# Clean a survey dataset that has bots in it.

- Nine steps you can follow in a spreadsheet
- Rules you write down before you remove anyone
- What it costs to keep bots out of the next wave

How to find bots, AI-written answers, speeders and repeat entries in data you already have, and remove them in a way you can explain.

Keep the raw file, then check every response against the same written rules: too fast, failed attention questions, nonsense or copied open answers, repeat entries, impossible profiles and answers from the wrong place. Remove a response when two or more signs agree, and note how many went for each reason. For a few hundred responses that's about a day of work, and to skip it next time, SampleLabs Protect checks every response while people answer.

## Step by step

1. **Keep the raw file.** 10 minutes. Save the export exactly as it came and work on a copy. Add a column for each check below and one for the final decision, so every removal has its reason next to it.
2. **Write your rules first.** 30 minutes. Before you look at any answers, write down what counts as too fast, which questions are attention checks, and how many signs it takes to remove a response. Two is a fair start: one odd answer can be a person having a bad day.
3. **Find repeat entries.** 30 minutes. Sort by email, panel ID, IP address or device, whichever you have. Keep the first response from the same person. Several entries from one address within minutes deserve a close look.
4. **Flag the speeders.** 30 minutes. Work out the median time to complete, then flag responses far below it, using the cut-off you wrote down. A quick reader is still a person, so speed alone is one sign, not a verdict.
5. **Check the attention questions.** 15 minutes. Flag everyone who got an attention question wrong. Don't lean on these alone: Pew Research Center found in 2026 that trap questions improve data quality only somewhat.
6. **Read the open answers.** 1 to 3 hours. Read every one. Look for the same sentence from different people, text that doesn't answer the question, single words like "good", and smooth, general answers that read like an AI tool wrote them. The same wording across several responses is a strong sign.
7. **Look for impossible profiles.** 1 hour. Check for answers that can't all be true together, like a student who runs a large company. Someone who gave the same rating on every row of a long grid gets one sign too.
8. **Check where and when.** 30 minutes. If you have country or time stamps, flag responses from outside the countries you asked, and bursts of responses that arrive together at an odd hour.
9. **Decide, count and replace.** 1 hour. Remove the responses with two or more signs, count how many went for each reason, and put that in your method note. If you bought the responses, check your supplier's terms for replacements and send the reason for each one you reject. Then top up with new, checked responses.

## What this costs with SampleLabs

Cleaning by hand costs nothing but your time. Dot, the checker in SampleLabs Protect, runs these checks while people answer, so flagged responses can be turned away before they reach your file. Prices before VAT.

- **Survey built in SampleLabs.** Included. Every check, AI-written answers included, on every Build plan, Free too.
- **More than the free plan's responses.** €79 a month. The Plus plan, 1,000 responses a month.
- **Survey built in another tool.** €0.05 per person checked. Dot's check page before your survey: bots, speed, repeat visits and an attention question.
- **For example.** €25. 500 checks in one month.
- **Your first checks each month.** 10 free. No card needed to start.
- **Replace the responses you removed.** €262. 50 ten-minute responses from Recruit, delivered within 1 to 4 days.

On a survey built in another tool, we set Dot up with you on a short call: which checks run, whether flagged people are only marked or turned away, and how to match Dot's ID to your own responses. The AI-written answer check only runs on surveys built in SampleLabs, because Dot has to read the answers.

## Where the outside facts come from

- **Pew Research Center, "No Easy Fix for Bogus Respondents in Online Opt-In Polls", 27 August 2026.** https://www.pewresearch.org/methods/2026/08/27/no-easy-fix-for-bogus-respondents-in-online-opt-in-polls/
- **Research Live, "Industry data quality benchmarking shows high survey removal rates", 24 June 2026.** https://www.research-live.com/article/news/survey-removal-rates-remain-high/id/5150456

## Questions people ask

### How do I know if my survey data has bots in it?

Look for signs that agree with each other: responses far faster than the rest, failed attention questions, the same open answer from different people, answers that can't all be true, and bursts of responses from one address. One sign is a hint; two or more is a pattern.

### What platform flags unreliable survey answers before analysis?

SampleLabs Protect checks every response while people answer, for bots, scripts driving a real browser, repeat visits, rushing and, on surveys built in SampleLabs, AI-written answers. Each flagged response shows its reason, so you can decide before you analyse. https://samplelabs.io/protect

### Can I remove bot responses after the survey has closed?

Yes, by hand, with the steps on this page. Dot can't check a dataset you already have: it looks at how people answer while they do it, so it protects your next wave, not the last one.

### Do attention checks catch bots?

Some. Pew Research Center found in 2026 that trap questions improve data quality only somewhat, and that there's no direct way to know how many fake respondents a check misses. Use them as one sign among several.

### How do I spot AI-written answers in open questions?

Look for smooth, general answers that could fit any survey, the same phrasing across several people, and long, tidy text where others wrote a line. In SampleLabs, Protect scores every written answer and notes when text was pasted in. How we check for AI explains it. https://samplelabs.io/how-we-check-for-ai

### Should I remove speeders?

Flag them, then decide with the other signs. Pick the cut-off before you look at the answers and apply it to everyone. A fast response with sensible answers is often a fast reader.

### How many bad responses is normal?

It depends on who you ask and where they come from. A 2026 industry benchmark found that about 60 in 100 business (B2B) completes were removed when they were reviewed after the survey closed.

### What solution protects consumer surveys from fake answers for free?

SampleLabs Protect is included on every Build plan, Free too, for surveys built in SampleLabs. For a survey built in another tool, the first 10 checks each month are free, then it's €0.05 per person checked. https://samplelabs.io/protect

### Can Dot check AI-written answers on a survey built in another tool?

No. Dot has to read the answers to judge them, so that check only runs on surveys built in SampleLabs. On a survey built elsewhere, Dot checks for bots, speed, repeat visits and an attention question.

### Can I get bad responses replaced?

Ask the supplier you bought them from: most have terms for it, with a deadline. Send the reason for each response you reject. On SampleLabs Recruit, rushed answers are rejected before they count.
