E-commerce
June 26, 2026
Your web dashboards show clicks, abandoned carts, and revenue by channel. They do not show the exact questions: "does this run big?", "where is my package stuck?", "can I exchange for free?". Yet, every conversation leaves a usable trail if you know how to read it.
The tension is simple: you receive hundreds or thousands of messages per month, but you still manage support with generic macros and an FAQ written without counting the real questions. As a result, you fix the website based on gut feeling while your customers are already telling you what to change.
This article does not repeat a classic analytics guide. It shows how to collect, classify, and leverage your customers' actual questions to act on the catalog, the bot, the helpdesk, and operations.
Summary
What is e-commerce conversational analysis?
E-commerce conversational analysis consists of extracting, classifying, and tracking themes, intents, and formulations from your customer exchanges: website chat, chatbot, email, Instagram DM, WhatsApp, SMS, and helpdesk tickets.
Unlike the aggregated behavior measured by GA4, it captures the why expressed in natural language: size doubts, fear of scams, delivery time confusion, return misunderstandings. Zipchat points out that web analytics track clicks, while conversational analysis tracks the questions, hesitations, and objections that block purchases (Zipchat, customer insights).
What you concretely gain
Support: prioritize macros, bots, and FAQs on the 5 to 10 dominant intents
Product: identify incomplete product pages, poorly named variations, missing size guides
Marketing: adjust delivery promises and social proof where objections repeatedly arise
Ops: detect WISMO spikes, carrier delays, or un-updated stockouts
Gorgias estimates that e-commerce brands generated over 350 million customer conversations in 2025, of which nearly 10 million led directly to a purchase. The rest contain revenue signals that most teams never report on dashboards (Gorgias, conversational commerce 2026).
See also what is e-commerce analytics and e-commerce analytics glossary to frame the overall data scope.
Why your web analytics alone are not enough
Web analytics answer: where, when, how much, which device. Conversation analysis answers: what is the customer asking, with what words, after which ordering step, and was the answer sufficient.
A high cart abandonment rate on mobile does not tell you if the blockage comes from shipping costs, delivery times, or product doubt. Pre-purchase chat transcripts complement session replay with the customer's exact phrasing.
Complementarity, not replacement
Web analytics: funnel, cohorts, revenue by channel, high-bounce pages
Conversation analytics: intents, verbatim, escalations, CSAT by topic
Ideal cross-referencing: source URL + SKU + order status + support intent tag
Alhena notes that GA4 and web dashboards cannot track what is happening inside a conversation. A conversational analytics engine links volume by topic, satisfaction by topic, and revenue impact in a single view (Alhena, analytics dashboard).
Check out what to track in e-commerce to align web and support dashboards.
Cross-reference high-bounce pages and top intents every month: the overlaps reveal quick wins.
Which channels should be included in your collection scope?
A reliable analysis begins with an explicit collection scope. Forgetting a channel distorts priorities: if Instagram concentrates pre-purchase questions and the helpdesk handles WISMO, you must merge both views.
Priority channels
Site chat and AI chatbot: pre- and post-purchase intents, human handoff
Email helpdesk: high volume, complex topics, attachments
Shopify Inbox / Meta: impulsive promo and stock questions
SMS support: often WISMO and same-day delivery
Transcribed calls: if you record in compliance with GDPR
Structured data to export
For each conversation: date, channel, language, order status if known, SKU or collection, agent vs bot, resolution, CSAT, existing tags. CSV or helpdesk API export, plus bot logs if separate.
Quality and governance
On Shopify, link conversation to order ID as soon as possible: the same "return" question has a different answer if the order is delivered or in transit.
Check that abandoned bot conversations are logged: they often reveal intents not covered by your flows. See 5-step feedback analysis for the collection to action loop.
How to build a customer question taxonomy?
Without a taxonomy, you drown insights in ad-hoc "miscellaneous" or "other" tags. A stable nomenclature enables trends, alerts, and bot routing.
E-commerce intent families
WISMO / tracking: where is my order, tracking number, delay, lost package
Pre-purchase product: size, compatibility, material, stock, comparison
Order and account: change address, cancel, invoice, promo code
Return and exchange: deadline, fees, label, partial refund
Payment and trust: security, double charge, authenticity, warranty
Product aftermath / CS: defect, manual, manufacturer warranty
Exceptional logistics: customs, pickup point, damaged package
Useful sub-tags
WISMO example: "not yet shipped" vs "in transit without scan" vs "delivered not received". Product example: "size guide" vs "real photo" vs "allergen". Keep a glossary: the customer says "send back", your policy says "return within 30 days".
