Sentiment analysis: turn feedback into fixes

Classify customer sentiment in chats and reviews, check mixed feedback and sarcasm, compare analysis tools and turn complaints into better answers.

Sentiment analysis classifies the tone expressed in written feedback as positive, neutral, negative or mixed. For a retailer, its useful result is a reviewed list of frustrating topics linked to the customer’s original words. Those examples help staff fix unclear product details and unhelpful answers.

A negative label alone can’t tell you which part of a shopping experience went wrong. Keep the product, delivery and service comments separate so you can assign the right correction.

What sentiment analysis tells you

A sentiment label describes expressed approval or criticism under the guide you use. In this hypothetical shopping message, “The coat is warm and fits well”, the expressed sentiment is positive. That label doesn’t prove the writer’s emotional state or explain what caused their experience.

Positive wording expresses approval. Negative wording expresses criticism or dissatisfaction. Neutral text states information without clear approval or criticism. Mixed text contains both positive and negative views.

Intent describes what someone wants to do, such as return a coat. Emotion detection tries to identify feelings such as anger or joy. Customer satisfaction is a separate judgement, often collected through a rating. A politely written return request can be neutral even when the customer wants a refund.

How sentiment analysis methods work

Automatic sentiment analysis takes written text, applies a classification method and returns labels or scores. Keep the source wording beside each result so a reviewer can check it.

Natural language processing (NLP) is the use of software to process human language. Text sentiment analysis applies that work to written tone. AI sentiment analysis uses trained models to predict a label from the wording and context they receive.

Compare classification methods

MethodSetup and costContext and correction
Manual codingStaff read comments using a shared guide. Review time grows with volume.Reviewers can inspect the whole exchange. Resolve disagreements and record reasons.
Fixed word rulesDefine words, phrases and exceptions. Software is simple, but maintaining rules takes time.“Not bad” and sarcasm can defeat word counts. Amend rules and recheck affected records.
Trained classifierUse a model trained on labelled examples. Allow for service usage and setup work.Predictions depend on its training and input. Review errors against your store’s guide.
Language-model promptingSupply instructions, examples and the text. Allow for usage charges and repeated runs.Instructions can include context and topic rules. Keep the prompt and model version with the output.

Automation reduces repeated reading. People still define the labels and correct mistakes. A model’s confidence value describes its confidence in a prediction under that system. It isn’t a universal scale of how strongly a customer feels.

An overall label describes the whole message. Sentence-level analysis labels each sentence. Aspect-based sentiment analysis separates views about attributes such as fabric, fit or delivery. One sentence can praise fabric and criticise delivery, so splitting sentences alone won’t always separate the topics.

Label retail feedback by topic and sentiment

Prepare a sample whose labels can be traced back to individual conversations or reviews. Start with written comments you already hold.

  1. Define a permitted sample, such as completed product-question chats and product reviews from equal-length periods. Confirm that your organisation permits the intended analysis and provider use.
  2. Keep one source identifier per conversation or review. Remove duplicates and unnecessary names, email addresses, order details and payment information from the analysis copy.
  3. Separate customer wording from agent replies. Preserve message order and speaker roles in a restricted original record, with access limited to staff who need it.
  4. Define the labels before classifying. Keep topic, product or brand reference, predicted sentiment, reviewed sentiment and correction reason in separate fields.
  5. Review a small sample manually before expanding. Confirm that each label has an identifiable phrase supporting it and that uncertainty remains visible.

Before uploading real text, approve the provider’s processing terms and retention settings for that use. For example, AWS’s Comprehend privacy guidance describes possible input storage and service-improvement use, with an opt-out policy. Qualtrics has organisation-level retention controls. Removing names alone doesn’t settle whether an upload is permitted.

A small retail label guide

These hypothetical examples use the same four sentiment labels across five topics. An unanswered factual question is neutral unless its wording expresses approval or criticism.

TopicHypothetical customer wordingLabelEvidence to retain
Sizing“The sleeves are too short.”NegativeSleeves and the named product
Delivery“It arrived when promised.”PositiveThe delivery statement
Quality“The fabric is lovely.”PositiveFabric praise
Price“What does this cost?”NeutralThe price question
Product suitability“It won’t fit my narrow shelf.”NegativeThe stated space requirement

For a mixed hypothetical review R1, use “The fabric is lovely, but delivery took longer than promised.” Record quality as positive and delivery as negative. The overall label is mixed under this guide. Keep the product reference separate from both topic labels.

