Negative customer feedback signals: what appears in conversations before a review

Short answer

A negative review rarely appears out of nowhere: beforehand, the customer talked to you and voiced their dissatisfaction. The most typical early signals are an explicit threat ("I'll write a review", "I'll tell everyone"), a repeated unresolved problem, a broken promise ("you said tomorrow, it's been three days"), a sense of unfairness ("you did it for someone else but not for me") and the conversation ending unanswered. These signals can be caught in conversations, their cause checked, and the right step taken in time.

The aim is not to discourage the customer from writing a review — that is unethical and prohibited in some jurisdictions. The aim is to solve the problem before the review: about a solved problem, the customer either does not write or writes a different review.

Why early signals matter

Once a public review is written it is hard to change, and a recovery call only limits the damage (negative review recovery call). In a conversation, the customer is still talking to you — the best moment for a fix. The problem is that these signals get lost among hundreds of ordinary conversations, and nobody looks at them in time.

Five early signals

1. 1An explicit threat"I'll write a review", "I'll post it on social media", "I'll go to consumer protection". The latest but clearest signal.
2. 2A repeated problemThe customer writes about the same problem a second or third time.
3. 3A broken promiseThe company promised a time or an action and did not deliver.
4. 4A sense of unfairnessThe customer compares themselves with someone else or feels a rule was applied differently to them.
5. 5An unanswered endingThe customer's last message expresses dissatisfaction and no reply came.

A signal is not a verdict

Not every "I'll write a review" becomes a review, and not every review follows these signals. A signal means only one thing: a person should look at this conversation. So write "conversations with an early warning signal" in reports, not "customers who will write negative reviews". Labelling a customer "difficult" and treating them differently is wrong.

Checking the cause

The person reading a flagged conversation answers three questions: is the customer's complaint confirmed by facts (order status, payment record), is the problem still current, and is there an error on the company's side? Sometimes the customer is right and the problem must be fixed immediately. Sometimes there is a misunderstanding and a clear explanation is enough. Sometimes the company applied its rule correctly — and even then the customer deserves a respectful, clear answer.

A response rule

1. 1TimingA conversation with an explicit threat or a broken promise is reviewed within the same working day.
2. 2WhoSomeone with authority to solve the problem: a shift lead, the support lead, an account manager.
3. 3WhatAcknowledge the problem, give a concrete fix or date, deliver what was promised.
4. 4What notOffering compensation in exchange for not writing a review; pressuring the customer.

How to answer a broken promise

A broken promise is the most common signal and the quickest to fix. A good answer has four parts: openly acknowledging the promise was broken ("we told you tomorrow, and we're late"), explaining the reason briefly and without excuses, giving a concrete new date or action, and naming the person responsible for keeping it. An apology without a reason and a vague promise ("we'll sort it soon") usually increase dissatisfaction, because the customer has already heard that sentence once.

The aggregate view: where reputational risk comes from

Beyond individual conversations, the topic split of signals shows where reputational risk originates. If most "broken promise" signals concern delivery time, the problem is in agents' promises or in logistics. Such a finding is worth more than individual recovery steps: it lets you fix the source (complaint root cause analysis).

Setting up the fields

• "Early warning signal": the five values plus "none". Multi-value.
• "Topic of dissatisfaction": delivery, quality, price, service, returns, other.
• "How the conversation ended": resolved / unresolved / left unanswered.
• In the instruction: choose the signal only from the customer's words.

Illustrative example

This is an illustrative example. A restaurant chain's delivery service collects early warning signals. In the first month, most "broken promise" signals relate to the same situation: agents say "in 30 minutes", while in the evening delivery takes over an hour. Some of these conversations end with "I'll write a review".

Decision: in the evening, agents and the bot state the real waiting time, and customers are messaged in advance when an order is running late. Next month, the share of "broken promise" signals is tracked.

The link with monitoring

Conversation signals show the stage before a review; the review itself is written on social networks, maps and review platforms. Tracking both together is useful: if a topic seen in conversations later grows in public reviews too, early warning is working but the response is late. Tracking public reviews needs a separate tool and process.

Typical mistakes

• Watching only the explicit threat signal — it is already late.
• Pressuring customers or offering gifts so they do not write a review.
• Labelling a customer with a signal as "difficult".
• Resolving individual conversations without fixing the source.

Limits

• Some customers who write reviews never talk to you beforehand.
• AI may misread irony and emotional phrasing — check signals with a sample.
• Seeing the link between a signal and a later review needs data from review platforms.

Early warning in Vexvon

In Vexvon the company sets up fields such as "early warning signal", "topic of dissatisfaction" and "how the conversation ended" itself; AI picks from the customer's words in each conversation and shows the evidence message. The panel separately counts conversations whose last message is from the customer and that went unanswered, and the AI assistant can list customers matching a condition. Because analysis runs 24 hours after a conversation's last message, agents also need their own escalation rule for urgent cases. More: analytics.

Next step

Find the authors of 10 negative reviews from the last three months in your conversation base and read their conversations from before the review: which signals were visible? For the link with churn, see churn signals in conversations; we can set up the fields together in a demo.

Further reading on this topic: sentiment vs customer satisfaction.