What predicts survey reversals: data from one GPT site
A survey complete is not final when it is credited. The panel behind the survey checks the answers afterwards and can reverse the complete days later, after the user has been paid. We followed every conversion one GPT site sent us in its first two weeks, 12 to 27 October 2026, with the reversals it reported up to 7 November, to see what actually predicted a reversal and what only looked like it did.
Short answer
On this site 38% of settled survey conversions were reversed, a median of 6.8 days after they were credited. The strongest predictors were the time between the offer click and the complete (74% reversed after one to two hours, 34% within ten minutes) and the country (56% in Australia, 13% in Germany). Counting conversions instead of accounts gave misleading answers, and a reversal already reported said little about the next conversion, because reports arrive in batches about a week late.
- Reversed
- 708 of 1,856 settled conversions (38%); 140 of the 314 accounts behind them had at least one.
- When
- A median of 6.8 days after the conversion; 10% within 2.7 days, 90% within 12 days.
- Slow completes
- 74% reversed when the complete came one to two hours after the click, 34% within ten minutes.
- Clean at the time
- About four in five reversed conversions came from accounts that looked clean when they converted.
The data
The site pays its users for surveys and offers from several survey walls and offerwalls, and reports every reversal back to CashoutGuard through the outcomes API: the same information a network sends when it takes a payout back. We kept only conversions old enough for their reversal to have arrived, 12 days or more, for the reason the next section explains.
That left 1,856 conversions from 314 accounts. 708 of them were reversed, 38%. Measured per account the picture is wider: 140 of the 314 accounts, 45%, had at least one reversal. Both numbers matter, and the gap between them comes back in the sections on device signals.
Reversals arrive about a week late
A reversal is reported when the panel has finished checking the answers, not when the user finishes the survey. On this site the reports came in batches, and the delay was long:
| Share of reversals reported | Time after the conversion |
|---|---|
| 10% | 2.7 days |
| 50% (median) | 6.8 days |
| 90% | 12 days |
Two consequences follow. Any report on the last week of conversions undercounts reversals, because most of them have not arrived yet. And an account can keep converting for a week before its first reversal is known. In Australia the busiest account made 49 conversions in the period and all 49 were reversed; the next ones had 30 of 30 and 24 of 24. None of them looked like that on the day they converted.
Much of that money was already gone. Of the $328 the networks reversed, $211 (64%) came from 71 users who had asked for a cashout between the conversion and its reversal: paid out on conversions that were taken back afterwards.
The time between click and complete
The clearest single predictor was how long the user took between clicking the survey and the complete arriving. Most surveys take a few minutes to half an hour, and completes credited long after the click were reversed far more often:
| Time from the click to the complete | Reversed |
|---|---|
| Under 10 minutes | 34% |
| 20 to 30 minutes | 55% |
| 30 to 60 minutes | 59% |
| 1 to 2 hours | 74% |
| All matched conversions | 37% |
The pattern held in both halves of the period, so it was not one bad week. Long gaps fit what panels reject: a survey left open while the user does something else and answered carelessly at the end, or a complete that arrived through a chain of routers the panel later refused.
What did not predict reversals is just as useful. The number of conversions an account made in a day made no difference: the first conversion of the day was reversed 38.5% of the time, the seventh or later 36.2%. On this site busy accounts were not riskier per conversion, so a daily cap would have cost honest users without saving much.
Country
Reversal rates differed by country far more than by anything the user did on the site. In the same two weeks, 306 of the 683 reversals came from Australia: 56% of Australian conversions were reversed and 72% of those from New Zealand, against 13% of German ones and 37% overall.
That is not a reason to block Australia. Panels set their quotas and quality checks per country, and a high-paying country attracts both stricter panels and more people pretending to live there. Use the country to decide how long to hold a reward and which conversions to look at first, not who may join.
Device signals: count accounts, not conversions
Device checks did predict reversals here. The strongest was a device shared with other accounts: 22 of the 30 accounts flagged for it had reversals, 73% against 45% of all accounts, and per conversion 47% were reversed against 35%.
Another signal looked like the opposite. Conversions from browsers whose language did not match the IP country were reversed only 19% of the time, half the site average, which suggested the check was useless here and its weight should come down. Counted per account it said something else: 70 of its 137 conversions came from a single account with one reversal, and 9 of the 13 accounts behind it, 69%, had reversals, well above the 45% of all accounts.
| Signal | Reversed per conversion | Accounts with a reversal | Accounts |
|---|---|---|---|
| Device shared with other accounts | 47% (site 35%) | 73% (site 45%) | 30 |
| Browser language not matching the IP country | 19% (site 38%) | 69% (site 45%) | 13 |
A few heavy accounts can carry most of a signal's conversions. Before acting on a figure, check how many different accounts it rests on. Under about 30, it says more about those accounts than about the signal.
A past reversal does not clear an account
You might expect an account that has already had a reversal reported to be reversed again. In this data it seemed to be the other way round: of 358 conversions made after the account's first reported reversal, only 5.3% were reversed, against 35.2% for all the others.
That is an artefact of the delay. The conversions made after a report are the account's newest, and their own reversals had not arrived when we looked. A reversal is evidence about the account, not a sign it has been cleaned: review the account, and look at its other conversions that are still inside the hold.
Most reversals are about the answers
About four in five reversed conversions came from accounts that looked clean when they converted. Survey panels reverse for answer quality: speeding, identical answers down a grid, failed attention checks, answers that contradict the profile. That happens inside the survey, where no device check can look.
Fraud checks remove the share that comes from farms, hidden VPNs and multi-accounts. The rest is lowered by holding rewards, by watching the wait-time pattern above, and by sending users to surveys that fit their profile. Our guide to survey fraud on GPT sites covers the patterns, and the one on offerwall chargebacks covers handling reversal postbacks in your code.
Measuring it on your own site
Every number in this guide comes from reversals the site reported to CashoutGuard. When you report yours through the outcomes API, the Accuracy page shows the same breakdowns for your site: how long your reversals take to arrive, which countries and wait times are reversed most, and which signals predict them, with the number of accounts behind each figure.
- Signals resting on fewer than 30 accounts are marked as too few to tell.
- Rules & lists turns the clear results into suggestions, such as counting completes that arrive long after the click, only when 30 or more accounts back them.
- Nothing changes until you apply a suggestion, and a weight you lowered is offered back if newer data stops supporting it.