What do customers ask most: a customer inquiry analysis method

Short answer

To find out what customers ask most, first build the topic list by hand, then count all conversations against it. The sequence: read 150–200 randomly chosen conversations, write each one's main question in a short sentence, group similar questions, make a list of 8–15 topics, add "other", and only then switch on automated analysis.

The most common mistake is asking AI directly "what do customers ask about?". The answer looks convincing, but it comes out differently every time and cannot be compared month to month.

Why this question comes first

The list of inquiry topics underpins almost every other analysis. Where to expand the knowledge base, what the bot must know, how many people the shift needs, which FAQ item is missing, which page of the website is unclear — all of it starts from "what do customers ask?"

Most companies know the list by guesswork: "mainly price and delivery". When they start counting, an unexpected topic often makes the top three.

Step 1: read 200 conversations

Pick 150–200 random conversations from the last month or two. Take them proportionally from all channels: if 60% of enquiries come through WhatsApp, about 60% of the sample should too. For each, write one sentence: "what does the customer want to know or do?" Keep the customer's own words; do not swap in your terms.

Step 2: group them

Sort the sentences into groups by similarity. Sentences in one group should need the same answer: "how many days does delivery take?" and "can you deliver by tomorrow?" are one group, while "is delivery free?" may be another, because the answer differs.

• If a group is very large (over 25% of the sample), split it into two or three subtopics.
• If a group has fewer than 3 conversations, merge it with a similar one or keep it in "other".
• If one conversation contains two separate questions, make the field multi-value rather than forcing one topic.

Step 3: write the topic list

Give each group a short name and a one-sentence definition. The definition should be clear enough that AI and a new employee would understand it the same way.

1. 1Name"Delivery time" — short, close to the customer's language.
2. 2Definition"The customer asks when the order will arrive or how many days delivery takes."
3. 3Not included"Delivery cost is a separate topic; a complaint about a late order is 'order problem'."
4. 4Examples2–3 phrases customers really wrote, including misspellings.

The spelling problem

Customers write the same thing in very different ways: Latin and Cyrillic scripts, missing special letters, words from other languages, abbreviations, voice-message transcripts. Keyword search does not catch this variety: three spellings of "delivery" count as three different words.

Analysis that chooses by meaning reduces the problem, because the topic is set by what the sentence means, not by a word. It still helps to add the most common variant spellings to the field's examples.

Step 4: track "other"

However good the list, some questions will fit no topic. They should land in "other" — not be forced into the nearest topic. The share of "other" is the list's quality indicator:

• Under 15%: the list reflects reality well.
• 20–30%: read the "other" conversations; one or two new topics have probably appeared.
• A sudden rise: there may be a new product, campaign or problem — that is a signal in itself.

Frequency, share and trend

Once the list is built, look at three measures. Frequency — in how many conversations the topic appears. Share — what percentage of all analysed conversations. Trend — how the share changes week to week or month to month. Frequency is for resource planning, share for priority, trend for spotting change.

When total volume changes, frequency misleads while share stays stable — this is covered in more detail in why conversations go deeper than a dashboard.

Human checking

For the first two weeks after automated analysis starts, open and check 10–15 conversations per topic. The question is simple: is this conversation really about this topic? If errors are systematic — say, "price" and "payment method" get mixed — strengthen the "not included" part of the definitions and re-analyse.

Illustrative example

This is an illustrative example. A gym chain believed customers mainly asked about price. Reading 180 conversations gave 11 topics. Price really was first (22%), but second, unexpectedly, was "freezing" (14%): members asked whether they could pause their membership during illness or travel. The rule was not on the website, and an agent explained it separately every time.

Decision: the rule was added to the website and the bot, and the topic's share began to be tracked.

Typical mistakes

• Building the list from internal department names ("sales question", "technical question") without reading.
• Mixing topic and intent: "price" is a topic, "wants to buy" is an intent.
• Adding new topics and changing old ones every month — the trend breaks.
• Counting the bot's or agent's sentences as customer questions.

The topic field in Vexvon

In Vexvon "question topic" is a field the company builds itself: name, instruction and closed value list. AI picks from that list for each conversation, a conversation can have several topics, and an answer that does not fit lands in "other". On the admin side, how often phrases in "other" repeat is shown separately; a frequent one can be added as a new value and old conversations re-analysed. The panel's AI assistant answers "most common questions last month" with links to conversations. More: Vexvon analytics.

Next step

Start this week with 50 conversations: write each one's question in one sentence and see how many distinct groups appear. Turning topics into FAQ is also discussed in chatbot vs FAQ page. We can build the list on your own conversations in a demo.