Almost every form ends with a box. Anything else we should know. Tell us about your project. What could we do better. It is the question where people say the thing they actually came to say, in their own words, without your categories getting in the way.
It is also, in most businesses, the question nobody reads in aggregate. Individual answers get read when a submission arrives, and then they sit in a column that grows by a few hundred rows a year and is never looked at as a whole.
That is a strange thing to be true of the most information-dense field on the form, and the reason is entirely practical. Reading two hundred paragraphs and working out what they collectively say is a genuine afternoon of work, and there is always something more urgent.
Why the aggregate is different from the individual
Reading answers one at a time as they arrive tells you about each customer. It does not tell you about your customers, and those are different things.
The pattern only exists in the pile. Eleven people in three months mentioning that they could not find your pricing is a finding. Each of those eleven, read on its own on a Tuesday morning, is a mild observation you nodded at and forgot. The signal is real and it is only visible from above.
This is why businesses are routinely surprised when they finally do the exercise. The most common reaction is not learning something new. It is realising they had been told the same thing forty times and had processed it forty separate times as an isolated remark.
Eleven people saying the same thing over three months is a finding. Read one at a time, it is eleven shrugs.
The old way, and why it stalled
The traditional approach is coding: read everything, invent categories, tag each response, count the tags. It is how qualitative research has been done for decades and it works.
It also takes three to four hours for a few hundred responses, requires somebody to hold the whole set in their head while inventing categories, and has to be redone every quarter or the categories go stale. For a business with one person doing marketing, that is not a task that gets scheduled. It is a task that gets intended.
The result is not that businesses choose not to analyse their free text. It is that the cost sits just above the threshold where it happens at all, which is the worst possible place for a cost to sit.
What a model is good at here
Grouping a few hundred short pieces of text by what they are about is close to the ideal use of a language model. It is not a task with a single correct answer, it requires reading comprehension rather than reasoning, and being roughly right is genuinely useful.
Ask for themes with counts and representative quotes and you get, in under a minute, something that looks a great deal like what a careful person would produce in three hours. It will not be identical and the differences will mostly not matter, because you are looking for what to pay attention to rather than for a number to publish.
The important reframing is that this is not a report. It is a reading aid. It tells you which forty of the two hundred answers are worth reading properly, and then you read those forty yourself.
Ask for quotes, not just categories
A summary that says thirty per cent of responses concerned pricing is nearly useless, because pricing covers too much to act on. Too expensive, unclear, not shown up front and different from what I was quoted are four different problems with four different fixes.
Always ask for representative quotes with each theme, in the customer's own words, and read them. The quotes are where the actionable specificity lives, and they are also the thing that will convince anybody else in the business, because a percentage is arguable and a customer sentence is not.
It is worth asking for the outliers too. The one person who said something nobody else did is sometimes noise and is sometimes the first sign of something that will be a theme by spring.
Do not let it decide anything on its own
There are failure modes worth knowing, and they are all versions of the same thing: a summary sounds authoritative regardless of how well it reflects the source.
A model will smooth over a response it does not understand rather than flagging it. It will occasionally invent a theme that is a plausible generalisation of two unrelated answers. And it is sensitive to how you ask, so a question framed around complaints will find complaints in a set of mostly neutral answers.
The discipline that handles all of this is simple. Never act on a theme without reading the underlying responses. The summary directs attention. The decision comes from the actual words somebody wrote.
The privacy question, which is real
Free text is the field most likely to contain something personal, because people write what is on their mind rather than what the field asked for. Health details, financial circumstances, the reason they are moving house, sometimes another person's name.
Before sending a batch of free text anywhere, be clear about what is in it. For an ordinary enquiry form the answer is usually nothing sensitive, and for a feedback form on a medical service it is very much not.
If the content is sensitive, the options are to strip identifying details first, to do the analysis somewhere with appropriate handling in place, or to read it yourself. All three are reasonable. Doing it without thinking about it is the option that is not.
