Skip to content
AI-SNS.IO — CHAT SUPPORTAI-SNS.IO — CHAT SUPPORT
KPIs

Reading Conversation Analytics: From Charts to Fixes

AI SNS CHAT EditorialPublished 6 min read

How to read conversation analytics: volume, topics, sentiment and ratings, and how to turn each pattern into a concrete fix for your support agent.

You open the analytics page, see a handful of charts, and wonder which of them should change what you do on Monday. Most teams either stare at one headline number or ignore the dashboard entirely. This guide shows how to read conversation analytics in a sensible order, which patterns matter on website chat, WhatsApp and Messenger, and how to turn each one into a specific fix.

Start with the question, not the chart

A dashboard answers questions you bring to it. Before you look at any graph, decide which of these you are asking:

  • Is the agent answering correctly?
  • Are customers getting what they came for?
  • Where is the work landing on my human team?
  • Is the cost of running this sustainable?

Each question maps to a different view. Mixing them is how people end up celebrating "more chats" when the real story is that customers keep asking the same thing because the answer is wrong.

A useful habit is a short weekly review with a fixed order: volume, topics, sentiment, individual conversations, then action. The rest of this article follows that order.

Volume: the baseline, not the verdict

Chat volume tells you how much demand arrives and when. It is context for everything else.

What to look for

  • Shape over time. Do chats cluster after a campaign email, a product launch or a billing run? Spikes with an obvious cause are fine. Spikes without one deserve a look at the conversations.
  • Channel mix. Website chat visitors often ask pre-purchase questions. WhatsApp and Messenger customers tend to be existing customers following up. The same agent may need different emphasis in each place.
  • Trend after a change. If you add a help article or retrain the agent, volume on that topic should shift. If it does not, the change did not land.

Rising volume is neither good nor bad on its own. Falling volume can mean customers found answers elsewhere, or that they gave up.

Topics: where the real information lives

Topic views group conversations by what customers were asking about. This is usually the most actionable part of any analytics page, because it tells you what your customers care about in their own words.

Reading a topic list

Look at three things for each topic:

  1. Size. How much of your traffic does it represent?
  2. Trend. Is it growing, steady or fading?
  3. Outcome. Do those conversations end well, or do they end in handoffs and thumbs down?

A large topic that resolves well is a success; leave it alone. A small topic that always escalates is a gap in knowledge or a case that truly needs a person. A growing topic is an early signal, often a product issue, a confusing policy or a broken page.

Order status
110
Returns and refunds
70
Product questions
60
Shipping costs
35
Other
25
Example: share of chats by topic (hypothetical shop with 300 chats a month)Illustrative example

In this made-up example, order status and returns make up well over half of the traffic. That tells you where to spend your effort: make sure the agent can answer those two topics flawlessly, and consider connecting order lookup so it does not have to guess.

Question history: the exact words customers use

Topics are summaries. Question history is the raw material. Reading actual questions shows you phrasing, typos, mixed languages and the assumptions customers make.

Use it to:

  • Spot questions your knowledge base answers poorly or not at all.
  • Notice vocabulary mismatches. Customers say "money back" while your documentation says "reimbursement".
  • Find questions that arrive in a language your content does not cover.
  • Collect real wording for new Q&A entries.

Fifteen minutes with this view each week usually produces more improvement than any chart.

Sentiment: read it as a smoke alarm

Sentiment analysis labels the tone of conversations. It is a useful signal, but a blunt one. Frustrated customers are not always unhappy with the agent; sometimes they are unhappy with the situation, such as a late parcel, and the agent is handling it fine.

How to use it well

  • Compare topics, not just totals. Overall sentiment hides a lot. Negative sentiment concentrated in one topic points to a specific problem.
  • Watch for change. A sudden dip after a policy update or a model change is worth investigating.
  • Always open examples. Never act on a sentiment score without reading several of the underlying conversations.

Negative sentiment paired with a high handoff rate on the same topic is a strong sign the agent lacks the right information or authority there. Negative sentiment with successful resolution often means the problem is upstream of support.

Ratings and handoffs: the human check

Automated labels are helpful, but direct feedback is better evidence. Thumbs up and thumbs down ratings tell you what customers thought of an answer. Handoffs tell you where the agent chose, or was configured, to step aside.

