Customer intent analysis vs simple keyword counting

Short answer

Keyword counting shows which word appears how often in conversations. Customer intent analysis identifies what the customer wanted to achieve in this conversation: to buy, compare prices, complain, cancel, or just gather information. The word "price" can appear both with a customer ready to buy and with one who wants to cancel — same word, completely different intent.

Keyword counting is cheap and transparent, but it falls into five typical traps. Intent analysis avoids them, but needs a clear definition and human checking.

When word counting is enough

There is no need to drop word counting entirely. It works well for precise, unambiguous terms: a product name, a competitor's brand, a campaign code, a specific service. To know how often "Model X" comes up, you do not need complex analysis. Platforms such as AWS Contact Lens keep exact-phrase matching as a separate option and recommend it for finding specific phrases, such as a scripted greeting.

Trap 1: negation

"I don't want a discount, I'm just asking about delivery" contains the word "discount". Word counting records it as interest in a discount. In many languages, Azerbaijani included, negation often sits at the end of the verb, so a search that finds the word misses it.

Trap 2: quotation and someone else's words

"My friend said your prices are high, but I want to take a look" — here "high prices" is not the customer's own objection. They are repeating someone else's view and, if anything, showing interest.

Trap 3: synonyms and spelling variants

One intent comes in dozens of forms: "too expensive", "over my budget", "cheaper elsewhere", "any discount?", in another language or mixed script. However long the word list, it always misses something, and the longer it gets, the more false matches it produces.

Trap 4: the agent's and the bot's words

"Warranty" appears 30 times in conversations, but the agent said every one of them because it is in the script. Word counting shows this as customers being interested in warranties. Intent must be taken only from the customer's words.

Trap 5: context and sequence

A customer first asks the price, then delivery, then says "I'll pick it up tomorrow". Word counting sees three topics. The intent is one: to buy. Conversely, "thanks a lot, I'll think about it" contains no keyword at all, yet the intent — for now at least — is not to buy.

How to define an intent field

1. 1Start from the decisionWhat decision do you need intent for? For a sales funnel, "buy / compare / information / complaint" may be enough.
2. 24–8 valuesToo many values blur into each other. Each value should call for a different next step.
3. 3Definition + boundary"Wants to buy: talks about a specific product, date or payment, or states agreement. Only asking the price is not included."
4. 4UnclearIf the customer did not state an intent, there should be an "unclear" value — no forced choice.
5. 5Customer onlySay explicitly in the instruction: decide only from the customer's words.

A hybrid approach: words and intent together

In practice the best results come from combining the two methods. Word counting catches exact names: which product, which competitor, which campaign code. The intent field says what the customer wanted in that same conversation. For example, looking at the intent split in conversations that mention a competitor shows whether customers mention them mostly with a "compare" intent, or already say "I bought from them and there's a problem". The first case calls for a sales step; the second is an opportunity for marketing.

A small test to validate the intent field

1. 1Pick 40 conversationsAbout 10 per intent value, so rare values are checked too.
2. 2Two people score separatelyEach writes their own value without looking at the AI's result.
3. 3Compare the three resultsIf the two people disagree, the problem is the definition, not AI. If the people agree and AI differs, sharpen the instruction.
4. 4Fix the definitionRewrite in one sentence the boundary between the two most disputed values, then repeat the test.

Combining topic and intent

The most useful result comes from crossing topic with intent. Of the conversations about "price", how many are "buy" and how many "compare"? How does the intent split differ between customers from Instagram and from WhatsApp? How to build the topic list is covered in customer inquiry analysis.

An illustrative comparison

This is an illustrative example. A furniture store sees "credit" in 18% of conversations and plans a credit campaign. Once an intent field is added, the picture changes: in only a third of the conversations mentioning credit does the customer ask about it themselves. In the rest the agent offers it, or the customer says "no credit, cash". The campaign is scaled down and a delivery discount for cash payment is tested instead.

The limits of intent analysis

• Intent is an estimate based on the conversation; the customer may change their mind later.
• In short, vague conversations intent will often be "unclear" — that is normal.
• AI can be wrong; irony and polite refusals are especially hard.
• Intent should not be the sole basis for a high-stakes decision about a customer (credit, refusal, different pricing).

Do not hide the "unclear" share in the report. If it grows, either conversations have become shorter or a new group of customers has arrived whose intent does not fit the existing values. Both deserve a separate look.

The intent field in Vexvon

Vexvon uses field-based analysis, not word counting: the company builds a field such as "intent" with its own values and instruction, AI reads the customer's messages in context for each conversation and picks one value from the list, keeping the customer's own sentence as evidence. The panel's AI assistant answers cross-tab questions such as "intent split by channel"; the cross-tab counts conversations, not fact rows, so one conversation is not counted twice. More: Vexvon analytics.

Next step

Take one keyword you track today and read 30 conversations where it appears: in how many is the customer really showing that intent? If the gap is large, it is time to move to an intent field. We can build the field on your own conversations in a demo.