was that the unpredictability of forecasting home prices had far exceeded what they anticipated. That was Zillow, holding the best housing data set in the country, betting real money on its own price forecasts and losing badly enough to close the business.

So when a Rent Zestimate insists your home is worth a number the live market keeps declining to pay, remember what the same class of model cost the company that built it. No owner should accept two or three months of vacancy defending a robot’s price when the machine’s own creators walked away from that exact bet. Vacancy on a $2,000 home runs roughly $67 a day; a stubborn estimate is an expensive thing to be loyal to.