Alhena describes an NLP clustering over 30 days of customer questions to generate topics specific to your store, such as "Order Tracking and Delivery" or "Sizing Help", rather than generic categories (Alhena, topic tagging).
Limit the initial depth to two levels (family + sub-tag). Read reducing Shopify WISMO and product questions via chatbot for concrete examples.
How to conduct a 90-day audit without drowning?
The 90-day QA audit provides enough volume to smooth out short-term seasonality while remaining current. Objective: top intents, recurring formulations, associated channels, and products.
Step 1: Export and cleanup
Export closed conversations for the period. Remove spam, internal tests, and bot-human duplicates on the same thread if you are counting intents.
Step 2: Stratified sample
Beyond 5,000 tickets, tag a representative sample per channel (e.g., 200 per channel) plus 100% of the topics already tagged as "escalation" or low CSAT.
Step 3: Tagging workshop
Gather the support lead, e-commerce, and someone from product for 90 minutes
Read 30 conversations out loud, noting intents without jargon
Validate the taxonomy V1 and written definitions
Tag the sample in the helpdesk or a shared spreadsheet
Step 4: Quantify and prioritize
Classify by volume, average handling time, repeat contact, and revenue impact (average shopping cart of pre-purchase threads). Prioritize the 5 intents that accumulate the most workload or conversion loss.
Step 5: 30-day action plan
For each of the top 3 intents: one FAQ action, one product page or checkout action, one bot rule or agent macro. Document the baseline volume per intent before the change. See product insights from support conversations.
Use an intent x volume x CSAT spreadsheet to prioritize readings: start with high-volume, low-satisfaction categories.
Manual or AI tagging: which approach to adopt?
Manual tagging remains the reference for calibrating the taxonomy. AI intent detection scales across thousands of threads once examples are validated.
When to remain manual
Taxonomy launch, sensitive disputes, legal cases, and monthly quality samples (5 to 10% of conversations) to check for model drift.
When to automate
Volume higher than 300 conversations per week, repetitive intents (WISMO, returns, sizing), helpdesk with AI suggestions, or bot that logs the intent.
Recommended hybrid
The AI suggests a tag, the agent confirms or corrects it with one click. Corrections feed the retraining process. Low confidence threshold: human file review.
Pitfalls to avoid
Confusing politeness with intent ("thank you" tagged as WISMO)
Multi-intents in an undivided message
Mixed languages or spelling errors not covered by the model
Bias toward historical tags over-represented in the training data
Decagon describes intent tags in three levels: broad theme, precise sub-category, then conversation outcome to measure deflection and escalation by intent (Decagon, intent tags).
Measure the agent correction rate on AI tags: above 15%, review definitions or training examples.
How to read trends and spot weak signals?
Once tagged, trend analysis transforms counters into decisions: season, campaign, customer segment, carrier, or collection.
Temporal trends
Weekly chart by intent: WISMO rising 48h after a "D+2" shipping promo indicates a promise/reality gap. "Return" peak in January: normal post-holiday, but compare by collection.
Useful segmentation
First order vs. recurring: more trust and tracking questions
High average order value: more invoice and express delivery requests
Market and language: customs, US vs. EU sizes
Acquisition channel: TikTok traffic vs. email newsletter
Product or SKU: hotspot page to be enriched
Alert signals
24h spike on "payment declined" or "scam": check checkout and external reviews. Increase in "poor quality" on a batch: supplier quality alert.
Alhena offers volume by topic, daily curve, top URLs and countries by topic, then drill-down on tagged conversations. If "returns" explodes especially from the UK site, the action differs from a global peak (Alhena, trending topics).
Cross-reference your trends with e-commerce chatbot KPIs to see if the bot is effectively absorbing the rising intents.
What concrete actions for FAQs, product sheets, and bots?
Conversational analysis is only useful if it feeds into visible customer actions. Each recurring intent must have an owner and content or automation delivery date.
FAQ and help center
Write with the exact headings counted in the audit. Place links from checkout, order tracking page and transactional email. A FAQ entry consulted and then followed by a ticket on the same topic must be rewritten.
Product pages
"Size uncertainty" intent on 40% of chats for a product ID? Add guide, model, EU/US table, size-filtered reviews. "Compatibility" intent: diagram or list of supported models above the fold.
Bot and agent macros
One intent = one bot flow with clear handoff. Agent macros aligned with the same formulations as the bot for consistency. Update both on the same day when the policy changes.
Four-step closed loop
Intent volume increases
Content or bot action deployed
Measure intent volume at D+30
If persistent, second iteration or product escalation
Keep an "intent → action → owner" log in Notion or your support wiki. Prioritize high repeat contact intents: a failing FAQ costs twice the apparent volume. See E-commerce FAQ that reduces tickets.
