British pollsters have a big problem to solve as the country heads towards an election: Brexit has scrambled traditional political allegiances and they say it is harder than ever to know whether voters are telling them the truth.
Polling companies, many of whom underestimated support for Brexit in the 2016 referendum and miscalled an election the previous year, are changing who they ask and how, and trying new ways to get to know voters better than they know themselves.
With the Brexit process still unresolved and financial markets on edge, the stakes are high. But the divisive national debate over whether and how to leave the European Union has complicated the pollsters' art of finding a few thousand people who reflect the mood of millions of voters.
"You can ask someone 'How certain are you to vote for that party?' but as a species we're not very good at explaining our own behaviour," said Joe Twyman, Director of Deltapoll, a new company set up by staff from some of the established players.
"It's far better to use underlying data to register that."
More voters switched between the two main parties at a 2017 election than in any ballot dating back to 1966, research by the British Election Study showed. The more people change their minds, the harder it is to draw a representative sample.
Now people are also switching between a greater number of parties, with Brexit propelling new and smaller parties, such as the Brexit Party and the pro-EU Liberal Democrats, to the fore and making voter behaviour more unpredictable.
Pollsters partly redeemed themselves in the 2017 election after miscalling an earlier one in 2015, but still failed to fully capture the swing which lost the governing Conservatives their majority.
One of the reasons was that some of the adjustments they made after 2015, particularly on predicting voter turnout, overcompensated for what they saw as failings in earlier models. Some said they had tweaked the models again to try to fix this.
Today, depending on which of some half dozen leading opinion polls you look at, Prime Minister Boris Johnson's Conservatives are either 15 percentage points ahead of Jeremy Corbyn's rival Labour, or dead level with them.
All parties want an early election but disagree over when it should be held.
Even information about how people voted last time, crucial to find a representative sample, can be unreliable.
Just after the 2017 election, 41% of respondents said they had voted Labour; two years later, the proportion of those same respondents who said they voted Labour had fallen to 33%, YouGov said, suggesting pure forgetfulness, misremembering tactical votes or a desire to be seen to have backed the winning party.
Opinium's Head of Political Polling, Adam Drummond, said his firm got more accurate results when it first asked whether someone had voted in a previous election rather than listing parties they may have voted for alongside a 'didn't vote' box.
Others are trying to gather new information to better measure underlying behaviour.
Deltapoll said it was looking at ways to gauge voters' "emotional resonance" with particular parties and policy issues to better understand their behaviour and give a more accurate picture of how they would vote.
The voting plans of someone who feels passionate about every issue, including their political allegiance, might be given less weight than those of a generally indifferent voter with a strong emotional attachment to their chosen party, it said.
Tweaking And Innovating
Typical political polls rely on between 1,000 and 2,000 responses, are conducted online and - as polling firms who often feel they are unfairly criticised are keen to point out - only provide a snapshot of public opinion.
"We want people to trust the polls but we want people to treat them with an appropriate awareness of their limitations," said Gideon Skinner, Research Director at Ipsos MORI.
Everything depends on the quality of the initial dataset.
Samples are selected from large pools of possible respondents according to variables ranging from the basics: age, sex and income, to much more in-depth data like past voting records, political awareness and education level.
YouGov and Ipsos MORI both said they were looking more closely at political engagement and education levels among their respondents - a way of countering the fact that those responding to polls tend to be more educated and politically active.
Survation, YouGov and other pollsters are also trying more sophisticated data analysis to navigate Britain's voting system, based on constituencies rather than proportional representation.
Multilevel Regression and Post-stratification', better-known as MRP, is a technique that was the first indicator in 2017 that the Conservatives could lose heartland seats and fall short of an overall majority.
YouGov, which published that poll, says it combines respondents' information with previous polling results, census data and official statistics to come up with a seat-by-seat analysis of how each constituency might vote.
These types of polls work best when carried out close to polling date, once candidates in each seat have been named - and while not perfect are likely to be closely watched whenever Britain next heads to the polls.