How to read an Australian election poll

A published poll carries more information than its headline number, and most of it sits in plain sight: the sample size, the margin of error, the fieldwork dates, the methodology statement filed alongside it. This guide sets out what each of those tells a reader, how Australia’s preferential system shapes what the numbers mean, and where the documentation behind every accredited poll can be found.

Every worked example on this page uses a settled result or a published poll identified by date. No current voting intention is reported here.

This guide was researched, written and visualised by Kami Rostami and the Komms-Haus™ team, and paid for by Ace Strategies. Ace Strategies commissions and publishes the South Australian DemosAU polls used as worked examples on this page. The federal DemosAU waves cited here were commissioned by Capital Brief or run by DemosAU on its own account, and each report names its commissioning client.

1. What a poll is

A poll estimates the view of millions by surveying a carefully assembled sample, usually between one and three thousand people, using sampling mathematics that long predates modern polling. You do not need to drain the ocean to measure its salinity; you need a flask drawn so that every part of the ocean had a fair chance of ending up in it.

The craft sits almost entirely in that drawing: recruiting respondents across ages, regions, education levels and voting histories, then weighting the responses so the finished sample matches the population on the characteristics that matter. Weighting adjusts the influence of each response; it adds nothing and removes nothing. The precision of the result is then a known quantity, which the poll states, and that stated figure is where careful reading begins.

2. The three numbers that travel with the headline

Every DemosAU/Ace Strategies poll publishes three figures alongside its results. The same three appear, in some form, with every accredited poll in the country.

Sample (n) link

How many people were surveyed. Larger samples are more precise, with sharply diminishing returns, which is why national polls cluster between 1,000 and 3,000, and why a 9,000-respondent study is an event.

Effective sample link

Weighting has a statistical cost, and the effective sample states it. In the June 2026 South Australian poll, 931 people were surveyed and the effective sample after weighting was 632: the weighted result carries the precision of an unweighted poll of 632. The margin of error is calculated from the effective figure, not the raw one, and publishing both lets the reader see exactly what the adjustment cost.

Margin of error link

The stated precision, typically ±3 to ±4 points at 95 per cent confidence: if the survey were run twenty times, about nineteen would land within that band of the true figure. The stated margin applies to a figure near 50 per cent, where sampling error is largest; for smaller figures the band is a little narrower, and the stated figure is the conservative one to use.

Three practical readings follow. A party on 26 per cent with a margin of ±3.9 sits somewhere around 22 to 30. A two-point gap between parties inside that band is a statistical tie. And a two-point movement since the last poll may be movement or may be noise; the trend across several polls settles it.

How these figures are produced for our own polls, wave by wave, is set out on the polling methodology page. Our reading of the August 2026 federal poll applies the test in this section to every party’s movement in turn, and shows which of the five clear it.

3. The Australian layer: primary votes, preferences, seats

Australia’s preferential system means the headline number is the beginning of the count. A reader of Australian polls holds three layers at once.

Primary vote link

First preferences: the most direct reading of each party’s own support, and the number that moves first when the electorate moves.

Two-party-preferred (TPP) link

The projection of where preferences settle when a contest reduces to two candidates. For most of the postwar era that pair was Labor versus the Coalition. In a growing share of contests it now involves other pairings, so a TPP figure is read alongside the pairing it assumes, and a poll states which pairing it has used.

Seats link

The number that forms governments, and the one most loosely related to the other two, because single-member electorates reward votes that are locally concentrated.

A two-party figure is not measured; it is calculated from the primary votes, and the calculation needs an estimate of where each minor party’s voters send their preferences. Polls arrive at that estimate in one of two ways, and the report states which. The treatment of people who have not decided changes the primary votes themselves, so it is published too.

Preference flows link

The share of each party’s voters whose preference reaches one side of the final pair rather than the other. A poll either applies the flows recorded at the previous election, or asks its own respondents how they would number their ballot and uses their answers. The first rests on real ballots, but from an electorate that voted differently; the second reflects the people surveyed, on a smaller and noisier base. When a minor party’s vote has moved a long way since the last election, the two methods can give different two-party figures from the same primary votes, and neither is wrong: they answer slightly different questions.

