Online panels in Vietnam: who answers an online survey, and how to read what it tells you

The request usually sounds simple. Can we run a quick online survey in Vietnam before we decide? Usually the answer is yes. An online panel can put a questionnaire in front of several hundred consumers within days, at a fraction of what face-to-face fieldwork costs.

Speed is the reason panels get used, and it is also why their numbers travel further than they should. A percentage from a two-week online study ends up in a board paper as a fact about Vietnam. Whether it is one depends on four things that rarely appear on the slide: who is in the panel, how many answers sit behind each figure, what the weighting did, and which answers were thrown away.

Choosing the supplier is a separate question, covered in the guide to choosing a research firm. This piece is about the study itself: how an online sample in Vietnam is put together, how big it needs to be, and how to read it once the deck arrives.

Who is in a Vietnamese online panel, and who is not

Most Vietnamese adults are online. DataReportal’s “Digital 2026: Vietnam” counts 85.6 million internet users in October 2025, 84.2% of the population. Being online is not the gap. Being on a panel is.

Panel members are people who signed up to answer surveys in return for points or vouchers, and they are not a cross-section of the online population. The clearest public picture comes from a panel owner’s own documentation. Rakuten Insight’s panel book, published for buyers, compares its Vietnam panel with the 2019 census:

That is not a criticism of one provider. Opt-in panels are built this way everywhere, and the owners say so. Q&Me states on its website that its online panel is not the right platform for reaching people over forty or in rural areas, and works best for 18 to 39-year-olds.

Set that against the population a plan usually means. Vietnam’s Statistics Office puts the 2025 population at 102.3 million, with 38.6% living in urban areas, and its population survey at 1 April 2025 counts 9.6% aged 65 or over. For most mass-market categories, the people a panel reaches easily are a minority of buyers. In practice:

• Urban adults under forty in the two largest cities: well covered. A panel is usually the right tool.
• Adults over forty, secondary cities and provinces: present but thin. Set quotas, and ask for active member counts by province before fieldwork.
• Rural households and people over sixty: sparse, and the members who exist are not typical. Telephone or face-to-face fieldwork is the honest option.
• Business decision-makers: a job title in a panel profile cannot be verified. Use recruited interviews.

Where the sample actually comes from

“Online panel” covers several different ways of finding respondents, and a single study often mixes them.

• A provider’s own panel. People who registered with one company to answer surveys, usually for points or small rewards. The provider knows their profile and history.
• A panel exchange. Software that fills one survey from several panels at once. It reaches more people, and it also makes it more likely that the same person appears twice, under two memberships.
• Recruitment for this study only. Respondents found through social media advertising or website intercepts and sent straight to the questionnaire. Fast, cheap, and shaped entirely by who clicks on the advertisement.
• Your own customer list. Useful, but it describes your customers. It is not a sample of the market and should never be presented as one.

ESOMAR’s “Questions to help buyers of online samples” (March 2021, 37 questions, replacing its earlier 28) describes how wide the sources have become: loyalty and rewards communities, customer lists, offer walls, affiliate networks and social media, as well as traditional panels. It also notes that different sources can produce different results. It is the most useful single document to send a supplier, because it asks exactly what this article raises: where each respondent came from, how a person registered in several panels is caught, and how random, inconsistent or too-rapid answers are removed.

The practical point is that “we will use an online panel” is not yet an answer. The proposal should name the source or sources, the share of completes expected from each, and how duplicates across sources will be removed.

How many answers each figure needs

The total sample is the number everyone looks at, and it is rarely the number that matters. What matters is the smallest group you intend to report on, because that is where a difference between two figures stops meaning anything.

These are the textbook margins at 95% confidence, and they assume a random sample. For an opt-in panel they are a floor, not a measurement. The American Association for Public Opinion Research states that the margin of sampling error applies only to probability-based surveys and does not apply to opt-in online surveys. Read each margin as the least uncertainty the figure carries; the real uncertainty also includes everyone the panel left out.

A study of 600 people reported by two cities and three age bands has about 100 people in each of six groups. A twelve-point gap between two of those groups can be noise. The same 600 reported only by city has 300 per group, and the same gap almost certainly is not. Decide the groups before fieldwork, write them into the brief, and size the sample from them.

Quotas and weighting: what they correct, and what they cannot

Two tools pull a panel sample toward the population it is meant to describe. Quotas are set during fieldwork: so many completes per city, age band and gender, after which the survey closes to that group. Weighting is applied afterwards: groups that ended up too small count for more, groups that ended up too large count for less.

Weighting has a price that rarely appears in the deck. When some respondents count three times and others half, the sample behaves like a smaller one. The usual measure is the effective sample size: the square of the sum of the weights, divided by the sum of the squared weights. Eight hundred completes with heavy weighting can carry the precision of six hundred, and every margin in the table above should be read against that smaller number.

There are three things weighting cannot do:

• Create people who are not there. A group with no respondents cannot be weighted up.
• Make a member typical of their group. A 55-year-old in a provincial city who answers surveys for rewards is not the average 55-year-old in that city, and counting their answer four times multiplies whatever makes them different.
• Correct for a variable it did not use. Weighting by city, age and gender leaves income, education and how often people shop online exactly as the panel delivered them.

