Back to all articlesCustomer feedback

Code feedback that contains several issues without losing the connections

Alactic Editorial10 min read

Prepared with AI assistance by Alactic Editorial. Examples identified as hypothetical are illustrative, not customer results.

A customer reports that checkout rejected their address, cleared their basket, and left them explaining the problem again to support. Calling the whole comment a checkout problem hides the support handoff. Adding three tags preserves more detail, but counting those tags as three affected customers exaggerates the evidence.

The practical answer is to separate three things: the original response, the meaning units inside it, and the relationships between those units. Code each meaning unit for the distinct experience it expresses. Keep every unit linked to its parent response. Then count issue mentions, responses containing an issue, and customers separately—not as interchangeable totals.

This approach needs a small annotation worksheet, not an unlimited tag list. The filled hypothetical worksheet below shows what to record, where to split a comment, and how to retain the connections that make feedback useful.

1. Choose the coding unit before choosing the labels

Use a meaning unit as the coding unit: the smallest passage that expresses a distinct experience while retaining enough context to interpret it. A unit may be a clause, a sentence, or several sentences about the same issue.

Do not make punctuation the rule. A single sentence can contain several issues. Several sentences can repeat one issue without adding another.

Keep the original response as the container. That lets an analyst return from an isolated label to the full account and lets a report count the response only once for a given issue.

  • Response: the complete submitted comment in the collection being analysed.
  • Meaning unit: a passage expressing a distinct issue, positive experience, request, or other relevant observation.
  • Code: a defined analytical label attached to a meaning unit.
  • Relationship: a recorded connection between units, such as reported cause, sequence, or contrast.

A customer is a different entity from a response. One customer may submit several responses. Unless the dataset already provides a reliable customer key and an agreed counting rule, report response counts rather than relabelling them as customer counts. Establishing customer identity is outside this coding exercise.

2. Split when the experience changes—not whenever a word changes

A useful split test is: Could this passage require a different explanation, investigation, or corrective action from the passage beside it? If yes, a separate meaning unit is usually warranted.

Address rejection and basket loss deserve separate units even when they occur in one checkout attempt. They describe different failures. Repeated references to the same rejected address, by contrast, need not become separate issue mentions.

Use these boundary rules consistently:

  1. Split distinct experiences. Separate an interface failure from a later support interaction.
  2. Keep repetition together. Rewording or emphasis does not automatically create another unit.
  3. Retain necessary context. If a passage says “it happened again,” preserve the referent in a context note.
  4. Separate praise from criticism when they concern different experiences. A helpful agent does not erase the original failure.
  5. Do not split an inseparable claim just to fit the code list. One unit can legitimately have more than one analytical label.

Distinguish recurrence from mention volume. A customer may describe several attempts at the same task. Record the reported recurrence as context or a separate field if needed; do not automatically turn each attempt into another mention.

3. Use multi-label coding without multiplying the evidence

Segmentation and multi-label coding solve different problems. Segmentation identifies separate experiences. Multi-label coding describes different relevant aspects of one experience.

For example, a unit about an agent asking someone to repeat information might carry the issue code support.repeat_information, a journey-stage label for support, and a negative evaluation. Those are three attributes of one unit—not three issue mentions.

Separate code families in the worksheet so this distinction remains visible:

  • Issue or experience code: what happened.
  • Context labels: where or under what circumstances it happened.
  • Evaluation: whether the stated experience was negative, positive, mixed, or unclear.
  • Relationship fields: how this unit connects to another unit.

Prefer the most specific supported issue code. If checkout.address_rejected sits beneath a broader checkout category, treat checkout as a roll-up rather than another independent issue mention.

Two issue codes on one unit are acceptable when the claim genuinely expresses both and splitting would damage its meaning. Count each unique unit–issue-code pairing once, but retain a separate unit count. If double-coding becomes common, review whether the codes overlap or the segmentation rule is unclear.

4. Filled hypothetical annotation worksheet

Fictional worked example: Every response, identifier, code assignment, and number in this example is hypothetical. These are not customer results. The worksheet uses one specific experience code per unit to make the counting rules easy to inspect.

Original responses

R1: Checkout rejected my address. That error reset my basket, so I had to add everything again. When I contacted support, the agent asked me to repeat the details I had already sent.

R2: My address was rejected twice at checkout; both attempts gave the same result. Support replied quickly and helped me finish the order.

R3: The parcel arrived late. Tracking showed delivered before it actually arrived.

R4: My delivery was late, and support could not explain the delay.

Worksheet header and completed rows

The passage field preserves the customer's wording. Context notes are analyst annotations, not additional customer statements. The relationship field points to other unit IDs rather than relying on row order.

Hypothetical worksheet — codebook version: pilot-A
Fields: unit | parent | passage | code | evaluation | context | relationship

R1-U1 | R1
Passage: Checkout rejected my address.
Code: checkout.address_rejected
Evaluation: negative
Context: checkout attempt
Relationship: reported cause of R1-U2

R1-U2 | R1
Passage: That error reset my basket, so I had to add everything again.
Code: checkout.basket_reset
Evaluation: negative
Context: “That error” refers to address rejection in R1-U1;
         rebuilding the basket is the reported consequence.
Relationship: reportedly caused by R1-U1

R1-U3 | R1
Passage: When I contacted support, the agent asked me to repeat
         the details I had already sent.
Code: support.repeat_information
Evaluation: negative
Context: support contact following the checkout problem
Relationship: follows R1-U1 and R1-U2; no causal claim added

R2-U1 | R2
Passage: My address was rejected twice at checkout;
         both attempts gave the same result.
Code: checkout.address_rejected
Evaluation: negative
Context: repeated attempts at the same task; recurrence reported
Relationship: precedes R2-U2

