Fieldwork metrics: the 12 numbers a survey produces, and which ones you can actually compute

What it isThe share of the respondents a supplier routed to a study that its own quality layer terminated before the survey existed to them: duplicates, technical fraud (anti-detect browsers, proxies, automation), blocklisted or hyperactive accounts, failed traps in the pre-screener. Pre-screen terminates are not part of it: they are a fit decision, not a quality one, and they surface only in the supplier's incidence.

What it tells youWhether the supplier pre-cleans its traffic at all, and how hard. A high rate can mean a careful supplier, an aggressive one, or a supply defective enough to need it; the rate alone cannot say which. A rate near zero can mean clean traffic or no checks.

CaveatsOnly the supplier can compute it, and it appears in no standard report. A catch rate is not an accuracy rate: nobody measures the false positives, the honest people caught by mistake. It is not comparable across suppliers, because each runs its own fraud vendor, thresholds and upstream filtering. When the supplier hosts, its gate is the only pre-survey gate; when the buyer hosts, a second gate follows at the link.

What it isThe share of entrants at the survey link terminated before the first question by the survey side's own quality checks, usually a fraud-detection tool plugged into the link: duplicates, including the same person arriving through two suppliers, technical fraud, blocklisted or hyperactive accounts.

What it tells youWhat the survey side's checks flag among traffic the supplier already passed. It is not a verdict on the supplier: the two systems see different signals, the supplier holding identity, account and IP history across many studies, the tool at the link seeing a device and a session, and they run different thresholds and disagree. It is the one number a buyer can compute alone about the supply side. A rate that climbs across waves, or differs sharply between suppliers on the same study, is a question worth asking, not a conclusion.

CaveatsIt exists only when the buyer hosts the questionnaire: the survey link is the buyer's door, and its checks run there. When the supplier hosts, there is no link and no second gate. Some reports call it in-survey fraud removal, because the termination is often executed after the first question; in this mapping a removal that rests on fraud or technical signals is pre-survey whenever it is executed, and in-survey is reserved for removals that rest on the answers. The supplier is not blind to it: most platforms return these people on a fraud or duplicate exit redirect, so it sees the count without the reason. Duplicates across suppliers are the one category only this gate can catch, since neither supplier knows the other is in the study.

What it isThe share of survey starts terminated by a quality check on the answers given during the survey, in the screener or in the questionnaire: speeding, straight-lining, failed red herrings or low-incidence traps, contradictions with stored profile data. A quality terminate is its own disposition, returned on a quality exit redirect. Removals that rest on fraud or technical signals belong to the pre-survey rates in this mapping, even when the platform executes them during the questionnaire.

What it tells youHow much of the traffic that started failed the survey's own quality checks. Catching it here is cheaper than after completion: a terminate costs a click; a bad complete costs an incentive, a replacement and a re-field.

CaveatsA high rate says something about the check before it says something about the people: a check that a large share of a sample fails may be badly calibrated. The split by screener and questionnaire, and by check, is more useful than the total. Some benchmarks divide by all records at the door rather than by starts, and read lower on the same data.

What it isThe share of completes judged bad quality on review and removed before delivery: gibberish or AI-written open ends, straight-lined grids, internal contradictions, duplicates surfacing in the data file.

What it tells youThe clean-up bill: bad quality that survived every earlier gate, by the reviewer's own standard. Every point of it was already paid for once, and each removal restarts a journey at the far-left edge of the map, through the same chain that produced the problem.

CaveatsWho reviews depends on who hosts: the buyer when it hosts, the supplier when it does. When the buyer hosts, this review is also the reconciliation the supplier receives. Completes and delivered completes are different numbers; agree which one the CPI is priced on before fieldwork. B2B studies run far higher here than consumer studies.

What it isThe share of completes delivered to the buyer that the buyer identifies as bad quality and refuses to pay for: IDs it does not recognize, duplicates against its own records, completes its review threw out.

What it tells youWhere the two sides' counts meet for the first time, and how far apart they are. It is the number the CPI is finally settled on.

CaveatsA low rate is not a clean sample. It says the buyer found little to reject in what it received, which depends as much on the cleaning that ran before delivery as on the traffic: a supplier that removed 60% of its completes before delivering will show a low reconciliation rate on a sample that was anything but clean. Read it with the post-survey removal rate. When the buyer hosts, reconciliation is the whole post-survey review: the same completes, seen from the supplier's seat as completes it will not be paid for. When the supplier hosts, it is the buyer's second review, of a file it did not count itself. Agree the CPI base and the dispute rule before fieldwork, not in the post-field email.

