Key Takeaways
- Respondent identity is the weak point now, not question wording. NORC at the University of Chicago put fraud rates across the market research industry at 15% to 30% in April 2026, reaching 45% on some platforms.
- Standard quality checks no longer separate people from models. A 2025 study in PNAS built a synthetic respondent that passed 99.8% of them.
- Your own customer list is the highest-integrity sampling frame you have. Authenticated contacts cannot be farmed the way an open survey link can.
- Surveys tell you what, follow-up calls tell you why. Book the interview inside the survey itself, while the respondent still remembers what they answered.
Most guides on customer research surveys spend their length on question types. That was the right emphasis when the hard part was asking well. The hard part in 2026 is knowing whether a person answered at all. NORC at the University of Chicago reviewed the evidence in an April 2026 research brief on fraudulent respondents and bots and put fraud rates across the market research industry at 15% to 30%, as high as 45% on some survey platforms.
So this guide treats identity as part of the method. Sampling frame first, questionnaire second. If you get the frame wrong, a beautifully written questionnaire just produces confident nonsense faster.
What survey fraud does to a research read
The failure mode is not that a few junk rows sneak in. It is that entire datasets turn out to be synthetic, and nothing in the data itself announces this.
NORC's brief collects the case evidence. Krawczyk and Siek (2024) received 981 responses within four hours and found three authentic. Goodrich and colleagues (2023) could verify 4% of 2,622 responses to a US beekeeping survey.
Pozzar and colleagues (2020) reported 94.5% fraudulent responses in a health study recruited through social media. Across several case studies, usable responses fell from roughly 75% to around 10%.
The screening tools most teams rely on are the part that has aged worst. Westwood's 2025 paper in the Proceedings of the National Academy of Sciences, The potential existential threat of large language models to online survey research, describes an autonomous language-model respondent that passed 99.8% of standard data-quality checks and produced coherent, persona-consistent answers. Attention checks, straight-lining detection and response-time thresholds were designed to catch a bored human clicking through. They were not designed to catch a model that reads the question.
The tell you can still see
Implausible-combination questions still work, because a bogus respondent's incentive is to say yes and move on. Pew Research Center demonstrated this cleanly in its 2024 analysis of online opt-in polls: asked whether they held a licence to operate a nuclear submarine, 12% of opt-in respondents under 30 said yes, against a true population rate that rounds to zero. Among opt-in cases identifying as Hispanic, 24% claimed the licence, against 2% of non-Hispanic cases.
The downstream damage is the point. In the same comparison, an opt-in sample put 20% of adults under 30 as agreeing that the Holocaust is a myth, where a probability-based panel put it at 3%. That gap is not a rounding error. At that scale it is the difference between a product decision and its opposite.
Five survey types and what each can honestly answer
Research surveys answer point-in-time questions. For the operational side of running feedback programs, from event triggers to keeping responses on the customer record, see our customer survey and feedback guide.
Pick the type from the decision you need to make, not from the template gallery. Each one has a frame it needs and a way it goes wrong.
| Survey type | The question it answers | Frame it needs | How it usually breaks |
|---|---|---|---|
Problem discovery | Which pains are frequent and expensive enough to build for | Your own users plus recent churned accounts | Leading questions turn a mild annoyance into a roadmap item |
Survey typeProblem discovery The question it answersWhich pains are frequent and expensive enough to build for Frame it needsYour own users plus recent churned accounts How it usually breaksLeading questions turn a mild annoyance into a roadmap item | |||
Willingness to pay | What a segment says a price change would do | Buyers with real budget authority | Stated intent overstates actual purchase behaviour |
Survey typeWillingness to pay The question it answersWhat a segment says a price change would do Frame it needsBuyers with real budget authority How it usually breaksStated intent overstates actual purchase behaviour | |||
Concept and message test | Which positioning a target segment understands fastest | A screened segment, not your newsletter | Fans of the brand rate every concept generously |
Survey typeConcept and message test The question it answersWhich positioning a target segment understands fastest Frame it needsA screened segment, not your newsletter How it usually breaksFans of the brand rate every concept generously | |||
Post-purchase and onboarding | Where new customers stall in their first weeks | Recent signups, triggered by event | Sent too late, so the memory is already reconstructed |
Survey typePost-purchase and onboarding The question it answersWhere new customers stall in their first weeks Frame it needsRecent signups, triggered by event How it usually breaksSent too late, so the memory is already reconstructed | |||
Market sizing and landscape | How a category behaves outside your customer base | A bought panel, because you have no other frame | This is where fraud rates hit hardest |
Survey typeMarket sizing and landscape The question it answersHow a category behaves outside your customer base Frame it needsA bought panel, because you have no other frame How it usually breaksThis is where fraud rates hit hardest | |||
The last row is the honest one. Market sizing is where you genuinely cannot use your own list, so you pay a panel and accept the risk. Everything above it runs against people you already know.