R2-U2 | R2
Passage: Support replied quickly and helped me finish the order.
Code: support.helpful_resolution
Evaluation: positive
Context: customer reports help completing the checkout task;
         speed is retained as an attribute, not another issue.
Relationship: resolution follows R2-U1

R3-U1 | R3
Passage: The parcel arrived late.
Code: delivery.late
Evaluation: negative
Context: parcel delivery
Relationship: co-occurs with R3-U2; causation not stated

R3-U2 | R3
Passage: Tracking showed delivered before it actually arrived.
Code: delivery.premature_delivered_status
Evaluation: negative
Context: status timing relative to actual arrival
Relationship: co-occurs with R3-U1; causation not stated

R4-U1 | R4
Passage: My delivery was late
Code: delivery.late
Evaluation: negative
Context: delivery delay
Relationship: R4-U2 concerns this delay

R4-U2 | R4
Passage: support could not explain the delay.
Code: support.delay_unexplained
Evaluation: negative
Context: “the delay” refers to R4-U1
Relationship: explanation gap about R4-U1

This worksheet makes two deliberate choices. Rebuilding the basket stays with basket reset as its stated consequence rather than becoming a separate failure code. Fast support and successful help remain one positive resolution unit for this analysis. If support speed were a separate research question, the codebook could capture it as a defined attribute or separate unit—but that choice should apply across the dataset, not only to this comment.

5. Count the worksheet in distinct ways

The hypothetical worksheet contains nine meaning units across four responses. Eight units describe issues; one describes a positive support experience. None of those totals establishes a customer count.

For this example, an issue mention means one unique pairing of a meaning unit and a specific issue code. Context labels, parent categories, and relationship links do not create additional mentions.

Hypothetical count ledger

Issue code                              Mentions   Distinct responses
checkout.address_rejected                  2               2
checkout.basket_reset                      1               1
support.repeat_information                 1               1
delivery.late                              2               2
delivery.premature_delivered_status         1               1
support.delay_unexplained                   1               1

Total issue mentions                       8
Responses containing any issue                             4
Positive-experience units                  1
All meaning units                          9
All responses                                              4
Distinct customers                         Not established

Do not sum the distinct-response column and describe the result as the number of responses with problems. A response can appear under several issue codes. Instead, take the unique parent response IDs across the relevant rows.

Likewise, count a category roll-up using unique response IDs. In this hypothetical example, checkout appears in two responses, support in three, and delivery in two. Those category counts overlap and must not be added to produce a total response count. The support category also includes a positive experience, so its presence count is not a complaint count.

Choose the count that answers the question. Use issue mentions to examine coded detail, distinct responses to describe how widely an issue appears in the collected comments, and relationship counts to inspect linked patterns. Label each explicitly.

6. Preserve connections without inventing causes

Keeping two codes under the same response preserves co-occurrence, but not the nature of their connection. The relationship field adds that missing distinction.

  • Reported cause: the customer explicitly attributes one event to another.
  • Sequence: one event happened before another, without an asserted cause.
  • Same experience: two units concern the same event or task.
  • Contrast or resolution: a later experience offsets, resolves, or differs from an earlier one.

In R1, the customer attributes basket reset to the address error. Record that as a reported causal link, not a verified technical diagnosis. In R3, late arrival and premature delivery status occur in the same account, but the comment does not say either caused the other.

When reporting linked issues, count the relationship once per relevant response unless the analysis explicitly uses another unit. Two reciprocal worksheet references describe one link, not two incidents. A useful summary would say that one hypothetical response explicitly connected address rejection with basket reset, then retain the passages for investigation.

7. Repair disagreements at the rule level

Before scaling the worksheet, have two reviewers independently segment and code the same small pilot set. Review boundary disagreements separately from code disagreements: analysts cannot meaningfully compare labels if they have divided the text differently.

Common failure patterns call for different repairs:

  • One unit per sentence: revisit passages containing several experiences and add boundary examples to the codebook.
  • Every repeated phrase counted again: clarify how repetition and recurrence are handled.
  • Broad and narrow codes counted together: designate parent codes as roll-ups and recalculate counts.
  • Positive support erased by an overall negative label: assign evaluation at unit level.
  • Analyst-inferred causes recorded as facts: downgrade links to sequence or co-occurrence unless the passage supports attribution.

Record the agreed rule, its rationale, and a fictional boundary example. If a rule changes, identify and recode earlier affected rows before comparing totals. Keep the original passages and a codebook version so the correction is traceable.

8. Run a small, auditable pilot tomorrow

For a practical starting exercise, choose a hypothetical pilot target of twenty responses, including comments with repeated issues, mixed praise and criticism, and linked events. This is a suggested workload, not a validated sample-size requirement.

  1. Write the meaning-unit rule and define the initial issue codes.
  2. Fill the worksheet, including parent IDs, context, and explicit relationship types.
  3. Have a second reviewer annotate the same responses independently.
  4. Log disagreements about boundaries, labels, and relationships separately.
  5. Reconcile the pilot and produce separate ledgers for units, issue mentions, and distinct responses.

Measure whether every unit traces back to its response, every relationship points to an existing unit, and every reported count can be reproduced from the rows. Record unresolved coding decisions rather than hiding them in an other category. The immediate deliverable is a checked worksheet and a short set of counting rules—not a larger total of tags.

Customer feedbackQualitative codingFeedback analysisResearch operations

Put the next conversation to work.

Explore Alactic Communications for WhatsApp conversations, campaigns and AI-assisted workflows. Choose the product and plan that fit your team.

Explore Communications