Review Text as a Leading Indicator of Displayed Reputation in Platform Rating Systems: Evidence from 34 U.S. Short-Term Rental Markets

Imagine you are walking through a massive, bustling marketplace where everyone sells something, from lemonade to vintage guitars. To help you decide who to trust, every stall has a big, glowing sign showing a star rating. Usually, these stars are the ultimate judge: five stars means "amazing," and four stars means "pretty good." But here's the twist: in this specific marketplace, almost every stall has a rating between 4.5 and 5.0 stars. It's like a classroom where every single student gets an A+. When everyone is perfect, the star rating stops being useful because it can't tell you who is truly the best. This is a problem for the "reputation system," which is just a fancy way of saying "the trust machine" that helps strangers feel safe buying from each other.

So, if the stars are stuck at the ceiling, is there any other way to tell a great stall from a good one? Maybe the written reviews—the long, chatty paragraphs people write about their experience—still hold the secret? This paper asks a clever, time-traveling question: If we read the words people wrote yesterday, can we predict how the star rating will move tomorrow? It's like checking the weather forecast based on how the birds are acting before the clouds even roll in. The researchers wanted to see if the "chatter" in the reviews is actually a leading indicator, a sneak peek at the future score, even when the current score looks frozen and unhelpful.

---

The Story of the "Star Freeze" and the Whispering Text

In the world of online rentals, there's a weird glitch. Almost every listing looks perfect. Whether a place is truly amazing or just "okay," it usually ends up with a rating of 4.8 or 4.9 stars. It's like a video game where everyone hits the high score, so the leaderboard stops telling you who is actually winning. Because the numbers are so crowded at the top, they stop separating the good from the great.

But the researchers, led by Ali Safari, wondered: Is the text underneath those stars just empty noise, or is it actually whispering secrets about the future? They treated the star rating and the written review as two different channels of information. Think of the star rating as a slow, heavy truck that takes a long time to turn. The written text, on the other hand, is a speedy motorcycle that zips ahead. The big question was: Does the motorcycle's path predict where the truck will go next?

To find out, they didn't just guess; they built a massive time machine using data from 34 different U.S. cities. They looked at over 200,000 rental listings and millions of reviews. They took all the text people wrote before a specific date (a "snapshot" in 2025) and turned it into a "text index." This index wasn't magic; it was a carefully built tool that counted how positive or negative people were about specific things like cleanliness, checking in, and communication.

Then, they waited a year. They checked the same listings in 2026 to see how their star ratings had moved.

The Discovery: The Text is a Crystal Ball

The results were clear and surprisingly precise. The researchers found that the "warmth" of the past text did predict the future movement of the star rating.

Here is the magic number: For every standard deviation of "warmer" text (meaning the reviews were slightly more positive than usual), the displayed rating moved up by about 0.00928 of a star over the next year.

Now, don't let the small number fool you. On a scale where most ratings are stuck between 4.5 and 5.0, a movement of about one-hundredth of a star is a real, measurable shift. It's like a slow-moving clock that finally ticks forward. The text was acting as a crystal ball, hinting at where the rating was heading before the rating actually got there.

Ruling Out the Ghosts

But wait, could this be a trick? The researchers were very careful to make sure they weren't seeing things that weren't there. They ran a "falsification battery," which is just a fancy term for a series of stress tests to break their own theory.

First, they looked at a group of listings that got no new reviews between the two years. If the text was just a random guess, it might have predicted a change even here. But it didn't. For listings with no new activity, the text predicted zero movement. This proved that the text wasn't just a lucky guess; it needed new information to work.

Second, they checked if the results were just because some hosts were naturally better than others. They used a special math trick called "host fixed effects" to compare listings owned by the same person. Even when they looked at the same host's different apartments, the text still predicted the rating changes. This meant the effect wasn't just about "good hosts" vs. "bad hosts."

Finally, they split their data in half. They used 17 cities to build their theory (the "discovery" half) and saved the other 17 cities completely untouched (the "confirmation" half). They ran their test on the untouched cities only once, and the result held up perfectly. This is like a scientist predicting an eclipse and then waiting to see if it happens exactly as planned, without peeking at the sky beforehand.

What This Means for You

So, what's the takeaway? The paper suggests that the star rating on a rental app is actually lagging behind. It's the slow truck, while the written reviews are the fast motorcycle. The text contains information that the star rating hasn't absorbed yet.

This doesn't mean the stars are useless, but it does mean they are "stuck" and slow to update. The written reviews are the leading indicator, the early warning system that tells you where the reputation is heading. For the platforms that run these apps, it's a hint that they might be able to design their displays better by listening to the text more closely. For the rest of us, it's a reminder that when you read a review, you aren't just reading about the past; you might be reading the future of that listing's score.

The researchers are honest about the limits, though. They didn't prove that the text causes the rating to change; they just proved that the text predicts it. And their text tool wasn't checked by human judges, so it's a "defined instrument" rather than a perfect human translation. But the pattern is real, precise, and consistent across dozens of cities. The text is speaking, and the stars are just slow to answer.