After the work, did AI’s description change?

After a communications team has corrected a product page, published supporting evidence, or clarified an executive’s role, the next report should help it decide whether to continue, refine, or stop that work. A new citation alone cannot settle the decision. The team needs to know whether AI now gives a more accurate account of the relevant company, product, or person, and whether the remaining gap calls for a different response.

An evidence-backed before-and-after comparison serves that purpose: it connects the original finding, the work completed, and the later observation closely enough to support the next decision.

Start with the change the work was meant to achieve

Consider a hypothetical software company. Its audit-export feature was previously available only on Enterprise. The company has now added it to Team, while Starter still excludes it. Its agency is helping communicate the change to buyers comparing plans.

The intended result is precise: answers about current availability should identify Team and Enterprise and preserve the Starter exclusion. More favorable language about the company would not establish that result. Nor would a recommendation that mistakenly gives every plan the feature.

Before the work begins, keep the relevant answers, questions, product names, source passages, and current website pages. Agree which finding the follow-up will examine. Our article on turning AI visibility into a communications brief explains how to commission the initial work; the follow-up tests whether that assignment still fits the evidence.

Use the approved website as the primary reference for the company’s published offer. The client should not have to maintain a second fact database for the investigation. If the pricing page and help center disagree, resolving that conflict becomes part of the assignment. Picking the more convenient page without recording the disagreement would leave it in place.

Measure after implementation, within a defined window

In this example, the follow-up belongs after the revised plan pages are published. If product approval delays the release, counting from the agency’s original recommendation date would measure a period in which the intended change had not happened.

From our client work, we use roughly two to eight weeks as an implementation planning range; three to four weeks is often a useful starting point for discussing a comparison. The agreed interval should follow the actual work. It is neither a universal optimum nor a forecast of when a model will change.

The time between reports is separate from the time spent collecting each report’s answers. Gather each set within a concentrated, defined research window and state its actual dates. A collection spread across several weeks may mix different stages of the product launch. It needs to be described that way.

Retain a comparable core of questions and model configurations. Add questions when the brief develops, but identify them as expanded coverage. The question-design article explains why changing the mix can move an aggregate result. Here, the practical consequence is that a newly added question cannot show improvement against an answer never collected.

Separate a changed answer from a changed reference

Suppose the baseline correctly said, “Audit export requires Enterprise.” After the Team launch, the follow-up says exactly the same thing and cites the same historical FAQ.

The answer has persisted. The company’s offer has changed. A report should say that the observed statement now conflicts with the updated reference, rather than claim that AI developed a new negative narrative.

Now suppose the follow-up says, “Audit export is available on Team and Enterprise,” with a link to a different page. Both the qualification and the supporting evidence deserve review. If they hold, the observed account has become better aligned with the current offer in those comparable answers.

A third outcome is less consequential: the citation changes, but the answer still restricts the feature to Enterprise. Source churn has occurred; the substantive issue remains. Commissioning a new content program because several URLs moved would miss the unresolved finding.

Separating the three outcomes tells the agency whether to keep clarifying the launch, investigate an old source, or move its attention elsewhere.

Keep the subject and the missing message in view

The follow-up also needs to preserve who each statement concerns. A capability of the company’s older product does not establish the same capability in its current product. An executive’s experience at a previous employer should not become a delivery record for the present company. Clearer wording can still carry the wrong attribution.

For the hypothetical supplier, an answer that praises the company’s “audit capabilities” may leave the plan question unanswered. The follow-up should examine the product and plan assignment, including the condition that changes the buyer’s choice. Our article on reading the qualification develops that distinction.

An intended message can also be absent without being contradicted. Assess the new Team availability only in questions where that information is relevant, such as plan eligibility or feature comparisons. Its absence from a question about the CEO’s background is not a message gap.

Keep three outcomes separate: present, assessed and absent, or unassessable. A failed collection or incomplete analysis cannot count as proof that the message disappeared. The report should give the relevant question coverage so the client can see how far the finding extends.

Use search comparisons to sharpen the next action

Where comparable configurations are available, answers without requested live web search can be compared with search-enabled answers. In the hypothetical case, the former might retain the Enterprise-only account while the latter accurately describe Team. That would be a useful difference to investigate and track.

Search availability and actual search use are separate facts. Anthropic’s web-search documentation explains that Claude can decide whether to search when the tool is available. A comparison therefore needs the requested mode and the available record of what happened. A missing citation list alone does not establish that no retrieval occurred.

Treat the results as observations. An answer without requested search does not expose identifiable training examples. Different models or versions also introduce differences beyond search, so those comparisons need to be described accordingly. Clear current material gives the team something defensible to publish and assess; it does not grant control over a provider’s training or guarantee adoption in a future answer.

Close with a decision the team can defend

Even when the updated Team page supports a later answer, it does not prove that this page caused the change. The research paper Correctness is not Faithfulness in RAG Attributions distinguishes whether a citation supports a statement from whether the model actually relied on that source. Its distinction helps explain why passage verification and causal attribution need separate treatment. Our source-analysis article shows how to examine the observable trail.

For the agency and client, the follow-up should resolve a short set of decisions:

• If the plan distinction is represented accurately in comparable answers, retain the clarification and agree what would trigger another review.
• If current pages still conflict, ask the product owner to reconcile them before commissioning broader promotion.
• If a source still presents the old restriction as current, check its wording and prepare a documented correction request where warranted. An accurately dated historical page may simply need clearer current evidence alongside it.
• If the evidence or comparable coverage is insufficient, identify the unanswered question and the smallest useful follow-up.

An improvement after publication is an observed outcome. Crediting the intervention requires further evidence.

A useful investigation combines a brief shaped around the client’s decision, evidence, reviewed interpretation, and a comparison that preserves the findings from both periods. Judge its value by the decisions it makes clearer: what to correct, what to substantiate, what to investigate, and what to leave alone.

Our three-product example illustrates how claims, conditions, and evidence can be examined together. It uses anonymized October 2026 answers and is a single-period example, not a demonstration of change. See example report.

Narrella delivers the baseline report and a call with the analyst who wrote it, so your team can question the findings and discuss priorities. To explore a baseline and an agreed follow-up around your implementation cycle, bring the decision and the material in scope. Discuss your brief.