Filtering happens before, during, and after completion—not in a single post-hoc pass. Every checkpoint is documented. Every decision that affects your dataset is auditable.
A single aggregate quality metric tells you nothing about where problems were caught. These three categories describe where in the process each determination happens.
Respondents who entered the study flow and were evaluated. Screening happens at entry—before the first question. The screened count includes everyone who attempted to participate, whether or not they proceeded.
Respondents or responses removed by validation checkpoints across all three stages. Filtering is continuous—it happens at entry, during the survey, and in post-completion processing. A filtered respondent is recorded with a reason code at each stage.
Responses that cleared all quality checks across all three stages and meet the study's acceptance criteria. This is what enters your dataset. The quality audit documents how each response reached acceptance—or why it didn't.
Eleven checkpoints that run before the first survey question loads
Device IP, mobile carrier country code, and Wi-Fi SSID country hash are combined. A single-signal location match doesn't pass when secondary signals conflict. Mismatches above 500km trigger automatic exclusion from the screened pool. Mismatches between 120–500km require two corroborating signals to proceed.
A continuously refreshed database of 48,000+ known VPN exit nodes, data center IP ranges, and residential proxy services is checked at entry. Residential proxies—which route through genuine home IP addresses—are tracked separately from data center VPNs, since they require different detection signals. Database is updated every 6 hours from three threat-intelligence feeds.
Device identity is derived from 94 hardware attributes—persistent across browser clearing, cookie deletion, and VPN switches. The same physical device cannot contribute more than one response per study. Cookie-based deduplication clears in 30 seconds; hardware fingerprinting doesn't.
Profiles identified as fraudulent through our own detection and through industry data-sharing agreements with three major panel networks are matched at entry. The combined exclusion list covers approximately 340,000 profiles. Matching is done against hashed identifiers—no PII is shared across the network.
Panel profiles must have been updated or activity-confirmed within 180 days to be invitation-eligible. B2B profiles go stale faster than consumer profiles—job changes, company acquisitions, and role restructuring affect profile accuracy. Stale profiles are quarantined for re-verification, not silently removed.
Completion pace is checked: profiles completing more than three studies in the past 30 days are deprioritized in quota filling; more than six in a month triggers a temporary exclusion. Overly active panel members develop habitual responding patterns that don't show up in standard quality diagnostics but affect data reliability.
Seven checkpoints that run in real-time during completion
Completion time for each question type (Likert, multi-select, open-end, ranking) is scored against type-specific benchmarks derived from 4.2M validated completions. A respondent who passes the overall time threshold but shows implausible per-question patterns is flagged. Fast readers can complete surveys quickly without triggering this check—the pattern matters more than the total time.
Scale consistency index is calculated across all Likert-type items. Very high consistency (above 0.91) is flagged—genuine respondents holding uniform views rarely hit this threshold even when they're highly aligned. Seesaw patterns (alternating between extreme positions) are scored separately. Both trigger review, not automatic exclusion, because some legitimate respondents do have very consistent views.
Attention checks are topic-relevant with verifiable correct answers—not generic "select the third option" instructions. A B2B tech study might include a question with a verifiable answer about a widely known industry fact. Failure indicates disengagement rather than misunderstanding, which is the distinction that matters for data quality.
Five checkpoints that run after submission, before delivery
All open-ended responses are embedded and scored against three measures: cosine similarity to other responses in the same batch (cluster detection), semantic relevance to the question asked, and AI-generation likelihood. Responses flagging on two of three measures are not delivered. Single-flag responses are held for manual review and delivered within four hours with a scoring annotation in the data file—they're not silently dropped.
This is the primary defense against coordinated response fraud and AI-assisted completion, both of which have become substantially more common in the past 18 months.
For studies with logically related question pairs, consistency checks run on the completed response set. A respondent who reports no purchasing authority but claims to be the primary decision-maker for technology procurement triggers a flag. These are scored rather than automatically excluded—role definitions vary across companies, and aggressive exclusion produces false positives in legitimate cases.
Consistency scoring is calibrated per study type. B2B studies have more complex logical relationships between questions than consumer studies, and the scoring thresholds reflect that.
Every delivered dataset ships with a quality audit report. You don't request it separately—it's standard. The report documents:
Industry data from the Insights Association 2024 Panel Quality Report, alongside QRC Survey's own figures from 4,200 studies conducted in 2023–2024.
| Quality Dimension | Industry Average | Top-Quartile Providers | QRC Survey |
|---|---|---|---|
| Validation stages | 1 (post-hoc only) | 2 (entry + post) | 3 (entry + in-survey + post) |
| Checkpoints per study | 4–6 | 10–14 | 23 |
| Device deduplication | Cookie-based | Device fingerprint | Hardware-level (94 attributes) |
| Open-end quality scoring | Minimum length only | Keyword filter | LLM similarity + AI-generation scoring |
| Quality audit delivery | On request (3–5d) | On request (24h) | Automatic with every dataset |
| Proportion entering dataset (B2B studies) | 88–95% of completions | 78–88% of completions | 49–78% depending on study type and market |
Source: Insights Association Panel Quality Report 2024; QRC Survey internal data (Jan–Dec 2024, n=4,200 studies). The "proportion entering dataset" figure reflects multi-stage filtering across all three validation stages—not a failure rate.
We'll send you a redacted audit from a comparable B2B study—so you know exactly what the documentation looks like before you run anything with us.