Original research

We simulated 660,000 group votes. Single-choice picked a different winner 43% of the time.

Everyone knows vote splitting exists. Nobody seems to have measured how often it changes the answer in an ordinary group decision — a team lunch, a feature list, a meeting time. So we simulated it.

31 July 2026 · 660,000 simulated votes · methodology and code below
43.2%
of eight-option votes produced a different winner under single-choice than under ranked choice
84.3%
of five-option votes had a single-choice winner with no outright majority
66.7%
the share a preference bloc needs to survive splitting between two similar options
28%
disagreement rate, roughly constant whether the group is 6 people or 100
The short version

With three options, single-choice voting is mostly fine — the two methods disagreed 12.7% of the time. From four options upward it degrades quickly, and by five the winner almost never holds a real majority.

The sharpest finding is arithmetic rather than statistical: if a group who prefer similar things splits between two of them, they lose under single-choice voting unless they hold more than two-thirds of the vote. Below that threshold, the win goes to an option most of the group did not want.

Finding 1 — disagreement grows with the number of options

We ran 20,000 simulated votes at each option count with 25 voters and independently placed options, then compared the plurality winner against the instant-runoff winner.

How often the two methods picked different winners

25 voters · 20,000 trials per option count
312.7%
421.8%
528.7%
635.2%
740.0%
843.2%

How often the single-choice winner had no outright majority

25 voters · 20,000 trials per option count
337.6%
467.8%
584.3%
693.1%
796.6%
898.3%

The second chart is the more troubling one. At five options, the leading choice held less than half the first preferences in 84.3% of votes. At eight options, 98.3%. In other words, when you put eight things to a single-choice vote, the winner almost certainly represents a minority of the room.

OptionsDifferent winnerNo majorityIRV winner beats plurality head-to-headAvg winner shareAvg rounds
312.7%37.6%12.7%55.6%1.38
421.8%67.8%21.5%46.5%2.25
528.7%84.3%27.9%40.4%3.34
635.2%93.1%33.8%35.8%4.51
740.0%96.6%38.2%32.5%5.63
843.2%98.3%40.8%29.9%6.71

The fourth column is the strongest test. It is not merely that the methods disagreed — it is that in those cases, the ranked-choice winner would have beaten the single-choice winner in a direct head-to-head vote. At eight options that happened in 40.8% of all votes.

Finding 2 — the two-thirds rule

The classic vote-splitting worry is that two similar options divide their supporters and hand victory to a third. We isolated that scenario: a bloc of voters who prefer either of two similar options, against one distinct alternative.

Bloc shareSingle-choice picks the minority optionRanked choice picks the bloc's option
52%100.0%100.0%
55%100.0%100.0%
60%100.0%100.0%
65%40.9%100.0%
70%0.0%100.0%

At 52%, 55%, and 60% the single-choice vote handed the win to the minority-preferred option every time. At 65% it happened in 41% of runs. At 70% it never happened.

The threshold is exact and can be derived without simulation. A bloc holding share s, splitting evenly between two options, gives each s/2. The distinct alternative holds 1−s. The bloc loses when s/2 < 1−s, which resolves to s < 2/3.

Practical meaning

If your team has two similar options on the list — two Italian restaurants, two variants of the same feature, two adjacent time slots — and the people who prefer that general direction are fewer than two-thirds of the group, a single-choice vote will systematically pick something else.

Finding 3 — small teams are not safer

We expected small groups to be less affected, since fewer voters means less fragmentation. They were not.

VotersDifferent winnerNo majority
628.4%78.4%
1028.0%82.4%
1527.8%78.4%
2528.8%84.7%
5029.5%90.4%
10030.6%91.6%

Across groups from 6 to 100 with five options, disagreement stayed between 27.8% and 30.6%. A six-person team choosing between five options faces essentially the same risk as a hundred-person department. Group size is not the variable that matters — the number of options is.

Finding 4 — which method finds the head-to-head winner

A Condorcet winner is an option that beats every other option in a direct one-against-one comparison. Where one exists, it is a reasonable definition of the "right" answer. We measured how often each method found it.

OptionsCondorcet winner existsSingle-choice found itRanked choice found it
399.3%85.1%96.3%
498.4%73.4%90.6%
597.3%64.1%84.5%
696.4%55.4%77.9%
795.1%49.2%72.8%
894.3%44.3%67.7%

Ranked choice was closer in every condition. With five options it found the head-to-head winner 84.5% of the time against single-choice's 64.1%. With eight, the gap widened to 67.7% against 44.3% — meaning a single-choice vote with eight options found the head-to-head winner less than half the time.

Neither method is perfect, and no method can be. Arrow's impossibility theorem rules out a ranking system satisfying every fairness criterion simultaneously. The question is not which is flawless but which is closer, and in these conditions ranked choice was closer throughout.

Methodology

Preference model

Voters and options were placed in a two-dimensional latent preference space drawn from a standard normal distribution. Each voter ranked options by Euclidean distance from their own position, with Gaussian noise of σ = 0.10 added per voter-option pair to represent individual idiosyncrasy.

This is the spatial model standard in social choice research. It produces correlated preferences — voters who favour one option tend to favour nearby ones — which uniform random rankings do not. Using random rankings would have inflated the disagreement figures considerably.

Voting rules

Plurality: the option with the most first preferences. Ties broken at random.

Instant-runoff: count first preferences among continuing options; if one exceeds 50% of continuing ballots it wins; otherwise eliminate the lowest and redistribute to each ballot's next continuing preference. Repeat. Ties for elimination broken at random.

