A single survey shows one result and the average of many shows another. The gap comes from two different kinds of error, only one of which averaging can remove.
A poll measures a sample and estimates a population
No survey contacts everyone. It contacts a group and treats their answers as representative of a much larger population.
The margin of error published alongside a result describes only the uncertainty created by taking a sample at all, assuming everything else was done perfectly.
Two competent pollsters asking identical questions on the same day will still produce different numbers for this reason alone.
Averaging cancels random error
Random sampling error is as likely to overstate as understate. Combine many surveys and those deviations tend to offset.
The average therefore has less noise than any individual poll in it, which is the whole reason the method exists.
This is also why a single poll showing a dramatic move is usually treated cautiously until others move with it.
Systematic error survives the process
The errors that matter are the ones shared by everyone. If a group of people is systematically harder to reach, and every pollster misses them in the same direction, averaging preserves the miss.
Response rates to telephone surveys have fallen sharply over recent decades, which makes the composition of the group that does respond more consequential.
Pollsters correct for this by weighting their sample to match known population characteristics, and the choice of which characteristics to weight on is itself a judgement.
House effects come from methodology, not intent
Firms differ consistently in their results because they differ in method, including whether they call landlines, use online panels, screen for likely voters, or push undecided respondents.
These persistent tendencies are called house effects, and serious averages adjust for them rather than treating every survey as interchangeable.
Weighting by recency and by sample quality is common for the same reason.
Averages describe a moment, not a forecast
An average summarises what surveys measured over a recent window, and surveys measure stated opinion rather than eventual behaviour.
Turnout is the largest unknown, since a respondent's intention to participate is self-reported and imperfectly predictive.
Reading an average as a prediction attributes to it a job it is not doing, which is the most common misreading of polling data.