Asked to comment on the latest poll results, politicians who haven't done so well usually like to say they don't believe the results and that "the only poll that counts is on election day".
And that is true - it's only the votes counted on November 7 which will determine who is going to be in Parliament for the next three years.
But are they right to doubt the accuracy of the polls and question whether they give a useful insight into the mind of the nation?
An analysis of the final 1News poll before each of the last eight elections shows the last survey of the campaign has been within an average of 1.17 percentage points of the actual party vote, according to accuracy data compiled by the current pollster Verian.
The figures, covering every election from 2002 to 2023, translate to the poll being out by just one to two seats per party once run through the electoral system.
The most accurate final poll came in 2014, when the average error was just 0.66 points.
That year, Labour was polled at 25% and finished on 25.13%.
The least accurate was 2020, the Covid-era election held under shifting alert levels and heavy advance voting, when the average error rose to 1.72 points.
The final poll that year had Labour on 46% before the party actually swept to 50% and an outright majority. National was overestimated by 5.4 points, polling at 31% but finishing on just 25.6%.
The 2023 poll was among the closest yet, with an average error of just over one point. It put National on 37% against a final result of 38.06%, and NZ First on 6% against 6.1%.
Across parties polled over the eight elections, the surveyed result and the final count finished within a single percentage point of each other 33 times.
But some differences stand out across the two decades.
In 2011, NZ First was polled at 4.2% - below the threshold to return to Parliament - before winning 6.59% and eight seats. In 2017, the party was underestimated by 2.2 points in a result that ultimately made Winston Peters kingmaker.
These shifts are notable because seat allocation under MMP is highly sensitive around the 5% threshold, a two-point polling error on a minor party can shift the implied shape of the next government, while the same error on a major party is ultimately less significant.
Translated into seats, the poll has been out by an average of 1.5 seats per party per election, once overhang seats are set aside.
Verian said the comparison was adjusted to reflect what the poll is designed to measure.
The company's methods have changed significantly over the period, with every final poll from 2002 to 2017 using a landline-only sample.
The poll by Verian - formerly known as Kantar Public and before that operated as Colmar Brunton - moved to a blended 40% landline and 60% mobile sample in 2020. By 2023 Verian was drawing half its sample from an online panel and half from mobile random digit dialling.
SHARE ME