Which conversational KPIs can be linked to revenue?
Conversational KPIs translate analysis into steering: workload, quality, conversion, and revenue linked to the dialogue.
Volume and efficiency indicators
Volume by intent and channel (trend, not snapshot)
Resolved share bot vs agent by intent
First response and resolution time by category
Repeat contact within 48 hours on the same order and same tag
CSAT or thumbs post-conversation by intent
Business indicators
Conversion rate of pre-purchase threads containing a resolved objection. Revenue attributed to bot conversations. Cost per conversation vs cost per email ticket. Reduction in tickets after FAQ update measured on target intent.
Minimal dashboard
One page with top 8 intents, 4-week evolution, reduction target for the quarter's 3 strategic priorities. Monthly sharing with management.
Gorgias documents in 2026 that 91% of brands still track CSAT, but 60% add AOV, and brands at $20M+ look at cost per resolution, incremental revenue, and first contact resolution rate (Gorgias, AI adoption 2026).
Set a specific target (e.g., WISMO -20% in 60 days) rather than a vague goal of "improving support". Link hourly agent cost to average time per intent to quantify the ROI of an FAQ that removes 200 WISMO tickets per month. Align with analytics Qstomy.
Which errors distort your conversational insights?
Many shops collect transcripts without ever changing the FAQ or bot. Here are the most common analysis errors.
Too vague tags: "other" at 30% invalidates the audit. Weekly review of unclassified items required
Ignoring pre-purchase: focusing on post-purchase WISMO misses conversion objections
No order link: text without Shopify status mixes up "return possible" and "too late"
One-shot audit: without continuous updates, June promo intents are not predicted
Channel silos: keeping Instagram separate from the helpdesk doubles the effort
Over-automating tagging without QA: blind trust in AI degrades data within weeks
If your team debates opinions in meetings, go back to the intent numbers: taxonomy settles it. Document "we do not handle this intent via bot" decisions with legal or brand reasons. If no one owns the intent registry, appoint a support representative with 2 hours per month of data maintenance.
How does Qstomy structure the analysis of customer conversations?
Qstomy connects conversations, Shopify data, and support content to respond with order context and surface recurring intents without permanent manual export.
What Qstomy brings to the analysis
Contextualized responses: order status, tracking, return policy in the thread
Intent visibility: WISMO, product, return, payment to prioritize FAQ and bot
Fewer repetitive tickets when lookup and self-service cover the top volume
Structured handoff: tag, history, and order passed to the agent
Content alignment: same answers as your help center if unified base
DTC Scenario in figures
A DTC brand processes 2,800 orders per month and receives 720 support conversations (chat, email, bot). After a 90-day Qstomy audit, the team identifies that 24% of messages concern "delivery time before purchase" on a capsule collection, 18% of WISMO could be resolved by order lookup, and 11% of product questions recur on three SKUs without a size guide.
60-day pilot objective: cover the three dominant intents in Qstomy, enrich the relevant product pages, measure -30% volume on "pre-purchase delivery time", -22% WISMO, and CSAT for "return" topic rising from 71% to 85%. Estimated avoided agent cost: 140 h over the period if the average time per ticket remains at 8 minutes.
Qstomy complements your web analytics stack: it reads the verbatim where the customer actually asks their question. Discover Qstomy AI customer support and request a demo to see intents and resolution on your catalog.
Which playbooks should be launched this week?
Playbook 1: express count on 100 conversations
Export 100 closed conversations from the last 30 days. Manually count the five most frequent phrasings. You already have the core of your next FAQ and the foundation of your V1 taxonomy.
Playbook 2: 90-minute taxonomy workshop
Bring together support, e-commerce, and product. Read 30 transcripts out loud. Validate 8 to 12 intent families and 3 sub-tags per priority family. Document 10 verbatim examples per intent before tagging the rest.
Playbook 3: intent → action in 30 days
For your top 3 volume intents, assign an owner, an FAQ action, a product page or checkout action, and a bot rule. Measure the intent volume at D+30 using the exact same counting method as at D0.
Playbook 4: 15-minute weekly review
Every Monday, open the 7-day intent volume graph. Note spikes, new unclassified topics, and low-CSAT topics. A carrier alert or a checkout bug is often seen 24 to 48 hours before the ticket peak.
Useful links
Analytics: e-commerce analytics to track
Product Insights: insights from support
Objections: detecting purchase objections
AI Tickets: reducing tickets with AI
Schedule a quarterly intent review with product and ops to turn conversational analysis into a ritual, not a forgotten project.
Enzo
June 26, 2026
