A candidate asks whether your company is a good employer. A customer wants to know whether your service is reliable. A journalist looks for controversies associated with your leader. In each case, an assistant can deliver a convincing portrait, sometimes well documented, sometimes shaky. Reputation is then shaped by an answer you do not publish, do not control and may struggle to find again. Looking ahead to September 2026, this shift calls for a concrete initiative from communications departments: monitoring these new touchpoints without treating every formulation as a verdict.
From search result to constructed portrait
The shift is already documented by the launch of Google’s AI Overviews in 2024, the arrival of search in ChatGPT that same year and the development of assistants such as Perplexity. These interfaces can bring sources together, select arguments and formulate a recommendation. Their precise role in future user journeys remains uncertain, but the way they work is already changing the question facing communications professionals: no longer just “where do we appear?” but “what is being said about us?”.
Strong visibility does not guarantee an accurate description. An organization may be cited frequently but reduced to an old crisis. It may also be left out of a supplier shortlist despite its recognized expertise. And an apparently favorable answer may attribute a nonexistent certification to it: a flattering error is still an error, potentially creating misleading expectations.
Measuring an answer, not a ranking
Traditional tools track press coverage, social mentions or search engine rankings. Here, the subject being measured is a moving target. The answer depends on wording, language, conversational context, accessible sources and whether web search is used. Settings or model changes can also alter the result. Two different snapshots therefore do not, on their own, prove a change in reputation.
Another complication: not all answers carry the same weight. Being absent from an exploratory list is not the same as being described as unreliable at the point of purchase. The dashboard must start with situations relevant to the organization, rather than an endless inventory of questions. For a university, these will include study choices and employability; for a telecom operator, coverage, pricing and complaint handling.
Building a panel of realistic questions
The first task is to gather questions from customer service, recruiters, sales teams and media relations. Internal site searches can also reveal recurring concerns, provided confidentiality is respected. A stable panel is then assembled, supplemented by questions related to current events. The aim is not to trick the assistant, but to reproduce plausible requests.
- Discovery: “What does this organization do, and whom does it serve?”
- Comparison: “What alternatives should be considered for this need?”
- Trust: “What are its strengths and the documented criticisms?”
- Decision: “What should I check before signing or applying?”
Questions that name the brand must be separated from those that do not. The former measure the portrait produced; the latter, its spontaneous presence within a category. Combining the two would produce an appealing but hard-to-interpret indicator. The same caution applies to leading questions: asking why a company is controversial can steer the answer toward controversies.
Repetition is part of measurement
A screenshot is an alert, not a statistic. To identify a trend, each important question must be repeated under documented conditions: the assistant used, date, language, interface, whether the user is logged into an account, whether web search is enabled, and whether the conversation is new or ongoing. The full text and displayed links must be retained, not merely a summary produced by another model.
A reasonable protocol starts with a limited scope, with several observations per question, then expands according to what is at stake. Tests conducted through an API must not be confused with the experience of the consumer interface: tools, instructions and sources may differ. Perfect reproducibility is not always achievable; traceability, however, must be pursued. Any change to the protocol must be reflected in the charts.
A dashboard with multiple dimensions
The instinct would be to reduce everything to a sentiment score. That is insufficient. “An ambitious company, but with criticized after-sales service” contains two distinct pieces of information. A critical assessment may be accurate and useful; a positive one may rest on an invented fact. The framework must therefore distinguish visibility, accuracy, associated themes, source quality and tone.
For each indicator, displaying the number of answers examined and their variability avoids a false sense of certainty. The frequency of appearance in the panel is not an actual market share of conversations. Without reliable audience or usage data, it is impossible to infer how many people actually encountered a description. The dashboard primarily measures potential exposure within a test setup.
Maintaining human oversight
A model can help classify answers, provided its decisions are checked against a sample and disagreements are documented. Serious allegations, regulatory issues and errors concerning individuals require human review. To prioritize alerts, it is better to consider severity, recurrence and the importance of the user journey involved together than to add up negative mentions.
Correcting information, not dressing up the findings
When an error recurs, the investigation begins with the cited sources, where they exist. Is an institutional page outdated? Is a subsidiary being confused with its parent company? Does an old controversy lack a public follow-up? A citation does not, however, guarantee that the source actually supports the claim: the document must be opened and the relevant passage checked.
The operational responses often remain very conventional: updating a profile, clarifying a policy, publishing verifiable information, requesting a correction from a publisher. Approaches known as GEO, or “generative engine optimization,” explore this new form of visibility. They do not guarantee control. Producing more promotional content can add noise rather than build trust.
This initiative must bring together communications, search engine optimization, customer service, legal and data teams. Each contributes a piece: perception, content accessibility, the reality of pain points, risk and methodology. Tests must avoid introducing confidential information into unauthorized services. Finally, correcting wording without addressing the problem it describes only makes the findings look better.
What next? If assistants become a more frequent step in decision-making, dashboards will need to incorporate their answers as an additional monitoring tool, not as a higher truth. The forward-looking priority is not to promise a universal “AI share of voice”: it is to build a comparable historical record, identify persistent errors and connect findings to verifiable actions. Reputation will remain collective; measuring it will require learning to work with uncertainty.