How the drop was made
The Airbnb number is our own count of this year’s public listings. The grids come from published research, with a source for each. The lab is a simulation we wrote, and every rule and number in it is here. Open any part below.
The dataWhere the Airbnb and food numbers and the six grids come from
The Airbnb numbers
Inside Airbnb publishes snapshots of public Airbnb listings under a Creative Commons Attribution 4.0 license. We used the detailed listings files for San Francisco and New York City, both compiled on 14 June 2026, and kept listings with a guest rating and at least 10 reviews: 3,592 in San Francisco and 12,102 in New York. The rating is the overall score Airbnb shows on each listing. Their median is 4.84 and their mean 4.79, which gives the title its number. On the page, “typical” means that median: half of the listings are rated 4.84 or higher. Of these 15,694 listings, 92.3 percent sit at 4.5 or above, 60.6 percent at 4.8 or above, and 0.3 percent below 4.0, which the page rounds to fewer than 1 in 100. In San Francisco, anything under 4.7 lands in the bottom 15 percent.
Counting every rated listing (27,448), the median is 4.88 and 62.8 percent sit at 4.8 or above. We use the 10-review cut because a single five-star review can give a new listing a perfect score. Inside Airbnb no longer serves its older snapshots, so we make no claim about how these ratings changed over time. The script that computes every Airbnb number here is research/airbnb_measure.py.
The drift grids
| System | Top of the scale | Share | Source |
|---|---|---|---|
| US college grades | Letter grades that are A’s | 15% in 1960, 43% by the late 2000s | Rojstaczer and Healy (2012) |
| Harvard College | Grades that are A’s | 24% in 2005, 60.2% in 2025 | Harvard Office of Undergraduate Education (2025) |
| An online labor market | Workers rated 5 stars | 33%, then 85% six years later | Filippas, Horton and Golden (2022) |
| eBay sellers, 2011 to 2014 | Feedback that is positive | 99.3% for the average seller; median 100% | Nosko and Tadelis (2015) |
| Airbnb, 2015 | Listings rated 4.5 or higher | 94% | Zervas, Proserpio and Byers |
| TripAdvisor hotels, 2015 | Hotels rated 4.5 or higher | 26% | Same |
| Airbnb, San Francisco and New York, 2026 | Listings with 10+ reviews rated 4.5 or higher | 92.3% | Our count of Inside Airbnb’s listings |
| Packaged food, 2025 | Products with 10+ reviews averaging 4.5 or higher | 33.1% | Our count of 2025 store reviews |
| Beauty products, 2025 | Products with 10+ reviews averaging 4.5 or higher | 27.4% | Same |
| LinkedIn endorsements | There is no negative endorsement | 100%; a billion given in under six months | LinkedIn’s 2012 launch post |
| GitHub stars | There is no negative star | 100%; 6.0 million suspected fakes, July 2019 to December 2024 | He et al. (ICSE 2026) |
The page draws six of these in two rows. Rating people: Harvard, eBay and Airbnb in 2026. Rating things: TripAdvisor hotels, packaged food and beauty products. Hotels are from 2015, the most recent like-for-like count we found; the same study put Airbnb at 94 percent then. The rest are here for the record. LinkedIn reported in March 2013 that a billion endorsements had been given, less than six months after launch. Grids round to the nearest whole star. The labor-market paper reports the rise in five-star ratings over six years without naming the years; its average rating runs from 3.74 in early 2007 to 4.85 in May 2016. On GitHub, researchers also found that 16.66 percent of repositories that gained at least 50 stars in July 2024 had run a fake-star campaign. Zervas and colleagues note TripAdvisor does not let hotels rate guests back and suggest people rate “other individuals differently or more tactfully, than they rate firms.”
The lab’s rulesHow the 80 simulated coworkers rate each other, and what each setting finds
How the lab works
The lab is a simulation, written for this page, of how scores evolve under simple rules. It does not use real people or real data. Its code is in js/lab-model.js.
- People. 80 coworkers in ten teams of eight. Each has a hidden skill drawn from a bell curve and a visibility (how much of a presence they have) that is only loosely related to skill, a correlation of 0.2. Each person also picks two people at random on other teams as people they have worked with, rarely the same one twice, and those links go both ways.
- Rounds. Each of 30 rounds, every person gives three one-tap vouches, or one written vouch, or one vouch tied to a piece of work. Each vouch is as likely to go to someone they worked with as to whoever the feed shows them. No one vouches for the same person twice in a round, and everyone gives a round’s vouches before anyone sees them.
