Politicians say the only poll that matters is the one on election day, and right now, polls are pointing to a historically tight race between Kamala Harris and Donald Trump.
Polls might show Harris leads Trump by around 2%, but it's also true that Trump has never been closer in the national polls to a Democratic rival.
Even though most pollsters correctly predicted Joe Biden's 2020 election win, his eventual margin over Trump was far closer than pre-election polls suggested.
Trump also won the 2016 presidency convincingly despite polls suggesting Hilary Clinton would win in a landslide. Pew Research later found at least 88% of national polls overstated each Democratic candidate's popularity.
"National polls are misleading in trying to generalize what's going to happen," Thomas Gift, director of the Centre of US Politics at University College London, tells DW.
"It looks like, right now, for example, Kamala Harris is going to win the popular vote — she's up by a couple of percentage points. But I think it's unclear if she will win the Electoral College."
Right now, poll analysts believe America is evenly split between the candidates.
So, will they be right this time around? That depends on whether pollsters can find a particular Trump-voting section of the population.
The art and science of capturing the American voter
When assessing voter intent, pollsters try to account for as many variables as possible.
"We measure people that are quite different from one another, and we provide that information to the public," says Don Levy, the director of Siena College Research Institute, which produces what is considered one of America's best-quality polls with the New York Times newspaper.
As a rule, opinion polls will aim to randomly sample enough "likely voters," often abbreviated to LV, to produce a result within a 95% confidence level — which means the same value will occur 95 times out of 100 — and within a certain margin of error, often around 3-4%.
The sample size needed to meet these parameters is low. Around 600 people is all a pollster needs to represent a population of 100,000 people within a 4% margin of error. For a 3% margin, you need to clear 1,000 people. This is the science of election polling.
The art is finding the right representative mix in the sample to make the survey as accurate as possible, and each pollster has a unique method.
It begins with determining whether someone is an LV. With barely half of the eligible population turning out to cast a ballot in recent elections, there's little point polling someone who won't participate.
Siena does this by marrying a voter's attendance history at polling booths with a verbal interview over the phone.
Once they meet Siena's LV threshold, they'll be asked about their voting opinion and are then demographically categorized to build a quota for the poll — the more granular these voter samples go, the more robust the poll's findings will be.
Siena has about 40 unique quotas that it targets to represent an electorate's demographics accurately, including variations of gender, ethnicity, age, education level, and so on.
"We try extremely hard to quota those samples not only be the overall United States or the overall state of Pennsylvania but by regions of the state," Levy says.
Finding the hidden Trump voter
Getting representative samples across 40 quotas is no easy task. Clearly, there was a glitch in polling methodologies that substantially underestimated Trump's standing in American electorates in 2016 and 2020.
Levy puts down the gap between polls and the final election tallies to pollsters struggling to capture a particular subset of the population, one he calls the "anti-establishment, non-response bias" — Trump-supporting Americans who refuse to participate in the polls trying to include them.
"Virtually every [pollster] had the same error creeping in."
Levy suspects being unable to capture this "anti-establishment" voter was worth "three to seven points of error" in 2020 alone. The remedy? To count the "drop-offs."
"There were a meaningful percentage of respondents who I'd call up and say 'this is Don calling from Siena College Research Institute doing a survey today,' and they would simply go 'TRUMP!' and hang up,'" Levy says.
"In 2020 ... we kept track of them, but they didn't count, and when we looked back, we found that had we counted those people, it would have corrected about 40% of the error. So we count them now."
Polling, prediction and Pennsylvania
Unlike other systems, US presidents are decided by which candidate secures at least 270 of 538 Electoral College votes, not just the popular vote.
These votes are allocated to each state, corresponding to their number of Congressional members, which are then pledged (usually in a "winner takes all" arrangement, though Nebraska and Maine are exceptions) to its most popular candidate.
This complex system has previously led to both Trump and George W. Bush (in 2000) being elected with fewer paper votes than their opponents.
To predict the Electoral College, analysts and commentators translate poll data into maps of predicted red, blue and "toss-up" states by predicting which candidate will win the popular vote in each jurisdiction.
Right now, analysts generally agree seven states considered 50/50 contests will decide the presidency: Arizona, Georgia, Michigan, Pennsylvania, Nevada, North Carolina and Wisconsin.
Accepting that polls aren't perfect, the race's closeness means many analysts are watching one state closely: Pennsylvania, a recent bellwether state that has sided with the elected president in the last four elections and carries 19 prized electoral college votes.
"It's very hard to imagine either candidate getting to the White House without winning Pennsylvania," says Gift, the University College London expert, himself from Pennsylvania.
Gift says the amount of money being spent by both sides in the Keystone State — and the attention being paid to it — is indicative of its importance.
"Candidates are doing everything that they can to win Pennsylvania. I really think that it is the linchpin to this election," says Gift.
