Empty States and Error Copy: Test Before You Ship, or Trust the Design Review?
On this page
A product lead needs to approve recovery copy that describes the real problem and the available next action. Subconscious can compare complete, accurate messages under controlled conditions and connect the modeled response to a bounded copy recommendation. The accepted brief should name the error state, exact words, response, uncertainty and owner who can ship the change.
Relevant public message-selection proof comes from River City Federal Credit Union. A comparison using 140 modeled counterparts identified the staff-proposed slogan as weakest; that slogan later performed worst in Facebook and Instagram testing. The record supports the weakest-message ordering. The reported campaign improvement does not isolate a slogan effect, and the case does not validate error recovery, comprehension or trust. The useful lesson is to inspect a proposed message before committing around an internal favorite.
Establish the recovery facts before comparing words
A design review should first verify what actually failed, whether entered information survives and which recovery actions work. No preference result can justify promising an unsaved draft, an unavailable retry or a refund that the product cannot provide. Correct an established factual or accessibility defect directly rather than asking a panel whether the defect is popular.
Nielsen Norman Group's error-message guidance recommends visible, accessible indicators, familiar language, a precise problem description, constructive recovery advice and preserved input. Severity should reflect actual impact. That guidance supplies design criteria, not a measured abandonment reduction for a particular sentence.
"Help Users Recognize, Diagnose, and Recover from Errors."
Nielsen Norman Group, "Error-Message Guidelines" (source)
An empty state needs its own factual check. An account with no projects may need an accurate explanation and a working create-project action. An error after an attempted action needs the actual failure and recovery path. Different states can require different copy even when the visual component is shared.
Separate the response the research can measure
The Subconscious methodology connects defined attributes, assigned complete profiles and generated choices to estimated contrasts with uncertainty. For a copy decision, agree the exact stimulus, response and analysis before execution. A comparison of two accurate messages can support a modeled message recommendation without pretending to observe a person recovering from the error.
- Message preference asks which version a respondent selects under the stated task. Side-by-side selection differs from assigning one version and observing a subsequent response.
- Comprehension asks whether the reader understands the problem and available action. Agree the question, scoring rubric and delivery path explicitly; a choice task does not automatically produce validated comprehension scores.
- Proposed reasons can identify a confusing phrase or possible interpretation. Generated explanations remain hypotheses about actual people.
- Actual recovery concerns a real task attempt, such as correcting a field and submitting successfully. A generated choice or before/after movement alone does not establish that outcome or its causal attribution.
The advantage for the product lead is an inspectable comparison tied to a decision, with unsupported claims removed before launch. Existing adequate usability or operational evidence may already settle the change. Additional research belongs to a specific unresolved question, rather than an automatic sequence of studies.
Keep three error situations distinct
| Error type | What happened | What the copy needs to do |
|---|---|---|
| Recoverable | Form validation failed | State the specific field problem and the available correction |
| Transient | A request timed out | Describe the known state and offer a safe, available recovery action |
| Severe | Data loss or blocked access has consequential impact | State the verified severity and recovery path; a routine authentication retry may be less severe |
A generic message can obscure those differences. Check the actual system state before using reassurance. A timeout does not necessarily mean an operation failed: a payment or submission may have completed despite the missing response. Do not offer an unsafe repeat action without establishing what the system can verify.
Work through one recovery-copy commitment
Consider fictional TaskCove, a project tracker. The proposed change concerns a failed project-name save. The example assumes engineering has verified that the name must contain at least one non-space character, the attempted name remains in the field, no project has been created and editing then saving is safe. The field has an accessible label, the error is programmatically associated with that field and the same Save button remains available. These are conditions of this fictional brief, not claims about a deployed product.
The product owner must choose between two accurate messages:
- A: “Project name is required. Enter a name, then select Save.”
- B: “The project was not created because the name is blank. Enter a project name, then select Save.”
Both versions show the same empty-name field, label, layout and Save button. Neither promises automatic recovery or adds a new feature. A real defect in input preservation or the Save action must be corrected separately. Hold the actual workflow and every other interface detail fixed so the comparison concerns the complete message wording.
The proposed modeled audience represents US adults who create and name projects at work, subject to supported population construction. A configured role does not verify recruitment of actual TaskCove users or accessibility needs. Define the task context and supported response format before launch.
Randomize A/B screen positions in otherwise identical complete screenshots. Ask which screen better indicates the next step the respondent would take, with an explicit “Neither message gives me a clear next step” option. The measured outcome is a generated choice among those options. The experiment does not observe an actual correction, successful save, reading accuracy or time to recover. Preserve the exact instructions and screenshots; a preference for B cannot by itself establish that B improves completion.
Request counts, the defined B-versus-A modeled-choice contrast, neither-option responses and uncertainty appropriate to repeated tasks. Record the population configuration, assignments, run version, response artifacts and analysis. Avoid selecting the largest estimate without considering the baseline and unresolved uncertainty. No completed run or winning message is supplied here.
Turn the comparison into an accepted copy packet
Before results, the product owner sets the smallest useful modeled preference, acceptable uncertainty, implementation cost cap and unacceptable misunderstanding. Engineering confirms the recovery facts; the accessibility owner checks the actual component. Existing human or product evidence can address comprehension or completion when those outcomes control the commitment. Agree any additional check for the particular remaining uncertainty and downside.
Approve B for the bounded wording change when the prior comparison rule, available evidence, accessible implementation and cost conditions support that decision. Retain A when the change lacks useful support or the extra copy creates a material problem. Leave the choice unresolved when the neither-option response, uncertainty or a factual gap could reverse the recommendation. A low-risk wording decision and a severe data-loss warning can require different evidence because their consequences differ.
The delivered Subconscious brief places the exact approved message beside the error condition, unchanged field and button, audience, comparison, counts, uncertainty and acceptance rule. The product owner accepts the copy; engineering receives the exact string and verified recovery behavior; the analytics owner receives the actual correction and successful-save events to inspect after release. Modeled choice remains separately labeled from those observed events.
For a causal claim about actual recovery, use an appropriate assigned live comparison and account for competing changes. Monitoring can reveal a problem without isolating its cause. Preserve input, accessible communication and a truthful next action regardless of which wording the modeled audience favors.
Bring the error state to a decision review
Book a Subconscious decision review with the actual error, verified recovery actions, exact copy alternatives, intended audience, existing evidence, release owner and cost of a wrong next step. Agree the comparison and acceptance rule before commissioning the change. The useful output is an approved recovery-copy packet, a supported decision to retain the current wording or an exact unresolved question.