AI visibility tools can produce what look like reassuringly precise figures. They can tell you how often a brand appears, where it ranks and which competitors are recommended instead.
I’m not going to reiterate the flaws in AI visibility reporting, but if we assume the findings are directional then these figures are only useful if the questions being tested resemble the questions that commercially important audiences might actually ask.
This is the discipline too often missing from generative engine optimisation, or GEO. Companies invest in tracking platforms, upload a list of prompts and begin reporting percentage scores before establishing whether those prompts reflect genuine customer needs or buying decisions.
The result can be accurate measurement of completely the wrong thing.
Prompt selection is research design
In traditional search, marketers have years of keyword data showing what people type into Google. AI search is different. People can describe their circumstances, add constraints, ask follow-up questions and request a recommendation tailored to their needs.
Someone choosing a PR agency is unlikely to ask only:
What is the best PR agency?
A more realistic question might be:
I am the marketing director of a UK technology company. I need a mid-sized PR agency with international reach, strong media relations and a credible approach to evaluation. Which agencies should I consider?
The second prompt reveals far more. It establishes the audience, sector, location, requirements and stage of the buying process. It tests whether an AI system associates an agency with the qualities that could influence a real appointment.
This means building a prompt library is not an administrative job to be delegated to a tracking platform. It is a form of audience and market research.
Stop turning keyword lists into prompts
Keywords remain useful evidence. Search data can reveal recurring needs, topics and language. But adding a few words to a keyword does not necessarily create a useful AI prompt.
“Best employment law provider UK” may have value, but it does not describe the different circumstances in which an employer might seek support.
A growing business appointing its first HR adviser has different needs from a national organisation with an established internal HR team. A company facing an employment tribunal will ask different questions from one comparing long-term compliance providers.
Those distinctions affect which organisations are recommended and how their strengths are described.
A useful prompt should contain four elements:
- Audience – who is asking the question?
- Need – what are they trying to achieve or resolve?
- Constraints – what conditions affect the answer, such as sector, geography, budget, company size or risk?
- Decision stage – are they exploring a problem, comparing approaches, selecting a supplier or checking a named brand?
Not every prompt needs to be long. It does need to represent a recognisable situation.
Measure journeys rather than isolated answers
AI answers vary. The same system can produce different recommendations when a question is repeated, while relatively minor changes to the wording can alter the result substantially.
This makes it dangerous to celebrate one favourable answer or diagnose a serious problem from one omission.
Prompts should instead be organised around audience journeys:
- Discovery – understanding a problem and the available solutions.
- Consideration – identifying suitable approaches or suppliers.
- Selection – comparing providers against relevant criteria.
- Reputation – checking a named organisation’s expertise, record or weaknesses.
- Risk – looking for complaints, controversies or reasons not to buy.
Patterns across these groups are more informative than the result of any individual prompt.
A brand might perform well when users ask for a list of suppliers but disappear when specialist expertise is added. It might be recommended frequently but described in terms that do not support its intended position. It may be visible during discovery but absent when the user becomes ready to appoint a provider.
Each finding demands a different communications response.
Separate visibility from reputation
Being mentioned is not the same as being well represented.
An AI answer could name a business prominently while presenting it as expensive, unsuitable for smaller companies or weaker than a competitor in a strategically important area. A simple visibility score might record that as a success.
Prompt research must therefore examine two related questions:
- Does the brand appear in relevant answers?
- How is the brand described when it appears?
The second question is particularly important for PR. AI systems form answers using information from company websites, media coverage, reviews, directories, forums and other third-party sources. Weak or inconsistent representation may reflect a wider problem in the information available about the organisation.
Publishing more website copy will not necessarily correct it. The intervention might require stronger earned coverage, clearer evidence of expertise, better customer proof or action on a genuine reputational weakness.
Avoid the false comfort of a single score
Different AI platforms use different sources and produce different answers. Their audiences and uses also vary.
Combining every result into one headline visibility percentage can conceal more than it explains. A brand may perform strongly in one system and poorly in another. It may also appear consistently for low-value informational prompts while remaining absent from commercially important recommendations.
Reporting should therefore distinguish between:
- platforms;
- audiences;
- topics;
- stages of the buying process;
- products or services;
- neutral, positive and negative questions;
- brand appearance and brand description.
Commercial importance should also be considered. Being absent from a question commonly asked by high-value buyers should carry more weight than appearing in a broad question with little connection to a sale.
Build consistency without standing still
A useful prompt library needs a stable core. If the questions change constantly, it becomes impossible to tell whether performance has improved or whether the research has simply become easier.
However, the library cannot remain fixed indefinitely. Customer concerns change, new competitors enter the market and campaigns create new subjects to test.
The answer is to maintain two sets:
- A core library used consistently to monitor meaningful changes over time.
- An experimental library used to investigate new products, issues, campaigns and audience behaviour.
The core provides comparability. The experimental set keeps the research relevant.
Both should be reviewed by people who understand the customer, the market and the commercial consequences of the findings.
AI visibility is not an SEO score with a new label
AI visibility measurement can be valuable, but only when it is connected to the wider evidence.
Search analytics can show what audiences look for and how they behave. AI tracking can show how brands appear within generated answers. PR evaluation can reveal the third-party information, associations and reputation signals that may be influencing those answers.
Taken together, this helps organisations distinguish between several different problems:
- The brand is not recognised.
- The brand is associated with the wrong subjects.
- There is insufficient independent evidence to support its claims.
- Competitors have stronger authority in the sources AI systems use.
- The brand appears, but its reputation could discourage consideration.
- The prompt set itself bears little resemblance to the real market.
That final possibility should always be tested first.
AI visibility data can look scientific. But if the prompts have been generated without proper audience research, commercial judgement or a clear hypothesis, the decimal places are simply adding polish to a weak methodology.
Before asking how visible your brand is, ask a more fundamental question: visible for what, to whom and at which point in their decision?
