Asking AI systems about your brand can reveal useful blind spots, but a prompt set is a way of looking, not a complete model of the market.
A marketing team runs its first AI visibility audit on a Friday afternoon. Someone opens ChatGPT, someone else opens Perplexity, and a third person keeps a spreadsheet. They ask twenty questions that sound reasonable: what the company does, who its competitors are, what alternatives exist, which vendors help with the category, whether the brand is credible. A few answers are encouraging. A few are irritating. One answer names a competitor that nobody internally takes seriously. Another describes the company with language from an old product page. The team screenshots the strangest results and posts them in Slack. By Monday, the screenshots have become evidence.
The founder wants to know why the company “ranks below” a competitor. Sales wants to use the good answer in a deck. Marketing wants to rewrite pages around the prompts where the brand was absent. Someone asks whether there is now a ChatGPT ranking report the way there used to be a keyword ranking report.
This is where AI visibility work can go wrong. The audit was useful. The interpretation was too heavy.
A prompt audit is not a map of the market. It is a lens. It shows how a small set of questions, asked in a particular way, through particular systems, at a particular moment, can cause the public web and model behavior to assemble a version of the brand. That version may be revealing. It may also be unstable, incomplete, and overly dependent on the wording of the question.
The mistake is treating the answer as if it were a fixed position.
Prompts are not keywords with a new costume
It is tempting to think about prompts the way marketers learned to think about keywords. A keyword had search volume, intent, ranking position, click-through rate, and a page that could be optimized. The model was imperfect, but it gave teams a stable working fiction: this query exists in the market, this page ranks here, this competitor ranks there. Prompts do not behave as neatly.
A buyer might ask an AI system a long, messy question that contains context, constraints, preference, uncertainty, and memory from earlier turns. They might ask for “good options for a mid-market team that already uses HubSpot and does not want another reporting dashboard.” They might ask for alternatives to a competitor but exclude agencies. They might ask the same thing twice with different phrasing and get a different emphasis. They might ask the model to explain a vendor to a CFO, then to a technical lead, then to a procurement team.
A keyword is usually a fragment. A prompt is often a situation.
That difference matters because the answer system is not merely matching a query to documents. It is interpreting the task. ChatGPT Search can decide to search the web depending on the user’s question and can return answers with links to web sources. OpenAI’s documentation also says ChatGPT may rewrite a user’s prompt into one or more targeted queries sent to search providers. Google describes a related pattern in AI Overviews and AI Mode: “query fan-out,” where multiple related searches may be issued across subtopics and data sources to develop a response. Google Search Central gives this as part of how these AI features can assemble broader answers.
The prompt the user typed is therefore not always the search the system performed. It is the front door to a small investigation.
A brand can be absent from the visible answer for reasons that have little to do with one clean “ranking.” The system may have expanded the question toward a subcategory where the brand has weak evidence. It may have chosen sources where competitors are better represented. It may have interpreted the prompt as asking for software when the company sells a managed service. It may have used a public list that omits the brand. It may have decided the answer needed general education rather than vendor recommendations.
The output is still useful. It is simply not a ranking table.
The shape of the question changes the company that appears
A small wording change can expose a large positioning problem.
Consider a composite example from B2B service categories. A company helps enterprise brands audit and improve how AI systems describe them. When the prompt asks for “AI visibility audit providers,” the company appears occasionally. When the prompt asks for “GEO agencies,” it disappears behind SEO firms. When the prompt asks for “brand perception research for AI search,” it appears again, but the answer places it next to market research vendors. When the prompt asks for “tools to track ChatGPT mentions,” it vanishes, which may be correct because the company is not really a software tool.
None of these answers is the whole truth. Together, they show the category boundary.
This is the real value of prompt audits. They reveal the words under which the market, or at least the answer environment, can recognize the company. They also reveal the words under which the company becomes someone else.
If a brand only appears when the prompt uses its exact internal category language, the category is probably too fragile. If it appears for broad prompts but is misdescribed each time, the public source trail may be too vague. If it appears for software prompts even though the business sells a service, the website may be overusing platform language. If it disappears whenever the buyer adds a specific constraint, the company may lack public proof around that use case.
A prompt audit is useful precisely because it is sensitive. The sensitivity becomes dangerous only when the team pretends it is stability.
One answer is anecdote; repeated distortion is evidence
A single bad answer is easy to overinterpret.
The model may have retrieved an odd source. The prompt may have been worded poorly. The session may have carried context from earlier conversation. The tool may have decided not to search. The answer may change the next day. AI systems are variable enough that one screenshot should not become strategy. Repeated distortion is different.