Signal
What it suggests
Thumbs down on a clear topic
The answer is wrong, outdated or unhelpful
Frequent handoff on one topic
Missing knowledge, or a case that needs a person
Negative sentiment, no handoff
Frustration with the situation, or a tone issue
Repeated near-identical questions
The first answer did not satisfy, or the source page is unclear
What each signal tells youIllustrative example

Keep in mind that ratings are sparse. Only a fraction of customers click them, so a handful of thumbs down is a prompt to read conversations, not a statistic to report.

Turning patterns into fixes

Analytics only matters if it changes something. A simple loop works well:

  1. 01Spot a pattern
  2. 02Open the conversations
  3. 03Fix the source
  4. 04Check next week
The weekly improvement loopIllustrative example

Typical fixes, matched to what you saw:

  • Wrong or missing answer: add or correct a Q&A entry, or update the source document.
  • Right answer, poor delivery: adjust the persona and guardrails, for example shorter replies on WhatsApp.
  • Needs a person: tighten the handoff rules so the right cases reach a human quickly.
  • Needs live data: connect an integration so the agent looks things up instead of guessing.

Change one thing at a time where you can. If you rewrite the persona, add five documents and swap the model in the same week, you will not know which change helped.

Doing this in AI SNS CHAT

AI SNS CHAT puts these views in a few places, so a weekly review is a short tour rather than a hunt.

Analytics. The sidebar's Analytics section includes Chats, Topics, Question history and Sentiment, and Voice analytics for phone and voice conversations. Advanced analytics need the Pro plan; see pricing for what each tier includes.

Chat logs. Conversations from every channel except email are under Activity → Chat logs. Here you can filter, select several at once, and flag, escalate, resolve or delete them. You can also export them and see thumbs up and thumbs down ratings. This is where you check what sits behind a number on the dashboard. Email conversations are in the Inbox instead.

Fixing answers on the spot. In the Inbox and the chat logs you can take over a chat, which pauses the bot, and edit a reply into the ideal answer to train the agent. That closes the loop directly from the conversation where you spotted the problem. Q&A entries and Suggestions live under Data sources if you prefer to add knowledge deliberately.

Testing before you publish. Use Playground to try the changed agent against the questions that failed, including the A/B model compare if you want to see two models side by side.

Leads and cost. Leads collected by the agent appear in Activity → Leads, and each lead also creates a contact. The Usage page shows credits per agent, which helps you relate conversation volume to cost. One reply uses one credit.

Ask Synthia. If you cannot find a screen, Synthia, the built-in fleet manager, can guide you there by navigating to the screen and highlighting the control.

A practical weekly routine: open Analytics, note the biggest and fastest-growing topic, open the matching conversations in Chat logs, fix one source or reply, and check the same view next week. For a broader tour of what is available, see features.

Common reading mistakes

  • Chasing one number. A single metric rarely reflects quality. Read volume, topics, sentiment and ratings together.
  • Ignoring small topics. Rare but high-stakes questions, such as billing disputes, matter more than their size suggests.
  • Reading too little data. With a few dozen chats, one odd conversation can swing a chart. Wait for a meaningful sample, or read the conversations directly.
  • Never reading transcripts. Charts point; conversations explain.

Frequently asked questions

How often should I review conversation analytics?

Weekly is a good rhythm for most teams, with a closer look after any launch, campaign or change to the agent. Daily checks are rarely necessary unless you are in a launch period or testing a major change.

Is negative sentiment always a sign the AI agent is failing?

No. Customers are often frustrated by the situation, such as a delayed order, rather than by the answer. Open the conversations behind the score and check whether the agent handled the situation well.

Do I need the Pro plan to get value from analytics?

Not to start. Chats, Topics, Question history and Sentiment, combined with the conversations in Chat logs, are enough for a useful weekly review. Advanced analytics need Pro, and you can see the plan details on the pricing page.

Where do email conversations appear compared with WhatsApp or Messenger?

Email lives in the Inbox. Conversations from the other channels, including website chat, WhatsApp and Messenger, are in Activity → Chat logs.

#analytics#KPIs#customer support#AI agents#sentiment

Put this into practice with AI SNS CHAT

Build an AI agent trained on your own content and connect it to your website, LINE, WhatsApp and more. Free to start, no credit card required.