- Judgment. Coworkers judge your skill with some noise. People who never worked with you judge you 35 percent on skill and 65 percent on visibility, with more noise. Writing a vouch cuts either kind of noise by a fifth. A vouch tied to work is judged from the work, which shows skill with less noise, even to strangers.
- Stars. Raters judge people partly against the current average rating and add a little on top, because a below-average rating feels like a penalty; with no fixed standard, that extra compounds every round. Against the fixed standard, an average person earns 3.2 stars, plus one star for each standard deviation of skill the rater sees. A one-tap rating is judged entirely against the average, a written vouch 30 percent against the fixed standard and a vouch tied to work 75 percent, the slider under “A vouch is.” Ratings round to whole stars from 1 to 5. Twelve percent of one-tap ratings and four percent of written ones are reflexive fives. A score is the average of the last 12 vouches a person received.
- A named yes. There is no way to say no, only silence. A giver says yes when someone seems at least average (one tap) or at least half a standard deviation above it (written or tied to work), and a quarter of one-tap judgments, and 3 percent of written ones, are a yes given without judging. A person’s standing is the number of yeses they have received, weighted as the rules on how they know you and on the feed describe.
- Visible or blind. A visible star rating is pulled 40 percent toward what the other person last gave the rater, and shades up more (0.12 stars) than a blind one (0.05). A visible yes makes a yes back easier: when the other person last said yes to the giver, the giver’s bar drops by half a standard deviation.
- How they know you. When a vouch says how the giver knows you, a vouch from someone who never worked with you counts a quarter as much.
- Feed. A feed ranked by count shows people more often the more vouches they have. A feed ranked by who vouched weights each vouch by the giver’s track record, meaning how the people they vouched for (with a yes, or four or five stars) have turned out, which the model reveals a little each round with noise; that weighting sets the scores too, and a record built on few vouches counts for less. The track record is our stand-in for how a trusted reference would be weighed. In either ranked feed, a person’s place in it sways how others judge them, by the same amount for every kind of vouch. A plain feed ranks no one.
- Measures. “At 4.8 or above” counts people whose score is at least 4.8. “Top 10 by score” counts how many of the 10 highest scores or standings belong to the 10 most skilled people; chance alone gives about one. Ties are broken at random, once per run, and people with no vouches rank last. The lab shows one simulated company, world 374, picked from 600. Its first run keeps every result the text describes (the worst setting finds about half, tying ratings to work finds the most of any single switch, and on the yes scale a track record beats a raw count), and each of the seven sets of rules in scene 5 finds within one of its average there. On its first run the worst setting finds 5 of the 10 most skilled and tying ratings to work finds 7; the table below gives the averages, which are what scene 5’s bars show.
- The chart. On the stars scale the dots sit by score, from 1 to 5. On the yes scale they sit by standing, on an axis that ends at a round number at or above the 95th percentile of standings, so the few people far ahead wait at its end, marked “+”, instead of squeezing everyone else toward 0. Weighted yeses compound, so their axis is spaced by square root: its evenly spaced ticks read 0, 1, 4, 9 and 16 sixteenths of the end. People with no vouches, or no yes, sit in their own lane at the left.