Condorcet: the option beating every other in pairwise comparison, computed from a full preference matrix. Reported as absent when no such option exists.

Verification

Before generating any results, the instant-runoff implementation was tested against a worked example with a known outcome — a four-option vote with first preferences of 35 / 28 / 22 / 15 — and reproduced the published elimination sequence and winner exactly.

Scale

18 primary conditions × 20,000 trials = 360,000 simulated votes, plus 15 vote-splitting conditions × 20,000 = 300,000. Total 660,000. Seeded at 20260731 for reproducibility.

Limitations

  • These are simulated votes, not observed ones. The findings describe what the model implies, not what happened in any real decision.
  • The spatial model is an assumption. Real group preferences may cluster differently, and a different model would produce different numbers.
  • Sincere voting only. No strategic behaviour was modelled, though strategic voting is itself largely a response to the problem described here.
  • Random tie-breaking affects small groups disproportionately, which is visible in the six-voter row of Finding 3.
  • We publish a ranked-choice poll tool, so we have an interest in the conclusion. The methodology is stated in full so the arithmetic can be checked independently.

What this means for an actual decision

Two or three genuinely distinct options: single-choice is fine. Disagreement was 12.7% at three options and the extra effort of ranking is not worth it.

Four or more options: disagreement roughly doubles at four and the majority rate collapses. This is where ranking starts to earn its cost.

Any two options that are similar to each other: the two-thirds rule applies regardless of how many options there are in total. If the people who want that general direction are fewer than two-thirds of the group, single-choice will systematically choose against them.

Any decision where the outcome needs legitimacy: a winner holding 30% of first preferences is difficult to defend as the group's choice. If people have to live with the result, the margin matters as much as the answer.

Frequently Asked Questions

What did the simulation actually measure?
For each simulated vote we computed three things: the plurality winner (most first preferences), the instant-runoff winner (eliminate the lowest option and redistribute until one holds a majority), and the Condorcet winner (the option that beats every other in a direct head-to-head), where one exists. We then measured how often these disagreed across 660,000 simulated votes.
How were voter preferences generated?
Using a spatial model, which is standard in social choice research. Voters and options are placed in a two-dimensional preference space and each voter ranks options by distance from their own position, with a small amount of individual noise. This produces correlated preferences — people who like one option tend to like similar ones — which is realistic. Uniform random rankings would have overstated disagreement substantially.
Is this real data or simulated?
Simulated, and we want to be explicit about that. These are not observations of real votes. They are the mathematical consequences of a preference model, run at scale. The value is that the model is stated, the code is reproducible, and anyone can change the assumptions and re-run it. It is not a claim about what happened in any particular real vote.
Why does the plurality winner so rarely have a majority?
Arithmetic. With five options and reasonably spread preferences, the leading option typically takes around 40% of first preferences. For it to hold a majority, it would need to beat the combined total of four alternatives. Our simulation found the plurality winner lacked an outright majority in 84.3% of five-option votes and 98.3% of eight-option votes.
What is the two-thirds rule?
If a group of voters who all prefer similar options splits between two of them, they lose under single-choice voting unless they hold more than about two-thirds of the vote. The arithmetic is simple: a bloc of share s splitting evenly gives each option s/2, while a distinct alternative holds 1−s. The bloc loses when s/2 < 1−s, which resolves to s < 2/3. Our simulation confirms this, with the transition softened when the split is uneven.
Does group size change the result?
Much less than we expected. Disagreement between the two rules held between 27.8% and 30.6% across groups from 6 to 100 voters with five options. A six-person team faces essentially the same risk of a split-vote outcome as a hundred-person department. Small size does not protect you.
Which method agrees more with the Condorcet winner?
Instant-runoff, consistently. With five options, the plurality winner matched the Condorcet winner in 64.1% of cases and the instant-runoff winner in 84.5%. With eight options the gap widened to 44.3% against 67.7%. Neither method is perfect — no voting system can be, by Arrow's theorem — but instant-runoff was closer to the head-to-head winner in every condition tested.
How many options is too many?
The data suggests the trouble starts at four. With three options, single-choice and ranked-choice disagreed 12.7% of the time and the winner held a majority in 62% of votes. At four options, disagreement nearly doubles to 21.8% and the majority rate collapses to 32%. Beyond five, an outright majority under single-choice becomes rare.
Does this mean single-choice voting is broken?
No, and we would not put it that way. With two or three genuinely distinct options it works fine and is simpler for everyone. The problem is specific: it appears when you have four or more options, when some of them are similar to each other, or when you need the outcome to command actual majority support. Outside those conditions the extra complexity of ranking is not worth it.
Can I reproduce this?
Yes, and please do. The methodology is described in full above, including the preference model, the tie-breaking rules, and the number of trials per condition. The instant-runoff implementation was verified against a worked example with a known outcome before any results were generated. Anyone implementing the same model should reach the same figures within sampling error.
What are the limitations?
Several, and they matter. The spatial preference model is an assumption, not an observation — real group preferences may cluster differently. We modelled sincere voting only, with no strategic behaviour. Ties were broken randomly, which affects small groups more than large ones. And these are simulated votes, so the findings describe what the model implies, not what any particular team experienced.
Why did VoteGenerator run this?
We make a poll tool that includes ranked-choice voting for free, so we have an obvious interest in the answer. That is exactly why the methodology is stated in full and the results include the cases where the two methods agree — which, with three options, is most of the time. Read it with our interest in mind and check the arithmetic yourself.

Run a ranked-choice poll

Free, no account, results show every elimination round.

Try Ranked Choice →