The Short Answer
Almost all research sample comes from one of three sources:
- River sampling — respondents intercepted in the moment, directed to a survey after clicking an ad or link while visiting a website or app, with no prior relationship and often no return visit.
- Opt-in access panels — people who registered and confirmed interest in taking multiple surveys over time, profiled in a database and invited when they match criteria. This is the dominant model in commercial research.
- Probability-based panels — members recruited through random selection from a known population frame. These generally have far fewer members than the non-probability panels that dominate online research, because member acquisition costs much more.
They differ on cost, speed, profile depth, and — critically — whether you can make population-level claims from the results. Most 2026 sample is blended across at least the first two, which is why your quote rarely names a single source.
River Sampling
How it works. Traffic is intercepted live. Someone browsing a site sees an offer, clicks, gets screened, and either qualifies or does not. The exchange fills your quota from whatever traffic is flowing.
Strengths. Fast, scalable, and reaches people who would never join a panel. The original argument for river sampling was access to a much larger and more diverse pool than even a very large panel — and it genuinely defeats panel conditioning, since respondents have no survey-taking history to condition them.
Weaknesses. You typically know little about who they are beyond what your screener captures, and researchers may never see them again. That makes verification harder, re-contact impossible, and longitudinal work unworkable. Self-selection is strong: you get people willing to click an offer mid-browse.
Use it for: fast, high-incidence consumer work; boosting a quota cell that a panel cannot fill; anything where reach matters more than a known profile.
Opt-In Access Panels
How it works. Members register, complete profiling questions, and get invited to studies matching their attributes. Their data is kept in a database that can be vetted for multiple member qualities, and members are trackable and reachable at any point.
Strengths. Rich profiling means better targeting and lower screening waste — a real cost advantage at low incidence. Members are re-contactable, so trackers and longitudinal designs work.
Weaknesses. Panel members are, by definition, people who chose to take surveys for rewards. That introduces two well-documented issues: professional respondents and panel conditioning, where repeated exposure changes the attitudes or behaviour you are trying to measure. Fewer than 1% of members across the ten largest panels once accounted for 34% of completed questionnaires — a concentration problem we cover in professional respondents and panel conditioning.
Use it for: most commercial research — targeted B2C and B2B, trackers, anything needing profile depth or re-contact.
Probability-Based Panels
How it works. Members are recruited via random selection from a known frame (address-based sampling, for example), so every unit of the population has a non-zero, knowable chance of inclusion. Non-internet households are often given devices or access to avoid coverage bias.
Strengths. The only model that supports genuine statistical inference to a population with a defensible margin of error. This is what serious public opinion and government research uses.
Weaknesses. Expensive, slower, and much smaller. Because acquisition cost per member is high, probability panels may require members to complete a minimum number of surveys — which reintroduces some conditioning risk. Feasibility for narrow B2B audiences is usually poor.
Use it for: published statistics, policy work, academic research, anything where "representative of the population" must survive methodological scrutiny.
Side by Side
| River | Opt-in panel | Probability panel | |
|---|---|---|---|
| Speed | Fastest | Fast | Slow |
| Cost | Low | Low–moderate | High |
| Profile depth | Minimal | Rich | Rich |
| Re-contactable | No | Yes | Yes |
| Conditioning risk | Very low | High | Moderate |
| Population inference | No | No | Yes |
| Best for | Reach, quota fills | Most commercial work | Published statistics |
The honest framing: river and opt-in panels are non-probability samples. They can be excellent for comparing options, understanding reasons, and tracking direction of change. They do not license claims like "62% of Americans" with a real margin of error, however the crosstab is formatted. Weighting improves comparability; it does not convert non-probability sample into probability sample.
Blending, and Why Your Quote Doesn't Name a Source
Most providers now assemble sample in real time across owned panels, partner panels, exchanges, and river traffic. Blending is how the industry hits feasibility at speed, and it is not inherently bad — but it makes deduplication and validation harder, and it means source composition can shift between waves of the same tracker.
If wave-over-wave comparability matters, ask your provider to hold source mix constant and to evidence cross-supplier deduplication. Our ESOMAR 37 buyer's guide covers exactly how to ask.
The Fourth Option Most Teams Overlook
There is a sourcing model missing from every vendor's taxonomy: your own customers.
For a large share of the questions product, marketing, and CX teams actually ask — why people churned, where onboarding breaks, what they would pay, which feature to build, why a deal was lost — the correct sampling frame is your user base. Incidence is 100%, identity is known, deduplication is trivial, and conditioning is a non-issue.
Koji is built for that frame:
- AI-moderated voice and text interviews with people you invite directly — import from CSV or sync from your CRM, then send personalized interview links so every answer is attributable to a known account.
- Runs in parallel, around the clock — 40 interviews take about as long as one, with no scheduling and no field window.
- Six structured question types — open_ended, scale, single_choice, multiple_choice, ranking, yes_no — so you get the distribution and the reasoning in one study, with AI follow-ups probing where a survey would just log a number.
- Automatic thematic analysis and one-click reports turn raw conversations into decisions without a synthesis backlog.
- No moderator bias — the same framing, every session, however many you run.
- Published pricing: €29/month Insights, €79/month Interviews, credits included.
Use river or panel sample when you genuinely need strangers — category entry, competitor customers, population norming. When the question is about your own product, buying sample means paying a premium to reach people who merely resemble your customers. And if you need external recruitment platforms, choose on sourcing transparency, not panel size.
Start free at koji.so — from question to insight in hours, not weeks, no research expertise required.
Frequently Asked Questions
What is river sampling in market research? River sampling intercepts respondents in the moment — people are directed to a survey after clicking a link or ad while browsing a site or app, with no prior registration. It is fast and diverse but offers little profile information and no ability to re-contact.
What is the difference between river sampling and panel sampling? River respondents are intercepted live with no ongoing relationship; panel members registered in advance, are profiled in a database, and can be re-invited. Panels offer targeting depth and re-contact, while river offers reach and no conditioning history.
What is a probability-based panel? A panel whose members were recruited through random selection from a known population frame, so every unit has a knowable chance of inclusion. It is the only online model that supports true statistical inference to a population, but it is expensive, slower, and much smaller.
Can non-probability samples be representative? They can be balanced to population targets on known characteristics through quotas and weighting, which makes comparisons and trends useful. They do not support a true margin of error or population-level inference the way probability samples do.
What is blended sample? Sample assembled in real time from multiple sources — owned panel, partner panels, exchanges, and river traffic. It improves feasibility and speed but complicates deduplication and can shift source composition between tracker waves.
Which sample source should I use for product research? For questions about your own product and customers, the best frame is your own user base — 100% incidence, known identity, no conditioning. Reserve river and panel sample for category entry, competitor customers, and general population benchmarks.
