Get the responses into shape first
A word cloud is only as tidy as the text you give it. For survey answers, put one response on each line and paste a single column from your export, not the whole sheet. Take out respondent numbers, timestamps and test entries. Decide what to do with empty answers such as none, nothing and no comment: the tool counts them like any other word, so a dozen of them will become a prominent word that says only that people had nothing to say. Remove them, or keep them knowingly.
Leave the question itself out. If it is copied onto every row, its words will be the biggest in the cloud and tell you nothing.
Filler words matter more in surveys than in essays
The generator’s filter removes 180 common English words. In an essay that is helpful. In survey answers it can change the meaning, because the list includes not, no, nor, very, too, more, few, off and against. An answer such as “not helpful” arrives in the cloud as helpful. A sensible routine is to generate once with the filter on to see the themes, then once with it off, then read every line that contains a negative before you report anything.
Variant words are counted separately
The tool counts exact word forms. Reply and replies are two entries, as are fast, faster and quicker, and British and American spellings. If you want them to count together, use find and replace in your text editor before you paste, for example changing every quicker to faster. Do this openly: note in your report which words you merged.
Worked example: twelve invented responses
These twelve short answers to “What is the one thing we should change?” were invented for this page; they are not real survey data.
- Faster delivery please, delivery is slow.
- The app keeps logging me out.
- Cheaper delivery. Delivery costs too much.
- Better app search. I can't find anything in the app.
- Delivery was late twice.
- Nothing, it's great!
- Support took three days to reply.
- Faster support, I waited three days.
- The app is slow on my phone.
- More delivery slots at weekends.
- Prices are fine but delivery is expensive.
- Quicker replies from support.
The twelve lines hold 70 words. After the filler-word filter, 47 remain, and they make 32 different words. These are the repeated ones:
| Word | Mentions | Responses using it |
|---|---|---|
| delivery | 7 | 5 |
| app | 4 | 3 |
| support | 3 | 3 |
| faster | 2 | 2 |
| slow | 2 | 2 |
| three | 2 | 2 |
| days | 2 | 2 |
Four things are worth noticing. First, mentions are not people: delivery is mentioned seven times but by five respondents, because two of them said it twice. Second, slow appears twice but means two different things, a slow delivery and a slow app, and the cloud cannot tell you which. Third, the only positive answer, “Nothing, it’s great!”, produces great once, which would be drawn at the smallest size; a cloud buries the one voice that disagrees with the rest. Fourth, 25 of the 32 words appear once, so twelve responses give a cloud with a few big words and a crowd of small ones. That is normal for a small survey, and it is why a count table beside the picture is more honest than the picture alone.
What a word cloud cannot show in survey data
- Feeling. It cannot tell praise from complaint. Delivery could be the thing people love or the thing they hate.
- Negation. “Not slow” and “slow” both feed the same word.
- Who said it. A word that dominates may come from one outspoken group.
- Who did not answer. People who skip open questions may differ from those who write at length.
- Number of people. As the example shows, mentions and respondents differ.
Use the cloud to find themes quickly, then count the themes yourself by reading and tagging responses. If a decision depends on a number, report that number from your tagging, not the size of a word.
Anonymity and care with people’s words
Free-text answers often hold more than their authors realise: names, job titles, places, health details, and the odd phone number. The Information Commissioner’s Office explains that personal data is any information relating to an identified or identifiable person, and that a person can be identifiable indirectly, not only by name (ICO, What is personal data?, read 21 October 2026). In a team of eight, “the new person on nights” may point to one individual.
This generator works on your device and does not send your text to us, which helps, but it does not settle whether you may analyse the answers in this way. That depends on what respondents were told when they answered and on your organisation’s own rules, so check those. The ICO also publishes guidance on anonymisation (ICO, Anonymisation: about this guidance, read 21 October 2026). As habits: remove names and identifying detail before pasting, avoid publishing a cloud built from a very small group, and do not use quotes that give a person away.
When a cloud is the wrong tool
If you have several hundred responses, need to report percentages, or must show change over time, a hand-coded set of themes in a spreadsheet will serve you better. A cloud is a good opening look, nothing more. Pair it with three or four short quotes and a table of counts, and say plainly how many responses it came from.
Related: the general guide covers choosing words, shapes and colours.
- Should the question text be in the box?
- No. Paste the answers only. If the question is copied onto every row, its words will be counted every time and will dominate the cloud.
- How many responses do I need?
- The tool has no minimum, but with a dozen responses most words appear only once. In the worked example, 25 of 32 different words appear once. Counts become more telling as more people use the same words.
- Should I switch the filler-word filter off?
- Try it both ways. When it is on, it hides words such as not, no and very, which can reverse the meaning of a short answer. When it is off, you see them, but you also see the, is and a.
- Can I paste names or email addresses?
- It is better not to. Remove them first. The checker above flags obvious email addresses, phone numbers, web addresses and @handles, but it cannot spot names.
- Can it handle ratings from 1 to 5?
- Not usefully. Numbers are counted as words, but any single digit is dropped because words under two characters are ignored. Use a spreadsheet for ratings.