Almost every online survey uses quota sampling, not random sampling — including most market research and political polling. Here's the real difference, when each actually applies, and what it looks like to manage quotas in a live study rather than just define the term.
By the Wavefield Research team · Published Jul 27, 2026 · Updated Aug 27, 2026
Both methods try to make sure your sample looks like the population you care about. The difference is how respondents actually get selected.
Every member of the population has a known probability of being selected — classically, an equal one. Because selection is probabilistic, you can formally calculate sampling error and a true margin of error. It's the gold standard when it's feasible, but it needs a real sampling frame (a way to reach anyone in the population) and is slower and more expensive to field.
The population is split into cells (e.g. age × gender × region), each with a target count. Respondents are recruited — often from an online panel — until each cell fills, with no probability guarantee on who specifically fills it. It's fast, cheap, and it's what the large majority of commercial and political online research actually runs on.
Swipe the table sideways to see all columns →
| Random sampling | Quota sampling | |
|---|---|---|
| How respondents are chosen | Every member of the population has a known, typically equal, chance of selection | Respondents are recruited until each defined cell (e.g. age × gender) hits its target count |
| Statistical error | True sampling error is calculable — the basis for a formal margin of error | No true sampling error exists (it's non-probability); margin of error is reported by convention, not derived |
| Speed and cost | Slower and more expensive — needs a real sampling frame and often multiple contact attempts | Fast and cheap — the standard approach for online panels and most commercial research |
| Representativeness | Representative by design, before any weighting | Representative on the quota variables only (e.g. age, gender) — everything else can still be skewed, which is why post-survey weighting matters |
| Where it's the norm | Academic research, official statistics, some high-stakes probability-based panels | Market research, most political/opinion polling, virtually all online panel research |
A quota is a logic condition (who it applies to) plus a limit. Every completed response is checked against every quota live, and once a cell hits its limit, new respondents matching it are handled per how that quota's configured — cleanly terminated, or accepted and flagged for review.
True random sampling needs a sampling frame that reaches the whole population you care about — a real list, a working phone-number universe, something exhaustive. Most market research and online political polling doesn't have that, so quota sampling from an online panel or a client's own contact list is the practical default. The honest fix for quota sampling's representativeness gap isn't pretending it's random — it's weighting the completed sample afterward against known population targets, so the numbers you report reflect the population even though the raw sample doesn't perfectly.
Quota and simple random sampling aren't the only options — and knowing the neighbors clarifies what quota sampling actually is. In textbook terms this is the probability sampling vs. non-probability sampling divide: random, stratified, and systematic designs are probability sampling; quota, convenience, and river sampling are non-probability designs that control composition instead of selection odds.
Splits the population into groups (strata) like quota sampling does — but within each stratum, selection is still random with a known probability, so a true margin of error survives. It's what quota sampling imitates when a real sampling frame isn't available: same idea (control the mix), different mechanism (probability vs. availability). If you have a complete member list — employees, a customer database — stratified random beats quota and costs nothing extra.
Take every k-th person from an ordered list — every 10th customer, every 25th voter file record. Behaves like random sampling when the list order is unrelated to what you're measuring, and it's operationally simpler. The trap is periodicity: if the list has a repeating pattern (say, one manager per team of ten), a matching interval samples all managers or none.
Whoever shows up: social-media polls, website intercepts, open shared links. This is quota sampling with the quotas removed — no composition control at all — and it's where most self-run “research” quietly lives. The honest uses are exploration and pretesting. The moment a number will be quoted to a stakeholder, add quotas at minimum, and weight the result.
Quota sampling's weakness isn't the missing margin-of-error formula — it's that quotas only control the variables you set. A panel study quota'd on age, gender, and region can still over-represent the extremely online, the survey-professional respondent, or whoever the panel recruited last month. The failure shows up as results that shift when the supplier changes, even with identical quotas. Three defenses, in order: quota on the variable closest to your topic (past vote for polling, category usage for brand work — not just demographics); weight the completed sample to trustworthy population targets rather than assuming quotas finished the job; and keep field quality controls on, because a perfectly balanced sample of straightliners is still garbage. This is also why serious shops treat quota targets and weighting targets as two different things — quotas manage the field, weights fix what the field still missed.
Both split the population into groups and control the mix, but they differ in how individuals get selected. Stratified random sampling selects randomly within each stratum from a complete sampling frame, so every person has a known probability of selection and a true margin of error survives. Quota sampling fills each cell with whoever is available and qualifies — no frame, no known probabilities — which is faster and cheaper but borrows its error math by convention. Rule of thumb: with a complete population list (employees, a customer base), stratify; without one, quota and weight.
Advantages: it needs no sampling frame, fields fast, costs far less than probability designs, and guarantees the demographic mix you set — which is why nearly all commercial online research uses it. Disadvantages: selection within each cell is by availability, so anything correlated with being reachable and willing can still skew; there is no formally valid margin of error, only the conventional approximation; and quotas only control the variables you set. The standard mitigations are careful screeners, post-survey weighting, and honest reporting of effective bases.
It can be, and that's the honest tradeoff for its speed and cost. Quotas guarantee the right mix on the variables you set (age, gender, region), but everything else — attitudes, behaviors, anything correlated with who's easy to recruit online — can still skew. That's why quota sampling is almost always paired with post-survey weighting rather than used alone.
Not a true one — margin of error is formally a probability-sampling concept. In practice, the market research and polling industry reports a margin of error on quota samples anyway, as a conventional approximation, and good practice is to compute it on the weighted effective base size (accounting for how much the weights vary) rather than the raw completed count.
Stratified random sampling also splits the population into groups (strata) — but within each stratum, selection is still random with a known probability. Quota sampling splits into the same kind of groups, but fills each one with whoever's available until the quota is hit, with no randomization guarantee. Same idea (control the mix), different mechanism (probability vs. convenience).
Yes — quotas are defined as cells (a logic condition plus a limit), and every completed response is checked against them live. When a cell fills, respondents matching it are either cleanly terminated or accepted and flagged, per how the quota's configured, and the field monitor shows fill counts in real time. Screened-out responses never count toward a quota's fill — only completes do.
Free tool: sample size calculator · Comparing platforms: Wavefield vs. Qualtrics · Wavefield vs. SurveyMonkey · Read survey weighting, explained and what survey programming involves and how to analyze survey results · How the sample gets reached is its own decision: survey data collection methods, compared · Who gets into the sample at all is the screener's job: survey screening questions
Upload a brief and the AI agent programs your quotas, screeners, and logic. You don't pay until you field.
Start a project