Customer Research from Reviews and Support Tickets

Why your customers' words

There is peer-reviewed measurement for this, and it is older than you would guess. A 2013 study analysed 18,682 customer reviews across 591 books on Amazon over seventeen consecutive weeks, using text mining and a panel-data model.

It has three findings you can use directly.

• A matching language style raises conversion. When the language style of the reviews moves closer to that of the prospect group, the conversion rate goes up measurably (coefficient 0.004, p < 0.001). A small number, but statistically strong — and it is free.
• Negative content hits harder than positive. A change in negative sentiment has a larger effect (0.067) than a change in positive sentiment (0.055). Which makes complaints your most informative raw material.
• Moderate positive beats extreme positive. The positive effect "flattens out": past a certain point, gushing adds nothing.

That is why the answer to "what should the headline say" is not a creative brainstorm but the body of text you already own.

Where the raw material is

In order of value, and the order is not an accident:

1. Support tickets. Here the customer is not performing, just describing a problem. This is the cleanest source.
2. Sales calls and emails. The exact wording of the objections is worth gold.
3. Reviews and feedback forms on your own site. Yours, on your own legal basis.
4. Chat and messaging logs.
5. Reviews on third-party platforms. This one comes last, because it is the most restricted — see below.

What you find in practice: at most companies the first four hold more usable material than the team can process. Scraping third-party platforms is a bad idea not because it is difficult, but because it is usually needless risk for something you already have in house.

The workflow

1. EXPORT (human)
   3-12 months of tickets/reviews, as raw text.
   Strip out: name, email, phone, address, order number.
   If the system cannot do an anonymised export, run a
   search-and-replace over it - this step cannot be skipped.

2. CLUSTERING (machine)
   "Group these texts by WHAT OBSTACLE the customer ran
    into. Do NOT use pre-invented categories, let them come
    out of the text. For each group: how many, and two
    verbatim quotes. Whatever fits nowhere: 'mixed'."

3. VOCABULARY (machine)
   "Collect the words customers use themselves for the
    product, the problem and the result. Per group, the 10
    most frequent phrases, with counts. Do NOT rewrite them."
   -> THIS is the most valuable output. The headline comes from it.

4. CONTRADICTIONS (machine)
   "Where do two groups contradict each other? Where is the
    same thing praised that others complain about?"
   -> This is where the segments come from.

5. CHECK (human)
   Take 20 random texts and see whether each landed in the right group.
   If not, the clustering is wrong, not the text.

6. FEEDING IT BACK
   Vocabulary -> headline and ad copy
   Obstacles -> FAQ and objection handling
   Contradictions -> segments

Step five is the one most people skip. The model always hands you categories, even when there is no structure in them. Twenty re-checked texts tell you whether there is.

What the law sets out

There are two separate questions here, and they get mixed up constantly: whose data it is and whose content it is.

Reusing your own data

You collected the support tickets in order to handle complaints. Using them for marketing research is a change of purpose, and for that the data protection regulation prescribes a compatibility assessment. This is not a prohibition: direct marketing may rest on legitimate interest. But there are two things you cannot get out of.

The first: the three-step legitimate interest test has to be carried out, and written down. According to the data protection board's opinion from late 2024, that means the interest is lawful, real and present; that there is no less intrusive option; and that the balancing may also come out in the data subject's favour. The same opinion states separately that transparency on its own is not enough for the data subject's "reasonable expectation" to exist.

The second: the simplest and most effective measure is anonymisation before the analysis. If the text carries no identifier, most of the risk evaporates. Do that before you send anything anywhere.

And if you send the text to an outside provider: a data processing agreement is mandatory, in writing, with the content the regulation prescribes. For a transfer to a third country you also have to look at what legal basis it runs on.

Scraping someone else's content

This part got harder in 2026.

• The terms of use of Google's maps service expressly prohibit extracting content, and among the examples they name the copying and saving of user reviews.
• Trustpilot's March 2026 policy prohibits the use of automated tools — naming AI agents, bots and crawlers explicitly — for data collection, the commercial exploitation of platform data, the training of AI models on it, and the republication of scraped data.
• Terms are not just recommendations. In a 2015 judgment the Court of Justice of the European Union held that where a database enjoys neither copyright nor sui generis protection, the database directive does not preclude a contractual restriction on use — meaning terms of use are enforceable as a contract.
• The text and data mining exception does exist in Hungarian copyright law, but for content made available online the rightholder may reserve their objection in machine-readable form — and once they have, the exception is not available.
• The data protection board's guidelines, adopted in July 2026 for public consultation, on web scraping for generative AI say this: sites that object unambiguously by technical means — robots.txt, ai.txt, CAPTCHA — have to be excluded from collection, because there the data subjects cannot reasonably expect it.

The practical summary: your own channels are free, third-party platforms are not. If you do need outside reviews, the official API or the platform's own export is the way, not a scraper.

What to use the output for

• Headline and ad copy. In the customer's words, not yours. On the limits of testing variants, the sample size behind landing page variants post gives you concrete numbers.
• Segments. Not by demographics but by obstacle and intent. This is what the segmentation of newsletters post also calls the single most useful AI application.
• FAQ and objection handling. The twenty most frequent questions are in the data already, word for word.
• Product decisions. A recurring complaint is not a marketing task.

Where you do have to be careful: the fact that you know something about a customer does not mean you have to show that you know it. That is what the limit of personalization post is about.

Common misconceptions

• "Looking at the positive reviews is enough." The measurement says negative content has the stronger effect. The complaint is the better raw material.
• "A review is public, so it is free to use." Being public is not a licence. Platform terms are enforceable as a contract.
• "We will anonymise afterwards." Anonymisation protects you if it happens before the analysis.
• "The machine will find the pattern." The machine always hands you categories. Whether there is structure in them is what twenty re-checked texts tell you.
• "Let's rewrite the customer's sentences so they read better." Originality is precisely where the value is. A rewritten customer sentence is your own marketing copy again.

What to take away

The raw material for the best marketing copy is already there, in tickets and reviews. What the machine helps with is turning thousands of free-text pieces into categories and a vocabulary — but the categories have to be re-checked, the texts have to be anonymised before the analysis, and the content of third-party platforms must not be scraped. Your customers' words are not something to invent; they are something to find.