What it isEvery removal on quality grounds inside your window, over everyone who entered it. For the buyer the window opens at the link: the checks at the link, the in-survey terminates and the post-survey removals. For the supplier it opens at routing and adds its own pre-survey removals.

What it tells youHow much of the traffic aimed at a study was bad, wherever it was caught. The single number for the quality problem, before any question of who caught it.

CaveatsBoth sides can compute it, on different windows, so the two seats read different values on the same study. Suppliers usually report the survey window, because including their own gate opens conversations they do not necessarily want to have. A catch rate says nothing about accuracy, and one supplier's rate is not comparable to another's: different vendors, thresholds and upstream filtering.

What it isGood-quality completes delivered, over the bad-quality removals in your window: usable completes per bad entry caught. The window opens at the link for the buyer and at routing for the supplier, as for the bad quality rate.

What it tells youHow much bad traffic a study, or a source, made you absorb per usable complete. Computed the same way across sample sources on the same study, it is a direct way to compare them.

CaveatsThe two seats read different values on the same study, and below 1 is common once the pre-survey gates are counted. Ratios computed among completes only, good against bad among what was delivered, are a different measure with a different denominator; do not compare the two.

What it isThe share of screened people who fit the target. By definition, qualified ÷ (qualified + terminates), over-quotas and drop-offs excluded. The qualified point is rarely logged, so in practice most reports use completes ÷ (completes + terminates), which is what this page computes.

What it tells youHow well the sample that reached the screener matched the criteria, and therefore how much traffic a study will consume per complete. It drives feasibility, pricing and timing more than any other number.

CaveatsIt depends on the seat: the supplier counts its own pre-screen terminates among the terminates, the buyer cannot see them. It is measured on a population someone else already filtered against a copy of the same criteria, two or three times in a layered chain, which is why two suppliers report different incidence on the same study without either being wrong. Other formulas circulate, completes ÷ starts among them. Agree the formula before comparing two quotes: the gap between formulas is usually larger than the gap between suppliers.

What it isThe share of survey starts terminated because the answers did not match the target profile.

What it tells youHow much of what reached the screener did not fit, after the supplier had already pre-screened it for the same criteria. Split by criterion, it says which question is doing the terminating, which is the one thing a supplier can act on.

CaveatsIt moves with incidence but is not its complement: starts also contain over-quotas, partials and quality terminates. In some status vocabularies a screen-out and a drop-out come back as the same terminate code, so a rate read from redirects alone is overstated by the partials hiding inside it. The supplier sees the outcome, never the question that caused it, unless the buyer sends it back.

What it isThe share of survey starts who qualified and were turned away because their cell was already full, closed mid-session, or throttled by pace controls.

What it tells youA pacing and quota-design problem, not an incidence problem: qualified respondents lost, and goodwill spent for nothing.

CaveatsKeep it out of the termination rate you track, so the pacing problem has its own line. Separate the commercial case, allocation full, a limit on one supplier's share, from the targeting case. Per cell is the useful unit.

What it isThe share of survey starts that never reached the end and never returned on any exit redirect: sessions that simply stopped.

What it tells youRead as respondent behavior, it measures the questionnaire: length against the promise, grid load, the device it arrives on. But the band mixes two populations with opposite fixes: people who left, and sessions that broke.

CaveatsOnly the party hosting the questionnaire can split the two; if you do not host, the drop-off rate you receive is a mixture. A partial in the screener means something different from a partial in the questionnaire: the first never saw your questions. Some platforms report a partial as a terminate, and some benchmarks divide by records at the door rather than starts, so agree the definition before comparing two numbers.

What it isThe share of entrants at the link that ended as a delivered complete: the whole survey window in one number.

What it tells youWhat a study costs in traffic per usable complete, from the survey side. It is the number the supplier is paid on, so it is the one both sides watch.

CaveatsThe supplier often divides by what it routed instead of what arrived, which alone explains many disagreements about how a study went. Neither entrant nor start has a standard definition, first page load or first answered question, so agree the event before comparing conversion across suppliers. It says nothing about what the chain discarded upstream to produce it.