Market research tool
Open-end word cloud and word frequency
How results are checked
Named statistical review pending. Selected automated reference cases are published on the validation page.Most mentioned word
29.1%of answers
“delivery”, then “arrive” (20.0%)
Subscribers mention “subscription” in 28.1% of answers vs 4.7% of other customers (z = 7.15)
A word cloud of open-ended answers shows the words respondents use most, drawn larger the more answers mention them. For survey reporting, the useful number is the share of answers that mention a word — each answer counted once, however often it repeats the word — read next to the answers themselves. Comparing that share between segments shows which words set a group apart.
Analyse your open-ended answers
Paste answers one per line, or open a survey file and pick the open-ended column, an optional segment column and an optional weight. Everything runs in this browser tab — free, no account, nothing uploaded. "Use sample data" loads 480 synthetic answers to "What one thing would most improve our food delivery service?" from new customers, regular customers and subscribers.
CSV, Excel, SPSS, Stata, SAS, R, Parquet, SQLite or JSON. The file is read in this browser tab and never uploaded. For Triple-S, select the .sss or .xml metadata file and its data file together.
How words are counted
Answers are split into words with the browser's built-in word segmenter (the Unicode word-boundary rules plus dictionaries for Chinese, Japanese and Thai), so text without spaces is split into words rather than counted as one long string. Words are lowercased and Unicode-normalised; punctuation is dropped and numbers can be dropped too.
Stopwords — "the", "is", "的", "です" — are removed with compact standard lists for English, Chinese, Japanese, Korean, Spanish, French, German, Portuguese and Italian. In automatic mode each answer gets its own list: by script for Chinese, Japanese and Korean, and by whichever Latin-script list matches most of its words otherwise. English plurals can be folded (deliveries → delivery); no other stemming or lemmatisation is applied, so Korean particles stay attached (배달이, 배달은) and German compounds stay whole. Two- and three-word phrases are built from the remaining words within a clause.
Word clouds summarise vocabulary, not meaning: "not late" and "late" both count as "late". Click any word to read the answers behind it before you report it, and turn recurring topics into a codeframe with open-end coding.
The method
Formula
% of answers mentioning w = Σ weight of answers containing w ÷ Σ weight of answers with text; z = (p₁ − p₂) ÷ √(p̄(1 − p̄)(1/n₁ + 1/n₂))
where
- answers containing w
- answers with at least one occurrence of the word or phrase; repeats in the same answer count once (mentions counts every occurrence)
- weight
- 1 for every answer unless a weight column is chosen; rows with a missing or non-positive weight are dropped
- p₁, p₂
- share of answers mentioning w in the segment and in all other segments combined
- p̄
- pooled share (p₁n₁ + p₂n₂) ÷ (n₁ + n₂)
- n₁, n₂
- answers in each group; with weights, Kish effective sample sizes (Σw)² ÷ Σw²
Worked example
In the synthetic sample, 464 of 480 respondents typed an answer. With the default settings (automatic English stopwords, plurals folded, words mentioned by at least 2 answers), the most mentioned word is "delivery": 135 answers (29.1%) use it, 139 times in all. "arrive" follows with 20.0% and "service" with 18.3%.
Comparing customer types, subscribers (n = 121) mention "subscription" in 28.1% of answers against 4.7% of everyone else: z = 7.15, p < .001. New customers (n = 147) stand out on "app" (20.4% vs 9.1%, z = 3.39).
Reading the answers behind "subscription" turns the count into a finding: what about it bothers that group, in their words.
Using it well
When to use it
Use it when
- Use: A first read of a large open-ended question before building a codeframe.
- Use: Spotting words that one segment, wave or market uses much more than the rest.
- Use: Short answers in several languages, including Chinese, Japanese and Korean, without uploading them anywhere.
Not the right tool when
- Avoid: Reporting meaning or sentiment: word counts ignore negation and context — code the answers instead.
- Avoid: Very few answers (under about 30): differences between words are mostly noise.
- Avoid: Long free-text documents, where topic models or careful coding are a better fit than word counts.
How to interpret it
Rule of thumb
Read the share of answers, not the size of a word: a word in 20% of answers is a common topic, one in 2% is a detail. Check the answers behind every word you quote. For segments, a ▲ means that segment mentions the word significantly more often than the others at p < .05; with dozens of words some of these arise by chance, so prefer differences that are large, make sense and survive the Benjamini–Hochberg q-values in the CSV.
Questions
Frequently asked questions
What is the difference between answers and mentions?
Answers counts responses that use the word at least once, so one respondent who repeats a word five times still counts once; the percentage is out of all answers with text. Mentions counts every occurrence. Survey reports normally quote the share of answers.
Does the word cloud work for Chinese, Japanese and Korean answers?
Yes. Words are split with the browser's Intl.Segmenter, which uses dictionaries for Chinese, Japanese and Thai, and the tool has Chinese, Japanese and Korean stopword lists. Splitting follows the browser's dictionary, so a few compound words may be split differently between browsers, and Korean particles stay attached to the word.
Why is there no sentiment score?
There is no reliable, freely licensed sentiment lexicon that works across all the languages this tool supports, and word-level sentiment misreads negation, sarcasm and survey context. Rather than show a misleading score, the tool links to open-end coding, where you define themes such as complaints and praise and review each assignment.
How are a segment's distinctive words found?
For each of the 30 most mentioned words, the share of a segment's answers that mention it is compared with the share among all other segments using a pooled two-proportion z-test. Words mentioned significantly more often (p < .05) are listed as distinctive, strongest first. With weights, shares are weighted and the test uses Kish effective sample sizes.
Is the word cloud layout random?
No. Words are placed largest first along a spiral from the centre, starting at angles drawn from a seeded generator, so the same words and seed always give the same picture. Font size is proportional to the number of answers mentioning the word. New layout changes only the seed; the table below the cloud is the accessible version of the chart.
Sources and next steps
References & further reading
- [1]Unicode Standard Annex #29: Unicode Text Segmentation. The Unicode Consortium.unicode.org
- [2]Agresti, A. (2013). Categorical Data Analysis (3rd ed.). Wiley.
- [3]Benjamini, Y., & Hochberg, Y. (1995). Controlling the false discovery rate: A practical and powerful approach to multiple testing. Journal of the Royal Statistical Society: Series B, 57(1), 289–300.
- [4]Kish, L. (1965). Survey Sampling. Wiley.
- [5]Manning, C. D., & Schütze, H. (1999). Foundations of Statistical Natural Language Processing. MIT Press.
Try it on your own data — free, no signup
Upload a CSV, Excel, SPSS, Stata, SAS, R, Parquet or JSON file. Everything runs in your browser; your file never leaves your device.
Open the workspace →