Every buyer who has fielded online sample has sent a reconciliation file that looked like this: a column of respondent IDs, and a second column that says "quality", "bad OE", "speeder?" or nothing at all. Every supplier has received one and had no way to act on it beyond removing the incentive.
That file is where most of the value of data cleaning is lost. The buyer did the work of finding bad responses. The supplier cannot tell which sub-source, which recruitment channel or which fraud pattern produced them, so the same traffic comes back on the next study.
What the GDQ code frame is
The Global Data Quality initiative is a collaboration of research associations including the Insights Association, ESOMAR, MRS, CRIC and The Research Society. In January 2026 it introduced a feedback loop code frame: 18 standard reasons a respondent can be removed for quality, each with a short definition.
The codes run from bot detection (1) and third-party fraud tool failure (2), through geo-location, duplication and entry time checks, to in-survey behaviour such as speeding (6), straight-lining (8), trap questions (10) and knowledge questions (11), open-end problems (13 to 15), duplicate closed answers (16), ghost completes (17), and a catch-all (18). The full list with definitions is on our reconciliation page.
The value is not the list. The value is that a code means the same thing to the buyer who sent it and the supplier who received it.
Which codes get misused
Standard codes only help if people pick them honestly. From reading the definitions against how removals are usually described, these are the ones we expect to cause trouble.
18, other. It will become the new "quality". If more than a small share of your reversals are code 18, the reason is almost always one of the other 17 and someone did not want to choose. Treat a high code 18 rate as a reporting problem.
15, open end AI completed. The definition says the open end "appears to be" copied from AI-generated text. Detector scores are not evidence. Use code 15 when a reviewer can point to something concrete, such as the same unusual phrasing across several respondents, or a response that restates the question in a chatbot's register. When in doubt, code 13 (poor quality) is honest and still actionable.
6 versus 7. Speeding is too fast. Excessive interview length is too slow, which can indicate someone running several surveys in parallel or a session left open. Both are real; they point to different problems.
4 versus 16. Participant duplication (4) is based on an identifier such as an IP address or cookie. Duplicate closed responses (16) is when the answers themselves suggest the same person. A supplier fixes these in different places, so keep them apart.
8, straight-lining. As Cint pointed out when the code frame was introduced, similar answers across a grid can be perfectly logical. A respondent who genuinely rates every brand as unfamiliar will produce a straight line. Remove on straight-lining combined with another signal, not alone.
How to send reversals suppliers can use
- One code per removal, the most specific one that is true. If a response failed a trap question and also sped, choose the check that decided the removal.
- A short note when the pattern matters. "OE2 identical across 11 respondents from the same allocation" helps a supplier find a sub-source. "Bad" helps nobody.
- Reverse promptly. A reversal three months after fieldwork arrives after the supplier has paid its own sources.
- Look at your codes by supplier. If one supplier's removals are mostly codes 2 and 4 and another's are mostly 6 and 13, they have different problems and need different conversations.
Pre-survey and post-survey in one vocabulary
The code frame is useful before the survey too. When ResearchDart stops a session at the entry link, the router reason maps to a GDQ code: duplicate devices and respondent IDs to 4, location mismatches to 3, automation to 1, anonymised networks and provider risk decisions to 2. Speeders caught at the endlink map to 6.
That puts everything a supplier's traffic cost you in one report, whether it was blocked before the questionnaire or reversed after cleaning. In our console this is the quality by source report.
What this will not fix
A shared vocabulary does not make suppliers act on feedback. It removes the excuse that the feedback was unusable. Whether a supplier drops a bad sub-source is still a commercial decision, and some will not. The codes make that decision visible.
It also does not settle disputes about whether a removal was fair. A supplier can reasonably challenge a code 15 based on a detector score. That is a better argument to have than one about a spreadsheet cell that says "quality".