Ask for the weighting targets and where they came from, the range of the weights, and the effective sample size. A supplier that weights to official population figures should be able to name the table.

Checking the answers before reading them

Some of the people who complete an online survey are not really answering it. They click through as fast as the screens allow, give the same answer down every row of a grid, or type random characters into the open questions. Most online studies contain some unusable completes, and the real question is whether they were found.

The usual checks, which a proposal should list:

• Speed. Completion time far below the typical respondent’s, against a threshold set before fieldwork.
• Straight-lining. The same answer to every statement in a grid, including statements that contradict each other.
• Attention checks. An item that tells the respondent which answer to choose.
• Open answers. Empty, nonsensical or copied text where a sentence was asked for.
• Consistency. An age, a household or a purchase history that contradicts an earlier answer.
• Duplicates. The same device or identity appearing more than once, especially across sources.

A useful habit is a soft launch: review the first tenth of completes before releasing the rest, when a broken question or a bad source can still be fixed. Then ask for the number removed at each check. A study that removed nothing did not look; a study that removed a large share is telling you something about where its sample came from.

Ask, too, what share of completes came from smartphones; it is one of the figures ESOMAR’s checklist asks suppliers to report. On a phone, length and layout are part of quality. A long questionnaire with wide grids loses attention halfway through, and the answers in the second half are the ones that suffer.

Questions a panel answers well, and questions it answers badly

The pattern is consistent. Panels are strong at comparisons among people they reach well and weak at absolute numbers about people they reach badly. An entry plan built on a survey that was asked an absolute question inherits every gap in the panel.

Stated purchase intent deserves its own warning. “Would definitely or probably buy” is an answer about a hypothetical, given by someone who will not pay for it. It is useful for ranking two options against each other. It is not a share of the market, and multiplying it by a population figure produces a forecast nobody should defend. How a category is actually bought, including the role of marketplaces and chat, is usually better read from the places buyers compare products than from what they say they will do.

One study, read twice

An illustrative example, with invented figures. A European breakfast cereal brand commissions 800 online interviews, 400 in Hanoi and 400 in Ho Chi Minh City, among adults aged 18 to 54 who bought a packaged breakfast product in the last three months. The headline in the deck: 42% would definitely or probably buy the new range at the proposed price.

Read again, one step at a time:

The weighting arithmetic is simple. Younger respondents count 0.64 each and older respondents 1.83, so the weighted intent is 0.45 × 48% plus 0.55 × 28%, which is 37%. The effective sample size comes from the same weights: about 620 out of 800.

Nothing in the study was done badly. The panel delivered the people it has, the supplier weighted correctly, and the two cities are reported separately. The difference between 42% and 37% comes entirely from who answered, and the board paper had quoted the first figure without the second. The more serious misreading would come next: treating 37% as the share of urban households that will buy. It is a measure of how this concept compares with others shown to the same people, and nothing more.

What to receive besides the deck

A report is a summary of choices the reader cannot see. These are what make the choices visible, and they should be named as deliverables in the contract:

• The questionnaire as fielded in Vietnamese, with the English version used for sign-off.
• Respondent-level data, with the quality flags kept rather than the flagged rows deleted.
• The weighting targets, their source, the weights themselves and the effective sample size.
• The number of completes removed at each quality check.
• Field dates, the sample source or sources, and the incidence actually achieved.

Respondent-level data should arrive without names, phone numbers or other identifiers unless there is a reason to hold them. Vietnam’s Law on Personal Data Protection, in force since 1 January 2026, makes consent valid only when it is voluntary and the person knows what data is processed, for what purpose and by whom, and it does not treat silence as consent. A brand that later wants to recontact respondents, or match them against its own customer list, needs that purpose in the consent the supplier collected, not added afterwards.

Where the figures in this article come from

The internet-user figure is from DataReportal’s “Digital 2026: Vietnam”, using data for October 2025. Population, urban share and the share aged 65 or over are from Vietnam’s Statistics Office: its press release on population and labour for 2025, and the main results of its population change survey at 1 April 2025. The panel and census comparison is from the panel book Rakuten Insight publishes for buyers, which uses the 2019 census; the statement on panel age and rural reach is from Q&Me’s website. The checklist is ESOMAR’s “Questions to help buyers of online samples” (2021), and the position on margins of error is AAPOR’s. The margins in the sample table use the standard formula at 95% confidence. The breakfast cereal study is illustrative, with invented figures.

Where this work stops

MWY does not run fieldwork or own a panel. Vietnam Market Research is six weeks of desk research, platform and marketplace data and interviews with people who work in the category, written toward the decision the research has to support. Where that decision genuinely needs primary consumer data, the survey is scoped and quoted separately, or placed with a research firm we introduce and take no commission from.

What MWY adds to an online study is the reader’s side of it: writing the groups and the decision into the brief, checking the proposal against the questions above, and reading the delivered data before its headline reaches a board paper. The same research opens Go-to-Market Strategy, and the wider sequence of an entry is set out under market entry.