An AI answer is an observation, not a permanent ranking

Build a repeatable observation panel for AI discovery so the team can distinguish changing answers, inaccurate claims and useful commercial signals without inventing a universal visibility score.

The founder asks an AI assistant for recommended suppliers and finds a competitor. The marketing team tries again and gets a different list. A third attempt includes the company, but describes a service it no longer offers.

Each answer is worth inspecting. None is a sufficient basis for declaring that the brand has gained or lost a permanent position in AI discovery.

The CMO needs a method for observing a changing experience and deciding which findings deserve action. Begin by recording what was actually tested rather than treating a screenshot as a market-wide result.

Specify the experience

Record the product, visible mode or model label, date, language, location where relevant and exact question. Note whether the session had previous conversation and whether web search was visibly used.

You may not control every condition or know every system detail. Record those limitations instead of implying laboratory precision. A consumer interface, a developer API and a third-party monitoring tool may not reproduce the same experience.

Google’s AI-features documentation describes expanded querying and notes that its AI experiences can produce different responses and supporting links. This is a reason to define the tested surface, not a licence to claim that every variation has a known cause.

Keep observed facts separate from explanations about how a particular answer was generated. The latter may remain unknown.

Use questions drawn from buying decisions

Build a small panel from actual customer questions, sales objections and comparisons. Include problem discovery, approach selection, provider evaluation and factual questions about the brand.

Avoid a panel dominated by prompts that name your company. Those can assess how the system describes the brand, but they do not show whether unfamiliar buyers encounter it while exploring a category.

Also avoid questions written only to make the brand appear. A monitoring panel should resemble decisions customers might make, not a competition to produce flattering screenshots.

Record how each question was selected. A sample drawn from ten sales conversations has a different meaning from a statistically representative view of market demand. Both can be useful if described honestly.

Repeat without chasing a desired answer

Set a modest repeat schedule and keep the questions stable long enough to compare observations. Save the full answer, visible sources and any clear errors, not just whether the company appeared.

Decide in advance how many repetitions you will run. Otherwise the team may keep asking until the brand appears and present that final answer as typical.

Use additional exploratory questions separately. They can reveal opportunities, but changing the panel constantly makes a trend difficult to interpret. Version the panel when customer needs or the product scope change.

Respect the service’s access rules when choosing manual or automated collection. The quality of the observation matters more than producing an enormous unrepresentative prompt count.

Code the answer for useful distinctions

Separate a brand mention from a recommendation, a cited link and a correct product description. An answer can name a company negatively, cite its research without recommending its service, or recommend it without a direct link.

Record errors by type: obsolete pricing, unavailable geography, invented features, misattributed research or confusion with another company. These categories point toward different corrective work.

For citations, inspect the destination and the supported claim. A visible link does not establish that the answer accurately represents the page or that a user followed it.

The output should help prioritise repairs and investigations, rather than reduce every observation to a single score whose assumptions disappear from the board report.

Repair what the company controls

If an answer repeatedly reflects an obsolete public page, update that page and other owned materials that repeat the error. Make essential product and company information clear, consistent and easy for people to find.

Google’s generative AI search guidance emphasises useful original information and technical clarity. Those are documented principles; a promise that a particular edit will secure citations is not.

If the source of an error is an independent publication, use its normal correction process where appropriate. Do not rewrite your own content into an unsupported claim merely to compete with the answer.

Some errors may persist despite accurate public information. Record them and assess customer impact rather than promising immediate control over an external system.

Report a panel, not the entire market

A defensible summary says how often a defined event appeared within the recorded sample and what changed under the same method. It does not call that figure total market share or guaranteed customer exposure.

Pair the observation with a decision: correct product information, improve a useful comparison page, investigate a source, or gather evidence from actual buyers. Keep commercial outcomes in their own measures.

The aim is a repeatable view of relevant questions and avoidable information problems. That gives the CMO something actionable without pretending that a changing answer interface behaves like a fixed league table.