Step 1: Pick a frame you can authenticate
Rank your options by how hard they are to fake, not by how fast they fill:
- Authenticated in-product prompt. The respondent is already logged in, so identity comes from the session rather than a self-report.
- Personalised email or SMS invitation to a named contact. One link per person, no public URL, and you can see who opened and who finished.
- Customer list with an open link. Fine for volume, weak once the link is shared or scraped.
- Social or community post. The recruitment channel behind the worst fraud figures in NORC's review.
- Bought panel. Necessary for non-customers, and the reason your screener has to earn its place.
Sending to named recipients is the practical version of this for most B2B teams. meetergo's routing forms send a form directly to a person by email or SMS and track opens and completions per recipient, so a response is tied to a contact record rather than to whoever found the link. That does not make fraud impossible. It makes it uneconomic, which is the actual goal.
Skip this section if you're sizing a market. When the question is about people who have never bought from you, none of this applies and you need a panel. Budget for the cleaning work instead, and read the provider's own fraud-detection documentation before you sign.
Step 2: Write a screener that costs a bot more than it earns
A screener has one job: make a fraudulent completion expensive. Three tactics still hold up.
Ask for something only a real user knows. Not a password, an artefact. Which plan are you on, what does your admin screen call the settings tab, how many seats do you have. A model can guess a plausible answer, and you can check it against your own records.
Include one implausible-combination item. Borrow Pew's design: a low-frequency attribute that costs nothing to admit and almost nobody genuinely holds. It works because the fastest route through a survey is to agree with everything, which is how Pew's methodologists describe the behaviour. Anyone claiming the attribute goes in the review pile.
Put an open-ended item early, and make it specific. Generic prompts invite generic answers, which is exactly what a model produces well. Ask about the last time the thing happened, and you get either a concrete situation or an obvious paraphrase of your own question.
Two things not to do. Don't rely on attention checks alone, for the reason Westwood documented. And be careful with incentives: NORC's review notes fraud is most acute when a survey carries a financial reward, so a prize draw restricted to verified customers beats an open per-response payment.
Step 3: Ask less than you want to
Every extra question buys a lower completion rate and a worse tail. A survey that takes four minutes gets finished by people mildly annoyed with you. One that takes fifteen gets finished by people who are furious, plus whoever is being paid.
Practical limits that hold up in B2B:
- One decision per survey. If you cannot name the decision the results will change, the survey is a fishing trip.
- Six to ten questions for a customer list, three to five for an in-product prompt.
- One open-ended question, placed where the respondent still has energy for it, not stapled to the end as an afterthought. Ask for a word or a short sentence rather than a paragraph, because the size of the ask drives the skip rate.
- Ask about behaviour that already happened, not intentions. What did you do the last time, not what would you do.
Multi-step forms help here more than people expect, because a short first page commits the respondent before the harder questions arrive. meetergo's AI form and funnel builder breaks a survey across several steps with one shared theme, and covers 20+ field types including rating and NPS scales, so a mixed methodology fits in one form rather than three tools. If you're still choosing where to build, our comparisons of Typeform and Google Forms cover which parts of a research workflow each one handles.
Step 4: Book the follow-up interview inside the survey
This is the step most teams skip, and it is where the survey stops being a number and starts being a reason.
A survey tells you that 38% of respondents stalled during setup. It cannot tell you why, and by the time you have exported the results, picked interviewees and negotiated a slot, you are two weeks past the moment they remembered clearly. Put the calendar in the form instead, so a respondent who ticks the right answer sees available times on the last screen.
meetergo's forms branch on answers and assign points to choices, then route the respondent onward, so the qualifying logic and the booking live in one flow. Round robin spreads those calls across your researchers instead of stacking them on whoever owns the survey, and workflows sends the reminders that keep booked interviews from evaporating. The structure in our discovery call question template transfers almost directly to a research interview.