| Setting | Of the 10 most skilled, how many the top 10 finds | Average score, round 1 to 30, or share of yeses | People with no vouch or no yes |
|---|---|---|---|
| Worst: stars, one tap, seen right away, not said, feed by count | 4.50 | 3.5 to 4.4 | 0 |
| Tied to work, the rest as worst | 6.08 | 3.2 to 3.3 | 0 |
| Written, the rest as worst | 4.75 | 3.3 to 3.7 | 0 |
| After both write, the rest as worst | 4.83 | 3.4 to 4.2 | 0 |
| Says how they know you, the rest as worst | 4.75 | 3.5 to 4.5 | 0 |
| Feed ranks nothing, the rest as worst | 5.25 | 3.5 to 4.4 | 0 |
| Feed ranks by who vouched, the rest as worst | 4.67 | 3.5 to 4.4 | 0 |
| Tied to work, after both write, said, feed ranks nothing | 8.42 | 3.2 to 3.2 | 0 |
| Real work + trusted references: tied to work, after both write, said, feed by who vouched | 8.33 | 3.2 to 3.3 | 0 |
| LinkedIn-style endorsements: named yes, one tap, right away, not said, feed by count | 2.42 | 78% yes | 0 |
| Named yes: written, right away, said, feed by who vouched | 5.92 | 64% yes | 20.2 |
| Named yes, feed by count | 3.75 | 64% yes | 21.4 |
| Named yes, feed ranks nothing | 6.17 | 37% yes | 5.1 |
| Named yes, tied to work | 6.67 | 69% yes | 32.4 |
| Named yes, tied to work, feed ranks nothing | 6.25 | 37% yes | 13.8 |
| Picking 10 people at random | 1.25 |
Tying ratings to work is our most generous setting. We assume a vouch that points at work has no reflexive fives or yeses, reads skill with less noise, is judged on the work even by strangers, and is held to a fixed standard. That standard is what keeps the scores from drifting: with the other switches at their worst, setting it to 0 lets the average climb 0.75 over the run, against 0.96 for one-tap stars. If strangers judged a piece of work entirely by how well known its maker is, tying ratings to work would find 4.25, below the worst setting’s 4.50. On the yes scale, tied-to-work yeses leave more people with none, because none of them is reflexive. Having no yes there can mean nobody looked, so it is not a judgment about the person.
The rules follow the research under “Why this happens” and “The research,” but every number in them is our assumption, including the ones that make tying vouches to work look strong. Change them under “Try other rules” and the sizes of the effects change with them. What held across the settings we tried is the direction: scores judged only against each other drift up, scores anchored to the work do not, and on the yes scale, weighting who said yes beats counting yeses. Push the assumptions to their extremes and some effects reverse.
One question stays open: how much does pointing at real work help real vouches? Our lab says a lot, but only because of rules we chose. Real vouches and real outcomes would settle it.
The researchWhy ratings drift, what predicts job performance, and what keeps a vouch honest
Why this happens
No one has to lie. Four small, ordinary habits all push the same way.
-
1. Giving is free
A one-click endorsement costs nothing, so little stops anyone from giving one. Scarce signals work better. At two dating events, people got a few free “roses” to attach to date requests. A request with a rose was accepted about a fifth more often.
-
2. The other person sees what you said
If they can rate you back, a low score invites a low score in return. When Airbnb hid each review until both sides had written one, the link between guest and host ratings fell by about half.
-
3. Honesty costs you
A bad rating can cost someone income, so people hold back. On one big freelance site, 28.4 percent of employers who privately said they would never rehire a worker still gave that worker four stars or more in public.
-
4. Unhappy people go quiet
Happy customers leave feedback. Mildly unhappy ones mostly say nothing. On eBay, fewer than 1 in 100 comments was negative. One estimate says about 1 in 5 buyers was unhappy.
Then the feed makes it worse. On one news site, a single random up-vote made the next reader 32 percent more likely to up-vote too. Praise keeps flowing to whoever already has it.
Hotels are the exception. On TripAdvisor, only about 1 in 4 hotels is rated 4.5 or higher. The researchers suggest people rate a business more harshly than a person.
When writing got cheap, reputation got expensive
ChatGPT arrived in November 2022, and soon anyone could write a polished job proposal. Economists compared job posts on Freelancer.com from before ChatGPT with posts from 2024.
What an employer would pay for one step more of each
| For one standard deviation more of | Before ChatGPT | 2024 |
|---|---|---|
| A proposal tailored to the job | $25.67 | $14.85 |
| A higher reputation score on the site | $27.96 | $46.24 |
Employers would pay 42 percent less for a well-tailored proposal. What they would pay for a higher reputation score rose by about two thirds. We read that as employers leaning on a worker’s record once the writing tells them less.
Both sides use AI more each year. In a 2025 survey, 29.3 percent of job seekers had used it on a resume or cover letter. In SHRM’s 2025 survey, 43 percent of organizations used AI for HR tasks.
Watching someone work beats counting their years
Work samples, where a candidate does a piece of the real job, predict performance better than years of experience. A 2022 re-analysis lowered their score, but years of experience ended lower still.