Undecided voters link

Some respondents say they do not know how they would vote. A poll either asks them a follow-up question about which way they lean and counts the answer, sets them aside and reports the rest as shares of those who chose, or carries them as a line of their own. Each choice changes every other party’s figure, so two polls can publish different primary votes from similar answers. The report says which was done, and the reader’s check is simple: see whether the published figures add to 100 with the undecided in or out.

Our own polls ask the leaner question and exclude those who still decline; in the May 2026 federal wave that was 7 per cent of respondents. Where a two-party figure is published, the report states the basis on which preferences were distributed, and where none is calculated, none should be attributed to the wave.

4. Reading polls in the plural

The mathematics of a 95 per cent confidence interval means roughly one poll in twenty will land outside its own margin of error even when everything has been done correctly. An occasional outlier is expected, and the way to identify one is comparison: against the same series’ previous waves, and against other houses in the same window.

The January 2026 movement in One Nation’s national vote illustrates the standard analysts apply. The elevation appeared at the same time across five independent polling houses (DemosAU, Newspoll, Roy Morgan, Redbridge and YouGov) using panel, phone and online methodologies. As part 1 of our series put it: a single house can be wrong; five independent series moving together, across different survey modes, is a measurement.

Rapid movement is read by the same test. When one national series recorded a nine-point fall in a party’s support across a fortnight in mid-2026, the question was the one this section began with: is it movement or noise? The answer comes from the series’ own next waves and from the other houses in the same window, and from nowhere else. A single poll is a photograph of a moving object; a sequence of them shows the motion.

5. What the margin of error leaves out

The margin of error is a statement about one thing: the chance variation that comes from surveying a sample rather than everyone. It says nothing about error that would still be there if the sample were ten times larger. Four kinds of that error are worth knowing by name, because the documentation behind a poll is where each one shows.

House effect link

A pollster’s tendency to read a party higher or lower than the other houses in the same period. It comes from the choices a house makes in how it recruits, how it weights and how it asks, and it is often steady from wave to wave. A house effect is measured against the field, not against the truth, and the field can share a lean, as it did in 2019. A house that reads a party two points above the average every time is a house to read with those two points in mind, and a two-point movement within that house is still a movement.

Mode effect link

The difference that comes from how people were reached: an online panel, a phone call, a text message with a link, or an interviewer at the door. Each mode reaches a different mix of people and puts the respondent in a different frame of mind. A poll conducted in one mode is not wrong for it, but a comparison across modes is a comparison across two kinds of sample, and the mode is published for that reason.

Non-response link

The people a poll could not reach, or who declined, differ from the people it did. Weighting corrects for the differences it can see, such as age, region and education. It cannot correct for a difference it does not measure, and the people who agree to take surveys tend to be more engaged with politics than those who do not. The gap that remains is a bias, not noise, and no sample size shrinks it.

Herding link

A set of final polls more alike than independent samples would be. When the last polls before an election all land within a point of one another, either sampling error has taken a holiday or some houses have adjusted towards the consensus. Agreement between polls is reassuring only when each was arrived at on its own.

6. The other numbers in a poll

Voting intention is the headline, and it is rarely the only question. The same report usually carries a leader question, the results broken down by group, and sometimes the Senate. Each is read with its own caution.

Leader ratings link

Two questions travel together. A satisfaction or approval rating asks whether the respondent approves of a leader’s performance, and is usually published as a net figure, the share approving minus the share disapproving, so a fall in “don’t know” can move it without anyone changing their mind. A preferred-leader question asks the respondent to choose between named people, and it tends to favour the person who holds the office, whose name arrives with the title. For both, the movement from wave to wave says more than the level, and a leader’s rating is not a vote.

Subgroups and crosstabs link

The results broken down by state, age, gender, region or past vote. Each cell rests on a fraction of the sample and carries its own, wider margin. A national poll of 1,500 carries a margin near ±2.5 points on its topline, leaves a few hundred respondents in a mid-sized state and a few dozen in a small one, and a figure from 200 people carries a margin near ±7 points, one from 50 near ±14. A crosstab is for reading direction across several waves, not for reading a level from one.

