Why 90% Preferring One Option Doesn’t Mean It’s Much Better

Polls are everywhere: political races, product reviews, taste tests, headline-grabbing brand battles. And the message is usually loud and clear: if a large majority prefers one option over another, it must be far superior. But that “clear message” is very often actually an illusion.

According to a new study by Graham Overton (National University of Singapore), Ioannis Evangelidis (ESADE Business School, Barcelona), and Joachim Vosgerau (Bocconi University), we systematically misread consensus (such as 90% prefer A over B) as magnitude (A must be much better than B. In other words, we assume that the more people choose A over B, the more they must like it—when in reality, a high percentage in a poll may mask very small actual differences in liking.

“Consumers believe that 90% consensus implies A being much better than B.”

This idea—that large majorities imply large differences—can be true, but it is much more likely that it is wrong. And it has major consequences for how we judge products, people, and policies.

The surprising math behind majority opinion

Suppose you rate two movies on a scale from 1 to 10. There are many ways to end up with a small difference of 1—for example giving one movie an 8 and the other a 7, or giving one a 5 and the other a 4. There are far fewer ways to create a huge difference, for example a difference of 9 is possible only with rating one movie a 10 and the other a 1. Consequently, small differences in liking are much more likely than large differences in liking. 

That holds even when 90% of people say they prefer Movie A over Movie B, it is still much more likely that Movie A is just slightly more liked. The large majority doesn’t require a large gap in liking—it just means more people preferred it.

“At any given consensus level… it is more probable that both options are liked to a similar extent than it is that one option is liked vastly more than the other.”

This is not a quirky footnote—it’s a robust statistical regularity that holds across different rating scales, sample sizes, and even when preferences are correlated in complex ways.

The fieldwork

The researchers put their hypothesis through rigorous testing, using a blend of simulations, large-scale datasets, and controlled experiments. They started by modeling how preferences and differences in liking would distribute if people rated two options (like two beers or jokes) on a scale—say, from 1 to 10. 

From this, they created simulations showing that even with extreme consensus levels (like 90% choosing Option A), the most likely difference in average ratings between A and B is still small—often around 1 to 2 points out of 10. 

To test their simulations empirically, the authors turned to big datasets:

• Jokes: Over 600,000 ratings from nearly 25,000 people.
• Beers: More than 1.5 million ratings across 56,000 different beers.
• Movies: 25 million ratings from over 160,000 people.

They focused on cases where people rated both items in a pair, so they could measure actual differences in liking and compare them to the consensus preference (i.e., how many people picked A over B). Across all domains, they found the same thing: small differences in ratings were far more common than big ones, even when preference shares strongly favored one option.

Finally, to see if people’s intuitions aligned with reality (they didn’t), the researchers ran seven experiments with hundreds of participants.

In one, students in a master’s data analytics course were told that 90% of tasters preferred Wine A to Wine B. Then they were asked how much better they thought Wine A was rated on average. Nearly all overestimated the gap, choosing ranges like “3–4 points higher” or “4–5 points higher.” The actual difference? Around 1.

In another experiment, participants were shown real beer pairs with known consensus levels. Using an interactive tool, they were asked to “build” the distribution of how people likely rated the beers. Most created distributions that were skewed too far to the right—with exaggerated peaks and gaps—suggesting they imagined bigger differences in liking than actually existed.

“Participants overestimated the mean, mode, and probability of the maximum difference… and underestimated the probability of the smallest difference.”

Even when given incentives for accuracy, participants overestimated differences in liking.

Consensus information is more persuasive

Marketers, campaign managers, and influencers know this instinctive bias well. That’s why you often see preference statis—“9 out of 10 dentists recommend…” or “84% of people chose…”—without any information about how much better those options are rated, because consensus messaging is more persuasive than ratings, even when it’s misleading.

When all you know is that a large majority prefers Option A, your brain fills in the blanks—and it tends to inflate the difference in quality, appeal, or truth.

However, the researchers found that when participants were educated about how preference distributions work—specifically, that small differences are statistically more common—they made more accurate judgments. Indeed, when they were shown an interactive tool where they could guess how ratings might be distributed given a certain consensus level, people saw the skewed nature of the actual data and adjusted their assumptions significantly.

Bigger implications

This bias isn't limited to the marketing of products, but occurs in politics, science, moral judgments, in short wherever we may encounter consensus information. When we hear that “90% of people support X,” we assume they support it strongly. But in reality, many of those 90% may prefer X only to some degree—and would switch if offered something slightly better. The bias might inflate perceived polarization, lead to overestimating others’ preferences, or to underestimate the support that a minority candidate or option is receiving.

Graham Overton, Ioannis Evangelidis, Joachim Vosgerau, “People Believe If 90% Prefer A over B, A Must Be Much Better than B. Are They Wrong?”, Journal of Consumer Research, Volume 52, Issue 1, June 2025, Pages 135–156, DOI https://doi.org/10.1093/jcr/ucae055