How to Detect Fake Survey Respondents in Research

Review suspicious survey responses using available metadata, a documented worksheet and sensitivity checks. Separate review flags from evidence of fraud.

Review suspected fake survey responses by combining available evidence, checking alternative explanations and recording a decision for each case. A fast response, shared IP address or polished paragraph is a reason to investigate, not proof of a bot or dishonest participant.

This guide covers manual review of an exported dataset. For controls that must exist before responses are collected, use the separate survey collection and bot-control workflow.

Start with the data you actually have

Keep an access-controlled original export and create a review copy with stable response IDs. List the fields available, what each field measures and which checks your dataset can support.

For example, the Google Forms API response schema includes response IDs, submission timestamps, answers and email if collected. It does not expose respondent IP addresses, browser fingerprints or device IDs. Its submission timestamp is not a survey start time, so it cannot by itself establish completion duration. Additional instrumentation would be a separate implementation.

Other products and plans expose different fields. Inspect their current documentation and a sample export before planning a check. Treat a missing duration or device field as unavailable, not as zero or as evidence of suspicious behaviour.

Separate eligibility, response quality and fraud

An ineligible participant, a misunderstood question, an interrupted session and deliberate fabrication require different interpretations. A dataset may need an exclusion without establishing the respondent’s intent.

Pew Research Center’s study of bogus online responses found differences across the sample sources it tested and explains the difficulty of distinguishing bots from people answering carelessly. Its findings are tied to that study’s population, recruitment and field period; they are not a universal contamination rate for your UX survey.

Define eligibility and review rules before interpreting the substantive findings. Preserve exceptions and reasons when a rule proves inappropriate. Use behavioural screening criteria and review whether question wording itself caused confusion.

Use a signal as a review question

Signal, if available	What to investigate	Plausible alternative explanation
Short recorded duration	Whether the person saw the relevant questions and followed the route	Short branching path, familiarity or a timing implementation error
Identical matrix answers	Whether agreement across those particular statements is implausible	The person genuinely holds the same view across related items
Repeated or generic open text	Relevance to the question and consistency with the rest of the account	A short but valid answer, copied prompt wording or language support
Conflicting role or experience answers	Whether the questions refer to the same event and period	Multiple responsibilities, a role change or ambiguous wording
Shared IP or device signal	Whether independent records establish a duplicate submission	Shared workplace, network, household or equipment
Repeated recruitment identifier	Whether the identifier is intended to represent a single participation	A technical retry, resumption or an integration error

Compare duration against the relevant survey route and pilot conditions. Do not apply a universal “half the median” cutoff. An unusually long session may simply have been interrupted. Neither grammatical fluency nor an AI-detector label establishes authenticity.

Read open-ended answers in context and inspect apparent contradictions before excluding them. The qualitative analysis workflow can help organise that review. Several correlated signals may reflect the same underlying cause, so adding flags together does not automatically produce a validated fraud probability.

Keep a review log with defensible decisions

Download the survey response review worksheet. Its filled rows are hypothetical examples, not records from a client study. Replace them with your own response IDs and restricted evidence references; avoid copying unnecessary participant information into the log.

Example	Evidence and verification	Decision in this example	Reason
example-001	Short route and identical matrix answers; reviewer checks that the answers are relevant and mutually compatible	Retain	The flags have an adequate explanation
example-002	Same single-participation invitation appears twice; collection log confirms one completed response was exported twice	Remove the duplicate copy from analysis	This is a verified processing duplicate, not proof of respondent fraud
example-003	Role answers conflict; wording may refer to different periods and clarification is unavailable	Mark unresolved	Evidence does not justify an authenticity claim

Keep the review signal separate from the final decision. Record the rule version, reviewer, evidence checked, reason and whether the response enters the primary analysis. Give difficult cases a second review where practical. A reviewer should be able to understand an exclusion without knowing whether it helps the expected conclusion.

Check what exclusions change

Report received responses, verified duplicates, eligibility exclusions, quality exclusions, unresolved cases and the final analytical sample. Use non-overlapping categories or explain overlaps so the counts reconcile.

Compare coverage of the intended roles and situations before and after cleaning. For unresolved records, examine whether reasonable inclusion and exclusion choices change the finding. Describe that sensitivity alongside the primary result instead of hiding uncertainty behind a single clean total.

Weighting is not a repair for fabricated answers. A weighting plan can adjust for specified sampling imbalances under assumptions; it cannot make an invented account truthful. Do not retain a known-invalid response just to preserve a demographic quota.

Decide whether additional fieldwork is needed

The decision depends on which groups were affected, what evidence remains and the precision or coverage required by the research question. A small number of exclusions concentrated in a critical subgroup may matter more than a larger number spread across the sample.

There is no universal contamination percentage that automatically requires re-fielding, and no guaranteed time in which a spreadsheet review catches most fraud. Document why the remaining evidence is adequate, why a conclusion needs qualification, or which recruitment gap must be filled. Carry that decision into your research operations record.

Glasgow Research helps B2B SaaS teams turn research into product decisions. Discuss your study.