How well each method predicts job performance
| Method | 1998 | 2022 |
|---|---|---|
| Structured interviews | .51 | .42 |
| Job knowledge tests | .48 | .40 |
| Biodata questionnaires | .35 | .38 |
| Work samples | .54 | .33 |
| Cognitive ability tests | .51 | .31 |
| Integrity tests | .41 | .31 |
| Unstructured interviews | .38 | .19 |
| Conscientiousness | .31 | .19 |
| Years of experience | .18 | .07 |
| Peer ratings | .49 | not re-estimated |
| Reference checks | .26 | not re-estimated |
| Years of education | .10 | not re-estimated |
What keeps a vouch honest
The research points to four rules.
-
A vouch stays hidden until both people have written theirs
Dana, about Omar Written. Hidden until Omar writes his.Dana, about OmarOmar rebuilt our billing retries in three weeks. Failed charges dropped, and I stopped getting paged on weekends.
Omar, about Dana Written. Hidden until Dana writes hers.Omar, about DanaDana ran the launch checklist and caught the tax bug two days before we shipped. I’d want her on the next launch.
Neither card can be written to match the other, so there’s nothing to trade. On Airbnb, blind reviews cut the link between the two sides’ ratings by about half.
-
It names something specific the person did
A specific claim ties praise to the work itself, and a piece of real work can be checked.
-
It says how you know them
Expedia only lets guests who booked a stay review a hotel. TripAdvisor lets anyone. Comparing the same hotels, economists found more signs of fake reviews on TripAdvisor, most of all where a hotel had a reason to fake them. LinkedIn began highlighting endorsements from colleagues in 2016.
-
It costs the giver something
Scarce signals keep their meaning, as the roses suggest. The simplest cost is your name.
One more finding applies to networks that rank people. A feed that ranks by raw count hands the gains to whoever is already ahead. In the lab, ranking by the track record of who said yes finds more of the best people than ranking by count.
SourcesEvery source behind the numbers, and how this was made
Why this happens, the sources
- Roses: Lee and Niederle (2015), two dating events with 613 participants, each given two or eight roses that were free to send but limited in number. “About a fifth more often” is their estimate of 3.3 percentage points, a 20 percent increase, holding the sender’s and recipient’s desirability fixed. Job-market signals, which are scarce the same way: Coles et al. (2010), Table 3, where a candidate’s signal of interest came with a 6.8 point higher chance of an interview, an association in their data.
- eBay retaliation: Bolton, Greiner and Ockenfels (2013), on sellers answering negative feedback in kind, and a laboratory version of eBay’s system in which blind feedback cut the correlation between the two sides’ ratings from 0.68 to 0.41 (Table 4). Airbnb’s simultaneous reveal: Fradkin, Grewal and Holtz (2021), the introduction and Section 7: the correlation between guest and host ratings fell by 48 percent, the “about half” on the page, while the average guest rating moved only 0.25 percent.
- Public and private feedback: Filippas, Horton and Golden (2022), read in the NBER working paper version; the platform is not named in the paper. In its surveys, “some employers report they fear retaliation, while others claim to not want to harm the rated individual.”
- Restaurants and people: Yelp’s 2023 Trust & Safety Report gives the star ratings users submitted in 2023, from US data: 19 percent were one star, the “1 in 5” on the page, and the average was about 3.8. Yelp is an interested party, so we use only its own published counts. The freelance figure on the page comes from Filippas, Horton and Golden, read in the NBER working paper version: 28.4 percent of employers who privately said they would definitely not hire the same worker again publicly gave 4 or more stars out of 5, and more than 80 percent of public ratings fell between 4.75 and 5.
- Food and beauty on store shelves: our own count of public customer reviews posted on Amazon, Walmart, Target and Kroger product pages in 2025, taken from a commercial collection of retailer reviews, so we can’t republish the raw data. Across all products that came to 28.5 million reviews, 8.9 percent of them one star. A product counts as food if the collection tags it food, even when it carries another tag too, and the same for beauty. Food drew 8.1 million reviews, 10.9 percent of them one star. Of the 61,889 food products with at least 10 reviews that year, the median product averaged 4.29, 33.1 percent averaged 4.5 or higher, 7.4 percent averaged 4.8 or higher, the “7 in 100” on the page, and 30.2 percent averaged under 4.0. Of 34,550 beauty products, 27.4 percent averaged 4.5 or higher. Each share compares a product’s star total with the threshold times its review count, so no rounding moves a product across a line. The gold line on screen 2 is the food products’ histogram in the same 0.05-wide bands as the Airbnb bars, each band as a share of all food products, on the same scale. Two differences to keep in mind: the product counts use only reviews posted in 2025, while an Airbnb score covers a listing’s whole history, and products are judged by shoppers who chose to write a review, as listings are. The counts are in data/food.json.
