The question decides who you are compared with.

An AI visibility score can rise when a brand’s competitive position has barely changed. The investigation may simply be asking more questions where the brand already does well. Before treating an improvement as evidence of success, a communications team needs to establish what was measured, which decisions those questions represent, and whether the comparison stayed consistent. Question design belongs in the client brief because it sets what the resulting report can reasonably claim.

Define the decision before the question set

Category-level supplier questions test which businesses enter consideration. Branded questions examine a business the person asking already has in mind. Both belong in an AI/LLM visibility investigation, but they tell a team different things about the communications task.

“Which firms provide enterprise data migration?” tests whether a business appears among the options. “Can Company A migrate regulated workloads?” tests the account of an identified supplier. Pooling them without distinction hides whether the problem is discovery, credibility, or suitability for a particular job.

Conductor describes prompt generation organized around topics, audience personas, and search intents such as comparison and pricing, with separate configurations for markets. The capability can help a team build a deliberate investigation, but the team still decides which audiences and requirements to cover and what evidence supports those choices.

Sales objections, customer research, product documentation, and questions received by the press office can inform that brief. A generated persona can help organize questions. It cannot show how often real buyers ask them.

A changing question mix can move the result

Take a hypothetical supplier mentioned in 80 of 100 branded answers and 20 of 100 category answers. Across that balanced collection, its mention rate is 50 percent. In the next collection, suppose the rates within each group stay the same, but the team collects 150 branded answers and 50 category answers. The combined rate becomes 65 percent.

The combined rate rises fifteen percentage points although neither group improved. All of the change comes from the extra weight on questions that already named the supplier.

A campaign may legitimately shift attention toward buyers evaluating a shortlist. The new aggregate then describes a different question mix. A report should show the group results, explain any weighting, and distinguish a change in coverage from a change in representation.

For recurring monitoring, keep a stable set of core questions and report newly added questions separately until there is a comparable baseline. If the client’s priorities change, revise the scope openly. Swapping difficult questions for favorable ones without saying so makes the trend less informative just when leadership expects it to show progress.

Change the requirement deliberately

Some differences between answers are what the investigation is meant to find. A request for a supplier that can launch in four weeks is a different decision from a request for one capable of a complex, yearlong implementation. A different recommendation may be a sensible response to that requirement.

In a hypothetical software comparison, Brand A might be credited with extensive configuration and Brand B described as quicker to deploy. An agency should find the point where the recommendation changes, whether that is the implementation deadline, internal technical capacity, integration needs, or cost. That point shows which argument each brand can credibly make to which buyer.

Keep the competing products and conditions comparable. Testing the client against current competitors in one question and outdated editions in another confounds the comparison. Record the product, market, requirement, and original wording alongside the answer so a reviewer can reconstruct the decision being tested.

Separate substantive differences from measurement noise

The 2025 DOVE study examined combinations of prompt changes across evaluation benchmarks, including formatting and wording. It documented sensitivity across the models and tasks it tested. That supports caution about relying on one wording. It does not establish how often a particular brand’s competitive position changes in commercial AI search.

A second 2025 paper, Flaw or Artifact?, found that much of the apparent sensitivity in its experiments came from evaluation methods. Rigid answer matching could miss semantically correct responses phrased differently. Using a model-based judge reduced variation in measured performance across prompts. The finding applies to those evaluation settings and does not prove that all wording effects disappear.

For communications analysis, the practical lesson is to check meaning before declaring a narrative change. “Strong implementation support” and “helpful deployment specialists” may describe the same proposition. Adding “only for smaller organizations,” on the other hand, changes the claim even if most of the words stay the same. Automated grouping needs to be checked against those distinctions.

Audience framing has a limit

Asking a model to answer for a procurement director examines the answer produced under that framing. It does not measure what procurement directors believe, how they search, or whether they saw that answer.

Repeated collection can help establish whether a finding persists under comparable conditions. Record the model or service, collection period, market settings, and relevant context. A larger answer set can reduce uncertainty within the chosen questions while leaving an important audience entirely unexamined.

The aim is to cover the decision first, including any requirement that decides the purchase, then repeat enough to assess the pattern.

Commission the work the finding supports

Before collecting more answers, an agency should agree with its client on the decisions being examined, the audiences and products in scope, the comparison set, and the conditions that could change the recommendation. It should also state how results will be grouped and which questions remain exploratory.

For Google’s AI features, Google’s own guidance emphasizes established SEO practices, accessible content, and helpful information. It specifies no additional technical requirements for AI Overviews or AI Mode and does not guarantee inclusion. That guidance applies to Google’s systems and not necessarily to other answer engines.

A finding of prompt variation is weak grounds for producing many pages around paraphrased questions. If a finding concerns missing evidence for a capability, the next LLM optimization assignment should specify what evidence is needed and where it belongs.