A hypothetical written contact-centre chat C1 starts with “Will this coat fit over a jumper?” and ends with “You still haven’t given me the chest measurement.” The opening question is neutral. The later message expresses negative sentiment about sizing information. Preserve both turns and record a negative closing label for the conversation. The evidence here is written chat text.

Use the guide to collecting shopper feedback when you need to collect comments. Once labels are reviewed, the voice of customer programme is where the findings can inform broader decisions.

Check sarcasm, mixed feedback and classification errors

Check sentiment predictions against the customer’s meaning before using them to prioritise work. A count becomes misleading when the same error repeats across a common phrase.

Errors and their corrections

ErrorHypothetical wordingCheckCorrection
Sarcasm“Brilliant, another missing parcel.”Compare the praise word with the reported event.Label delivery negative and record the sarcasm.
Negation“The quality isn’t bad.”Read the negated phrase as a whole.Use positive when the context supports mild approval.
Quoted agent text“You said ‘easy returns’, but this is a headache.”Identify who wrote each quoted phrase.Label the customer’s returns criticism negative.
Shorthand“nah, too exxy”Confirm the phrase’s meaning in context.Record negative price sentiment once understood. Otherwise retain it for review.
Changing sentiment“That’s frustrating” followed by “Thanks, that solves it.”Read the ordered customer turns.Preserve the negative opening and positive closing labels separately.

Two reviewers can independently label the same manageable sample, such as 20 conversations, using the shared guide. They then compare labels and resolve disagreements by pointing to the wording. Record the agreed label and the reason. Add an example to the guide when the disagreement exposes an unclear rule.

Compare the automated labels with those agreed labels in a confusion table. Put reviewed labels in rows and predicted labels in columns. Each cell counts records. A concentration of mixed reviews in the predicted-positive column exposes lost complaints even when the overall match rate looks high.

Keep context-poor cases unresolved instead of forcing them into neutral. Review uncertain predictions and a selection of confident predictions too. If the model or label guide changes, reclassify the comparison sample or start a separate series. Don’t merge incompatible outputs into one trend.

Choose sentiment analysis tools for retail feedback

Choose software around where the original text lives and how staff will correct its labels. Helpdesk reporting, dedicated feedback analysis and a configurable analysis service suit different workloads.

Zendesk intelligent triage suits written support feedback already held in tickets. It supplies topic and sentiment fields, with agent corrections and confidence fields. Its sentiment labels run from very positive to very negative, without a mixed option. Preserve mixed aspects separately when your guide needs them. Dynamic classification can use later customer replies, with behaviour depending on the channel.

Classifications require an eligible paid subscription, and workflow use requires the Copilot add-on.

Qualtrics Text iQ suits feedback analysis where you need editable overall and topic-level sentiment. It includes a mixed label and exports topic sentiment with response data. English is among its optimised sentiment languages. Sentiment analysis requires Advanced Text iQ.

Basic Text iQ’s topic tools don’t establish access to that feature. Its functionality guide directs customers to their account executive for Advanced access. Include your import and licensing needs in that discussion.

Amazon Comprehend suits a retailer with technical help to build a text-analysis workflow. It returns overall sentiment labels and confidence scores. Its separate targeted sentiment function finds sentiment about entities and attributes in English text. You must retain source identifiers and build the review and correction process around its outputs.

It analyses text you supply. It isn’t an Amazon marketplace review scraper or an integration with shoppa.

Compare the operational requirements

RequirementZendeskQualtrics Text iQAmazon Comprehend
Original wordingTicket conversationFeedback responseYour retained source text
Topic detailTicket topic and sentiment are separate fieldsEditable topic sentimentEntity-level sentiment through targeted analysis
CorrectionsAgents update classification fieldsEdit overall and topic labelsKeep reviewed overrides in your own records
LanguagesValidate sentiment on your actual ticket languagesUse its sentiment-language guidance for your sampleTargeted sentiment supports English. Overall analysis has broader language support.
Export and scaleInclude required ticket fields and reporting access in the subscription evaluationResponse exports include topic and sentiment fieldsAPI outputs need your own reporting workflow and quota handling
Cost structureSubscription and relevant add-onsAdvanced feature access through your accountUsage-based processing plus implementation and review

Amazon Comprehend’s batch quotas allow up to 25 documents per synchronous batch, with a 5 KB limit per document. Longer records need a deliberate segmentation rule. Otherwise, splitting one conversation into several documents can change your counts.

Compare the full cost of preparing text, running analysis, reviewing errors and maintaining exports. Include access controls and deletion rules in the purchase decision. An online scorer that accepts pasted text is only useful for real customer data when its processing terms permit that use.