If several systems place the company in the wrong category, that is a signal. If different phrasings keep surfacing the same competitor set, that is a signal. If citation-based tools keep using the same third-party source, that source matters. If the company appears in branded prompts but not in category prompts, the market-facing evidence may be thin. If the AI answer gets the company’s current offer wrong in the same way a human buyer does, the problem is probably not only the model.
The team should look less like a rank tracker and more like a field researcher. The question is not “what is our position?” The better question is “what kinds of questions cause the system to understand us correctly, and what kinds cause it to drift?” That drift is often where the work lives.
A good prompt set has rough edges
Clean prompt sets are suspicious.
If every prompt is written in the company’s preferred language, the audit will flatter the brand. It will test whether AI systems understand the company when the buyer already talks like the company. Real buyers do not behave that way. They use old category names, competitor names, confused terms, procurement language, shorthand from analyst reports, and phrases picked up from colleagues.
A useful prompt set includes some awkwardness. It asks the question the founder dislikes. It uses the competitor’s framing. It tests the term the sales team hears even though marketing avoids it. It includes “agency” if buyers keep saying agency. It includes “software” if the market keeps assuming software. It includes “alternatives to [competitor]” because buyers often know the competitor before they know the category.
The work is not about creating a clean taxonomy. It is about approximating the messy ways people ask for help before they have learned the vendor’s language.
This is also why prompt audits should include human interpretation. A spreadsheet can record whether a brand appeared. It cannot easily tell whether the answer would help a buyer make a decision, whether the description feels current, whether the competitor set is commercially meaningful, or whether the answer sounds confident because it has evidence or merely because the model writes confidently.
Prompt data needs editorial judgment. Otherwise it becomes another dashboard pretending to know more than it does.
The cited sources are often more important than the answer
In citation-forward systems, the answer is only half the artifact. The sources are the other half.
Perplexity describes itself as searching the internet in real time and summarizing information from sources, with citations that allow users to verify the answer. Its help center emphasizes source-backed responses. When a system like this recommends a competitor, the useful question is not only why the competitor appeared. It is which source made the competitor easier to recommend.
Sometimes the answer cites a directory where the competitor has a complete profile and the company is absent. Sometimes it cites a category article that uses the competitor’s language. Sometimes it cites the competitor’s own page because that page is the clearest answer to the prompt. Sometimes the cited source is old, weak, or only partly relevant, but it is still the material the system found.
A prompt audit that stops at mentions misses this layer. The citation path can show the surfaces that matter in a category: review platforms, industry lists, comparison pages, partner pages, analyst writeups, documentation, customer stories, or community discussions. These surfaces may be more actionable than the answer itself.
The answer says what happened. The sources hint at why.
Prompt audits should change public material, not just the prompt bank
The weakest outcome of an AI visibility audit is a larger spreadsheet.
A team tests prompts, records appearances, calculates a rough share of voice, and then repeats the exercise next month. The measurement becomes the work. The public materials remain the same.
A useful audit should change the artifacts that answer systems and buyers encounter. If prompts expose category confusion, the homepage and service pages need clearer nouns. If competitor prompts reveal absent proof, the company may need case studies or third-party evidence. If citation paths keep pointing to directories, public profiles need attention. If the AI summary uses old language, the source trail needs cleanup. If the answer consistently misunderstands the service model, the site probably has the same problem for human readers.
The prompt set is not the asset. The improved public explanation is the asset.
There is no responsible way to guarantee that a particular AI system will mention a brand in a particular answer. The systems are too variable, and the mechanisms are too opaque. But a company can make the accurate version of itself easier to retrieve, easier to summarize, and easier to verify.
A prompt audit is valuable when it points toward that work.
The lens should be kept, but not worshipped
AI visibility teams need prompt audits because otherwise they are guessing. The answer environment is now part of how buyers encounter brands, and companies should know what happens there. Ignoring it would be naive.
The danger is turning the audit into a false map. A map implies a stable territory. A prompt audit is more like looking through a series of windows. Each window shows a part of the landscape, from a particular angle, in particular light. If several windows show the same broken fence, you should probably go outside and fix it. If one window shows a strange shadow, you should check before rebuilding the house.
That is the discipline. Use prompts to notice. Use sources to investigate. Use patterns to prioritize. Then change the public material that made the pattern possible.
The answer is not the market. It is a clue about how the market may be learning to ask.