This is the list we would work through before fielding an online study today. It is sorted by where each control runs, because that decides what it costs and what it can see.

Two changes since this kind of checklist was written a few years ago. Attention checks no longer prove a respondent is human, since AI agents pass them. And reconciliation now has a shared vocabulary in the GDQ removal codes, which makes the last section worth doing properly.

Before the survey opens

These run at the entry link. A respondent stopped here never costs the buyer a complete and never touches the data.

  • Signed entry links or server-to-server entry. The 2025 Insights Association and GDQ benchmark found 75.2% of supplier records used encrypted or server-to-server links, against 91.5% for research agencies. The gap is where edited IDs get in.
  • Respondent ID deduplication. One session per supplier ID per project. Non-negotiable.
  • Device deduplication. On for most studies; off for clinics, schools, shared offices.
  • Pre-survey screening for automation, anonymised networks and multi-accounting. Run it in monitor mode first.
  • Country check using the screening provider's true location, not only the IP address.
  • Closed quotas enforced at entry, so nobody spends ten minutes on a survey that cannot count.

Inside the questionnaire

Only the survey can judge whether a person is paying attention and telling the truth.

  • One or two fair attention checks. Keep them. They catch inattentive humans. Do not treat a pass as proof of a human.
  • Knowledge questions for professional audiences. Easy for someone in the role, awkward for someone pretending. In the benchmark, general B2B studies had the highest post-survey cleanout of any study type, 18.9% of qualified completes.
  • Consistency pairs. Age asked twice in different forms, or ownership of a product that does not exist.
  • At least one open end that needs a sentence.
  • A quality endlink for anyone who fails, so they are recorded as quality terminates rather than completes or screen-outs.

What to leave out: trick questions that confuse genuine respondents, and captcha pages, which modern agents solve and mobile respondents abandon.

At the end of the session

  • Minimum length of interview. Completes faster than the threshold become quality terminates. A common rule of thumb is about a third of the median LOI from soft launch; adjust for heavy skip logic.
  • Signed endlinks or server-to-server outcomes, so a complete URL cannot be opened by hand.
  • First outcome wins, so replaying an endlink cannot turn a terminate into a complete.

In the data, before analysis

  • Straight-lining, combined with another signal. A flat line alone can be an honest answer.
  • Open ends: gibberish, copied text, and identical phrasing across different respondents.
  • Implausible combinations, such as a 19-year-old chief financial officer with 25 years of experience.
  • Clusters: many completes from one allocation in a short window with similar answer patterns.
  • Ghost completes: completes recorded by the router with no matching record in your survey data. How they happen.

After cleaning

  • Reverse with the most specific GDQ code that is true, and a note when there is a pattern.
  • Reverse within days, not months.
  • Review removals by supplier, pre-survey blocks and reversals together. A supplier with removals mostly under codes 2 and 4 has a different problem from one with removals under 6 and 13.

The short version

  1. Stop duplicates, automation and closed-quota entries at the entry link.
  2. Ask fair attention, consistency and knowledge questions.
  3. Send failures to the quality endlink.
  4. Enforce a minimum LOI and signed outcomes.
  5. Clean open ends and patterns before analysis.
  6. Reverse with GDQ codes and look at the results by source.

Questions readers ask

What is a speeder in a survey?
A respondent who finishes implausibly fast, usually defined as completing in less than a set share of the median completion time, often around one third.
Should failed attention checks be terminates or quality terminates?
Quality terminates. The respondent may have qualified; the problem is the quality of their answers, and suppliers need to see that separately.
Do attention checks still work against bots?
They still catch inattentive humans, but AI agents pass them easily. Automation is better detected before the survey, through browser, network and device signals.
Data quality across the session lifecycle Every control ResearchDart applies before, during and after the survey.