- Why Airbnb ratings skew high: Airbnb’s support account, Airbnb Help, in a post on X on 5 October 2026: “Ratings on Airbnb skew high partly because we remove listings that fall short of guest expectations.” The same post says Airbnb removes reviews that violate its policies, that hosts can’t delete reviews, and that only guests who booked can leave one. Some users replied that hosts get bad reviews taken down; we found no reporting that confirms it, so the page doesn’t say it.
- Silent buyers: Dellarocas and Wood (2008), Tables 2, 5 and 6: 0.55 percent of comments were negative, the “fewer than 1 in 100” on the page, and two versions of their model put dissatisfied buyers at 18.5 and 21 percent, the “about 1 in 5.” Counting every sale, including the ones with no feedback, the average seller’s positive share falls from 99.3 percent to 64 percent: Nosko and Tadelis (2015).
- Feeds: Muchnik, Aral and Taylor (2013), where one random up-vote made the next viewer 32 percent more likely to up-vote and those comments ended 25 percent higher; Salganik, Dodds and Watts (2006), where showing download counts to 14,341 people made the hits bigger and harder to predict; and Merton (1968) on the Matthew effect.
- Verified reviews: Mayzlin, Dover and Chevalier (2014). Colleague endorsements: LinkedIn’s 2016 update.
Why now
Galdin and Silbert (2025) use Freelancer.com job posts from January 2021 to July 2024. Their Table 2 compares posts from before ChatGPT’s release on November 30, 2022 with posts from March 26, 2024 on, by which time the site had its own AI writing tool (from April 2023). It turns how much employers favor each part of an application into dollars: how much higher a bid an employer would accept for one step up, one standard deviation, in that part. We quote the columns limited to applications the employer looked at. The reputation score is the site’s own ranking of workers, built mostly from their record there; employers see the ranking and many of its inputs, but not the score itself. Employers would pay $14.85 more for a proposal one standard deviation better tailored after ChatGPT, against $25.67 before, which is 42 percent less, and $46.24 more for a higher reputation score, against $27.96, which is about two thirds more. The job seeker figures come from iHire’s State of Online Recruiting 2025, a survey by a job site of 1,421 job seekers, where 29.3 percent had used AI on a resume or cover letter, up from 17.3 percent a year earlier, and the HR figures from SHRM’s 2025 Talent Trends, a February 2025 survey of 2,040 HR professionals about their organizations: 43 percent of organizations used AI for HR tasks, up from 26 percent in 2024.
Job performance
The chart’s numbers are .54 to .33 for work samples and .18 to .07 for years of experience, among others. They are correlations between each method’s score and later job performance, after statistical corrections. Schmidt and Hunter (1998) corrected for imperfect measures of performance and for the fact that studies only see people who were hired, a narrower group than everyone who applied. Sackett and colleagues (2022) list the new values in their Table 3 and explain each larger change in their text. That second correction had often been made too large, which is the main reason cognitive ability tests, both kinds of interview and job knowledge tests fell. Work samples, integrity tests, conscientiousness and years of experience fell because newer meta-analyses, which pool the results of many earlier studies, found lower values. The 1998 work-sample figure came from a 1974 review written before such pooling was done, and the new experience figure looks at experience before hiring instead of time in the current job. Five of the 1998 values, including work samples, reference checks and peer ratings, never got that second correction, and eight could not be re-estimated in 2022, including reference checks, peer ratings (coworkers rating each other, at .49) and years of education. In a 2023 follow-up the same authors lowered cognitive ability again, to .23, using an unpublished analysis, and their method is disputed: see Oh, Le and Roth (2023) and the reply that followed. One more caution: in the 2022 table, Black candidates’ average work-sample scores sit about two thirds of a standard deviation below white candidates’, one of the larger gaps among the methods listed, so how work samples are designed and scored matters.
How this was made
We built this with Claude Code, Anthropic’s coding agent, over two days in September 2026. Agents traced each number to its primary source. The page is plain HTML, CSS and JavaScript, with no framework, cookies or tracking. Every source we checked, including the ones we cut, is listed in the sources ledger.
Credits
Airbnb data from Inside Airbnb, used under CC BY 4.0. Type is Newsreader by Production Type and IBM Plex Mono by IBM, both under the SIL Open Font License.