The questions worth asking of a pile of text
Once analysis is cheap, the useful move is to stop asking what does this say and start asking specific things you can act on.
- What do people mention that is not a question on the form? Those are the fields you are missing.
- What comes up from people who did not go on to buy, compared with those who did?
- Has anything appeared this quarter that was absent last quarter?
- Which competitors are named, and in what context?
- What are people explaining that the page should already have told them?
- Which answers suggest the person misunderstood what we sell?
The missing-field question is the best one
Of those, the first is the one that consistently pays. If people keep writing something in the open box, they are telling you it matters to them and you did not ask.
Access arrangements, parking, whether there are pets, whether the property is occupied, a deadline they are working to. Each one arrives as an aside in a paragraph and each one is something you would rather have as a structured answer you can sort and filter by.
Turning the top two or three of those into proper questions tends to shorten the free text box's contents as well, which makes the next round of reading easier. The box is a queue of questions you have not added yet.
How often to do it
Once analysis costs a few minutes rather than an afternoon, the tempting answer is constantly, and that is not quite right either. Free text moves slowly and reading it weekly produces noise and a sense that nothing is changing.
Quarterly suits most small businesses. It is long enough for patterns to form, short enough to act within the same season, and it lines up with the rhythm most people already use for looking at anything.
The exception is after a change. If you rewrote the pricing page or changed a service, read the following month's free text specifically against that change, because the effect will be visible there before it is visible anywhere else.
Write it down somewhere other than the tool
The analysis is only useful if it survives. A summary read once in a chat window and not recorded produces the same outcome as not doing it, which is a business that keeps rediscovering the same three things.
Keep a running document with the date, the themes, the counts and two or three quotes. It takes five minutes and after a year it is the single most useful document about your customers that you own, because it shows movement rather than a snapshot.
It is also the thing to read before any decision about pricing, positioning or what to build next. Most businesses make those decisions from memory and a handful of recent conversations, which over-weights whatever happened last week.
Compare against the structured answers
Free text gets more useful when it is read alongside the questions around it rather than on its own. The same paragraph means something different depending on who wrote it.
What people say in the open box differs sharply by service, by how they found you, and by whether they went on to buy. Splitting the pile before summarising it is often where the actual insight is. A theme that appears in a fifth of all responses and in two thirds of the ones that did not convert has just told you where you are losing work.
This is the one analysis that is genuinely hard to do by reading, because holding two hundred paragraphs and their associated categories in your head at once is beyond most people. It is also the one where a cheap summary pays for itself most obviously.
Ask a better open question
Before improving how you read the answers, it is worth improving the question, because the standard one is weak. Anything else we should know produces mostly nothing, some logistics, and occasionally something valuable.
Specific prompts produce far better text. What are you hoping this will solve. What have you tried already. What would make this a success. Each of these gets people writing about the thing you actually want to understand, and the answers are longer and more useful without the question feeling heavier.
Two boxes are almost never worth it. If you want two kinds of information, pick the one you will act on, because a second paragraph field roughly halves the quality of the first.
Watch what gets left blank
The absence of an answer is data and it is invisible in any summary of the answers that exist. If the open box is optional and eighty per cent of people skip it, that is worth knowing before you draw conclusions from the twenty who did not.
People who write in an optional box are self-selected, and they skew towards the unusually pleased, the unusually annoyed, and those with a complication. That is not a reason to distrust the text. It is a reason to be careful about phrases like most customers, which the summary will happily produce from a sample of the motivated.
If the blank rate is very high and you genuinely need the information, the fix is usually a more specific question rather than making the box required, which produces a lot of answers reading not applicable.
Do it once this week
If you have never done this, the bar is genuinely low. Export the free text column from your main form for the last three months. Ask for themes with counts and quotes. Read the quotes.
Expect to find one thing you did not know, one thing you knew and had been ignoring, and one question you should add to the form. That is a typical result and it is worth considerably more than the fifteen minutes it takes.
Then put a note in the calendar for three months' time, because the value is in the comparison and the comparison needs a second reading to exist at all.