The honest limitation: meetergo collects, routes and books. It does not recruit respondents for you, and it does not run significance testing or crosstabs. For market sizing against a panel you will still buy sample elsewhere and analyse in a stats tool. For research against your own customers, the collect-to-interview path is the part that usually leaks.
Step 5: Clean before you look at averages
Look at the distribution of behaviour before you look at the distribution of answers. Four checks, in the order they cost you time:
- Duration outliers. Flag anything under a third of your median completion time. Do not auto-delete, review.
- Duplicate signals. Same contact, same IP block, near-identical open-text answers across submissions.
- The implausible-combination item from your screener.
- Open-text homogeneity. Fluent, well-structured, oddly balanced answers that never name a specific feature are the current signature of a generated response. Work on AI use in open-ended answers points the same way: model-assisted text clusters together and loses the idiosyncrasy that makes verbatims worth reading.
Then report what you removed. A finding that says 210 responses, 34 excluded, criteria listed, survives a hostile stakeholder question. A finding that says 244 responses does not.
Where GDPR bites in customer research
Two things make research surveys different from a marketing form, and both are easy to get wrong.
The first is purpose limitation. Customer data you hold to deliver a service is not automatically available for research, and a research response is not automatically available for sales follow-up. A respondent who books a call from the survey agreed to that. Feeding the same person into a nurture sequence because they answered is a different purpose, and needs its own basis.
The second is that incentives make respondents identifiable. The moment you record who received a voucher, an anonymous survey has a name attached to it, which changes your retention and access obligations. Decide before launch, not after.
Where the data sits matters for the usual reason: a US-parented processor carries CLOUD Act exposure whichever region the storage bucket is in. meetergo keeps form data on European servers in Frankfurt with no US corporate parent behind it, which is the same reasoning behind our writeup on GDPR-compliant software selection and the detail on the security page.
Mistakes that quietly wreck a read
- Surveying only your happiest users. In-product prompts reach people who are still logged in. The customers who left hold the answer, and reaching them takes an email to a list you exported before they churned.
- Treating stated price sensitivity as a forecast. Willingness-to-pay answers are directionally useful and numerically optimistic. Rank options with them, don't set a number.
- Running the same tracker until it stops moving. Quarterly beats monthly for anything strategic. A monthly cadence trains your customer base to ignore you.
- Reporting averages across a contaminated sample. If you skipped Step 5, every percentage in the deck inherits the fraud rate underneath it.
Try the collect-and-book flow on your own list
Set up one short customer research survey, send it to fifty named customers rather than posting a public link, and let the qualifying answers open a calendar on the final screen. You'll learn more from the twelve conversations that produces than from a thousand panel responses. Start free with meetergo, no credit card required, and check what sits on each tier on the pricing page.
Booking + video conferencing in one tool.
Booking + video conferencing in one tool.
FAQs
How many responses does a customer research survey need?
For directional decisions inside a known customer base, 60 to 100 clean responses per segment is usually enough to see a pattern worth acting on. Whether the responses are real matters more than how many there are: fifty verified customers beat 500 panel completions for anything about your own product.
Should you pay respondents?
For your own customers, prefer a prize draw restricted to verified accounts, or an early look at what you build. NORC's review flags financial incentives as the condition under which fraud is most acute, so a per-response payout on an open link is the worst combination available. Panels are different: paying is how the sample exists.
Can you use AI to analyse open-ended answers?
For clustering and first-pass tagging, yes, and it saves real hours. Read a random sample of the raw text yourself before you trust the clusters. A model summarising generated answers produces something that reads well and means nothing, and the summary layer is where that becomes invisible.
How often should you run customer research surveys?
Tie the cadence to a decision, not the calendar. Event-triggered surveys after onboarding or churn can run continuously because each respondent only sees one. Broad tracking surveys work quarterly at most, and every additional send lowers the response rate of the next one.
Are NPS and satisfaction surveys customer research?
They are measurement, not research. A score tells you the direction of travel and nothing about the cause, which is why the follow-up question and the interview matter more than the number. If you collect ratings publicly as well, the reviews app covers that separate job.
Do you need consent to survey your existing customers?
It depends on your jurisdiction, your existing basis for holding the data, and whether the survey is genuinely research or marketing wearing a lab coat. Research on an existing customer relationship is often handled under legitimate interest in the EU, but the assessment is yours to document, and worth running past whoever owns privacy at your company first.