Free sentiment analysis tools also need a precise definition of “free”. Manual labels in an existing spreadsheet need staff time. A restricted free allowance caps usage, while a time-limited trial ends.

Amazon Comprehend’s pricing page lists a restricted, time-limited free tier for eligible analysis. Character-based charging and minimum request sizes affect paid usage. A free account elsewhere doesn’t establish access to sentiment analysis.

Measure sentiment without confusing it with satisfaction

Measure customer sentiment using a stated counting unit and denominator. Count each conversation or review once in an overall distribution. Topic counts can overlap because one record can discuss several aspects.

A hypothetical five-record sample

Source recordCustomer feedbackReviewed overall label
R1, product reviewLovely fabric, delayed deliveryMixed
C1, written chatStill missing the chest measurementNegative
R2, product reviewCoat is warm and fits wellPositive
C2, written chatWhat does this cost?Neutral
C3, written chat“Same again”, with no earlier context availableUnresolved

For this hypothetical sample, four records have resolved labels. Positive, neutral, negative and mixed each account for one of four resolved records, or 25%. Report one unresolved record separately, equal to 20% of all five sampled records. State both denominators so the chart doesn’t hide the unresolved case.

Delivery appears negatively in R1, and sizing information appears negatively in C1. Fabric praise remains visible in R1 despite its delivery complaint. Keep agent replies out of these customer counts. If you measure staff wording, use a separate series.

Customer satisfaction (CSAT) uses a separately collected satisfaction response and its own denominator. A model confidence value measures prediction confidence. Average polarity summarises numeric sentiment scores under a particular scale. None supplies a customer’s satisfaction rating. Compare reviewed sentiment with service performance measures using the same source and journey stage.

For trends, compare equal periods of product reviews with product reviews, and pre-purchase chats with pre-purchase chats. A larger share of delivery complaints can follow a change in the sampled source. Record the sample definition, guide version and classifier version beside each series.

Use shopping conversations to fix product answers

Use the original exchange to identify which answer needs changing. Follow the customer’s product and question through the reply to the next observable action.

shoppa is an AI shopping agent that helps customers find, compare and buy products in the chat, then supports them after the sale. It uses the retailer’s catalogue and documents. Its analytics include a step-by-step conversation trace, so staff can inspect what happened around an unhelpful answer.

In a hypothetical conversation, a customer views coat P17 and asks whether it fits over a thick jumper. The answer repeats the generic size chart. The customer asks for the chest measurement again, then leaves without adding the coat to the basket. Label the repeated frustration for review, and retain that trace as evidence of the unanswered question. Leaving alone doesn’t prove why the customer stopped.

Assign the catalogue owner to add the missing chest measurement for each size. If the measurement already exists, assign the agent-settings owner to correct the instruction that selects the sizing information. A delivery complaint can instead belong to the policy-document owner. Retain the source identifier, correction reason and responsible person with the action.

After the correction, ask a fresh question about layering the same coat. The intended result is an answer grounded in its chest measurements and the store’s sizing advice. Confirm the answer includes the missing detail. Keep any subsequent basket action separate from the sentiment label and avoid attributing a sales increase to this change alone.

Use the guide to AI customer service to decide when a person needs to take over an unresolved exchange.

Build a useful sentiment analysis dashboard

Show reviewed topic counts with a path back to the wording and the action owner. Keep predictions awaiting review visibly separate.

A reporting row for the worked sample

Topic and sourceReviewed topic labelsSample sizeUnresolvedExample sourceAction owner
Delivery, product reviews1 negative2 reviews0R1, original reviewDelivery lead
Sizing information, written chats1 negative3 conversations1C1, original chatCatalogue owner
Quality, product reviews2 positive2 reviews0R1 and R2, original reviewsProduct lead

These rows use source-specific sample sizes, not the count of all topic mentions. The unresolved chat C3 stays visible in the chat sample until its context is recovered. A record without a topic label isn’t automatically neutral for that topic.

A spreadsheet can group reviewed rows by topic and sentiment. Power BI can import the same reviewed-label file through its Text/CSV connector. That reporting workflow uses labels already supplied in the file. Keep source identifiers as text and restrict drill-through access to the original records. The report doesn’t need to run a sentiment model to display reviewed counts.

Before sharing it, ask a colleague to reproduce a displayed topic count from its source records. They must be able to explain why R1 is mixed while its fabric label is positive. Then confirm that the assigned sizing correction answers a fresh shopping question with the intended product detail.

Give your store its own shopping agent

Customers ask in their own words. Your agent finds, compares and checks out from your catalogue, right on your site.

See it in actionFree trial · Simple setup