The Senate link

Most polls ask about the House of Representatives. A Senate question is asked less often, and the Senate is counted differently: seats are shared in proportion to the vote within each state, so the gap between votes and seats that section 3 describes for the lower house is much smaller in the Senate. A Senate figure is a state-by-state question, and a national figure for it is of limited use.

7. MRP, briefly

Multilevel regression and post-stratification addresses a limit of conventional polling: a national sample of 1,500 says little about any single electorate of 110,000, where perhaps ten respondents live. MRP starts from a much larger sample — DemosAU’s May–July 2026 federal model drew on 9,343 respondents — and models how demographics relate to vote choice: how age, education, income, region and past vote combine. It then applies that model to the census composition of each of the 150 electorates.

The output is a projection with stated uncertainty, which is why credible MRP results appear as seat ranges drawn from thousands of simulations rather than as a single seat count. A range such as “60 to 72 seats” is the method reporting the uncertainty that attaches to it. An MRP model reports no effective sample size, because the estimate for any one seat draws on the whole model rather than on the respondents in that seat; the variables and the seat-level margin for our own models are on the methodology page. A worked set of those ranges, and what a five-point vote change does to them, is in our reading of the August 2026 federal poll.

The method has its own limits, and they differ from a conventional poll’s. An MRP estimate for a seat is only as good as the model’s assumption that people of a given age, education, income and past vote behave alike wherever they live. A seat whose politics differ from its demographics, because of a strong local candidate or a local issue the model has no variable for, is the seat the model will get most wrong, and it will get it wrong with a confident-looking range. The sample beneath the model is still a sample, so a bias in who was recruited is modelled, not removed, and every seat inherits it in the same direction. The reader’s check is the same as for any poll: the range the model states, and its record against declared results.

8. What a poll measures, and when

A poll measures stated intention on the days of fieldwork. Between a poll and an election sit campaigns, candidates, preference decisions and events, so a survey taken eighteen months out is a reading of conditions at the time, and pollsters publish fieldwork dates precisely so it can be read that way.

The connected question is whether stated intention converts into behaviour when an election arrives, and Australia has a clean recent data point. At the March 2026 South Australian election, the first statewide contest after the sharpest polling movement in a generation, the final DemosAU/Ace Strategies pre-election poll (fielded 12–18 March, n = 1,242) identified the ordering and the approximate magnitude of every major party’s vote, with the two leading parties each within about a point of the near-final count. Polled support in the low twenties converted into votes at that level.

9. The documentation behind the numbers

In 2019, the final published national polls collectively pointed to the wrong federal election winner, for the reasons section 5 sets out. The industry reviewed its methods, and in 2020 the major research houses established the Australian Polling Council, whose Code of Conduct binds members to public disclosure whenever their political polling enters the public domain.

The disclosure is specific. A poll carrying the APC Quality Mark must publish, within two business days, a long methodology statement covering the full question wording and preambles, who conducted and who paid for the research, the fieldwork dates, how respondents were recruited, the weighting variables and method, the effective sample size, and the margin of error attached to it. The statements are public, so any reader can check the work directly.

The record since 2019 is part of the public documentation too. Polling at the 2022 federal election landed close to the result. The 2025 federal polls identified the winner and the direction of the swing, and understated its size. DemosAU’s final 2026 South Australian pre-election poll came within about a point on the two leading parties. DemosAU is a member of the Australian Polling Council and the Research Society; its methodology statements are published at demosau.com.

10. The reader’s checklist

These are the questions a poll’s documentation answers, and where each answer is published. Work through them against a poll you are reading; the ticks are kept in this browser, so you can come back to a half-finished one.

Every DemosAU/Ace Strategies poll publishes the documentation these questions are answered from. DemosAU’s methodology statements are published at demosau.com; our own polling practice is set out on the polling methodology page, and every wave is listed in the polling archive.

Common questions