Why One Poll Percentage Misleads (and What a Crosstab Shows Instead)

You post a question, the votes roll in, and the headline number says 55% agree. It feels like an answer. Then someone in the replies says "that's not what my friends think," and they might be right, because that one number can't tell you whose friends it describes. A topline percentage blends everyone into a single figure. The crosstab is the table underneath that pulls the blend apart, and once you know how to read one, a lot of confusing poll results start to make sense.

What is a poll crosstab?

A crosstab (cross-tabulation) is a table that shows a poll's results broken out by group: gender, age band, region, or any other trait the poll recorded. Instead of one "55% agree" figure, you see how each group voted, how many people are in each group, and how those groups combine to produce the overall number.

Think of the topline as the average and the crosstab as the ingredients. Two polls can both land at 55% while being made of very different ingredients. In one, every group sits close to 55%. In the other, one group is at 80% and another is at 30%, and the mix just happens to average out. Those are two different stories, and only the crosstab tells you which one you're looking at.

Why does a single poll percentage mislead?

A single percentage is a weighted average, so it hides how far apart the groups are and how many people in each group voted. A 50/50 result can mean everyone is torn, or it can mean two groups strongly disagree with each other. The number is the same; what it describes is completely different.

Three things get lost when a result is collapsed into one figure:

• Spread. You can't tell whether opinion is uniform or polarized.
• Composition. If one group made up most of the voters, the topline mostly reflects that group.
• Surprise. The interesting finding is often that a group you'd expect to agree didn't. That never shows up in the headline.

This is why two people can read the same "55% agree" and walk away with opposite conclusions. Each one fills in the missing ingredients with their own assumptions.

How do you read a crosstab, step by step?

Start with the topline, then read each row against it, check the group sizes, and only then look for patterns. The goal is to find where groups differ from the overall number and to judge whether those differences rest on enough votes to mean anything.

A simple routine that works for most crosstabs:

1. Read the topline first. Note the overall split so you have a baseline.
2. Scan each row against the baseline. Which groups sit well above it, and which sit well below?
3. Check the group sizes (often shown as n). A group with very few votes can swing wildly from one vote to the next.
4. Look for the biggest gap. The largest difference between two groups is usually the real story of the poll.
5. Ask what the gap could mean. Different experiences, different readings of the question's wording, or just noise in a small group.
6. Resist picking a winner. A gap tells you that groups see the question differently, not which group is correct.

What does a crosstab look like in practice? (A worked example)

Here's an invented toy example. Imagine a question that asks: "Is it okay to recline your seat on a short flight?" The numbers below are made up purely to show how the reading works; they are not real results.

Topline: 50% yes, 50% no (200 votes).

That looks like a perfectly divided question. Now the crosstab:

Reading it with the routine above:

• The topline says "evenly split." The age rows say something more specific: younger voters in this toy sample lean yes, older voters lean no, and the middle group is the one that's actually torn.
• The region gap is smaller, about 12 points, and Region B has fewer votes, so it deserves less weight than the age pattern.
• None of this says which age group has better manners. It says the question lands differently depending on who you ask. Maybe it comes down to legroom, flight habits, or how people picture "short."

Without the crosstab, you'd conclude the question is a coin flip. With it, you'd conclude it's an age question, which is a much more useful thing to know if you were about to post your take on it.

How many votes does a group need before you trust its number?

There's no magic cutoff, but small groups are much noisier than large ones. When a row has only a handful of votes, a single person changing their mind can move the percentage by ten points or more. Treat small-group numbers as hints, not findings, and look for patterns that hold across several rows.

A few practical rules of thumb:

• Compare sizes before comparing percentages. A 20-point gap between two groups of 300 is far more convincing than the same gap between two groups of 8.
• Prefer consistent patterns. If each age band steps steadily in one direction, that's more believable than one odd row.
• Don't over-slice. Splitting by gender and age and region at once can leave each cell with very few votes. Read one split at a time first.
• Watch for missing groups. If a group barely voted, the topline mostly speaks for everyone else.

What other splits are worth looking at?

The classic splits are gender, age, and region, because those often track real differences in experience. But any trait that changes how someone might see the question is worth a row. One newer split is human vs AI: comparing how people voted with how AI models voted on the same question.

Here's what each split tends to surface:

The human vs AI row works a little differently from the others. Models don't have a hometown or an age, so their votes act more like a second population to compare against. When model votes sit close to one human group and far from another, that's a prompt to ask why, not a verdict on who's right.

What are the most common mistakes when reading crosstabs?

The most common mistakes are trusting tiny groups, treating a gap as proof of who's right, and assuming that the group you belong to voted the way you'd expect. Crosstabs are good at raising questions. They're weaker at settling them on their own.

Other traps to watch for:

• Confusing "most of a group" with "all of a group." 70% of a group still leaves 30% who disagree.
• Reading cause into a gap. A difference between regions doesn't tell you why the regions differ.
• Ignoring the question's wording. A slanted question can produce a slanted crosstab. Check the exact phrasing before reading the rows.
• Cherry-picking one row. It's easy to pull out the single group that agrees with you and ignore the rest of the table.

The fairest way to read a crosstab is to describe every split you see, including the ones that cut against your own view, before deciding what the result means.

Try Repollo

If you'd rather see the ingredients than the average, Repollo is built around that idea. You post a case (a text take, a photo, or a YouTube video), and both real people and AI models vote on it. The result comes back as a crosstab instead of a single percentage, so you can see which groups agreed, which didn't, and where human and AI opinion part ways.

Repollo splits every verdict by gender, age, region, and human vs AI — so you see who actually agreed with you.

• Web: repollo.eodin.app
• iOS: App Store
• Android: Google Play