Most businesses start by asking what ChatGPT says about them by name. That is the wrong starting point, because customers usually ask unnamed questions about the kind of company they should buy from. This page is about those buying questions, not sector research such as market size, trends or regulation, and how they shape a company’s AI visibility.

What questions do people ask ChatGPT about a business?

People ask AI full questions, not short keyword strings. Most ask about the kind of company they need, the problem they want solved or the choice they are trying to make, rather than naming a company. Whether they phrase it as a question about a business or an industry, the complete set of buying questions is what a company should track.

AiDisco calls that complete set the Prompt Universe: the full set of questions real customers ask AI assistants about a category, spanning that category’s real buying stages. It is the scoreboard. If a question is not part of that set, appearing for it does not represent a meaningful visibility win.

The set can include discovery, problem, evaluation, comparison, risk, price and final-choice questions. Someone looking for accounting software might ask, “What is the easiest accounting software for a five-person company?”, “Which tools connect to Shopify?” or “What should I check before moving from spreadsheets?”

AI responds to the whole question. “Best payroll software” and “best payroll software for a company with contractors in three countries” may produce different answers because the second question adds a use case, operating model and risk profile.

How do you find out what customers ask AI about a business?

Start with real customer language from sales calls, support conversations, site search, search-query data, forums, reviews and competitor pages. Group the questions by what the buyer is trying to decide, then test the wording in several AI assistants. The aim is not to collect every imaginable sentence; it is to map the recurring decisions that could lead someone towards or away from a company.

The concept is general, but the working set is always built for one category at a time. A plumbing company and a payroll platform share the structure of discovery, evaluation and decision questions, but none of the useful content. One set asks about burst pipes, emergency response and local coverage; the other asks about tax jurisdictions, integrations and contractor compliance. The questions must be assembled for the category rather than borrowed.

A practical build has five steps:

1. Collect questions customers ask before contacting or buying from a company.
2. Extract repeated objections, decision criteria and comparison language.
3. Review search data for phrases that signal the same decisions.
4. Expand each decision into realistic conversational variants.
5. Remove questions that are informational but unlikely to influence a purchase.

A built set is not a one-time asset. It goes stale as the category changes, new products appear, regulations shift, competitors reposition and customers adopt new language. The idea remains valid, but each working set needs regular review.

Do the questions people ask AI change as they learn more?

Yes. Early questions tend to define the problem or identify possible approaches. Later questions become more specific about suitability, evidence, cost, risk and trade-offs. The sequence is rarely a neat funnel, but the wording usually becomes more constrained as the customer learns what matters and rules out options.

Consider someone planning residential solar. They might begin with “Is solar worth it for a house with high daytime electricity use?” Later they may ask, “What size system suits a four-bedroom home?” Evaluation questions follow: “What warranties should a solar company provide?” or “Is a battery worthwhile without an electric vehicle?” Near a decision, the questions narrow to installation timing, finance, local aftercare and two-company comparisons.

The important point here is that the questions change. The separate issue of visibility at the point of buying measures whether a company remains present as those questions become more commercially decisive; it belongs on its own page and should be linked when that page is published.

How do you know which questions are worth showing up for in ChatGPT?

A question is worth tracking when a credible answer could influence which company the customer considers, trusts, compares or buys from. High-value questions reveal a need, constraint, decision criterion or purchase risk. Questions that are popular but unrelated to choosing a company may generate attention without improving commercial visibility.

Use four tests:

• Relevance: Could the company genuinely satisfy the need?
• Decision influence: Could the answer change the shortlist or next action?
• Specificity: Does the question give AI enough context to distinguish suitable companies?
• Frequency or importance: Is it asked often, or is it rare but commercially decisive?

A mention for “What does this company do?” is less meaningful than an appearance for “Which companies can solve this problem under these constraints?” The separate calculation of how often AI names a business belongs on its measurement page and should be linked when published.

This page defines which questions belong on the scoreboard; it does not explain the scoring formula. For the wider selection mechanics, see how AI recommends businesses.

How many questions does a business need to show up for?

There is no universal number. A business needs enough questions to cover the meaningful ways customers discover the category, describe their problem, evaluate options, compare companies and make a final choice. Ten broad questions are usually too shallow; thousands of near-duplicates create noise without adding useful coverage.

A narrow local category may need dozens of distinct questions, while a complex national or international category may need hundreds. The useful unit is not the raw count but the number of different decisions represented. “Best payroll software”, “top payroll platform” and “recommended payroll system” are wording variants; “payroll software for multi-country contractors” represents a different need.

Cover each buying stage, major use case, customer type, location constraint, price concern, risk concern and comparison pattern. Add variants only when they reflect real customer language or materially change the answer.

Does it count if ChatGPT mentions a company only when asked about it by name?

No. If the question already contains the company name, a mention proves only that the model can recognise or repeat the name supplied in the prompt. The real test is whether the company appears when the question names nobody and asks for a suitable business, product or choice.

AiDisco separates branded, non-branded and comparison prompts. A branded prompt names the company being checked. A non-branded prompt names no company and tests whether the assistant selects it without being led. A comparison prompt names two rivals and reveals the head-to-head framing the assistant uses.

Only non-branded questions provide a clean test of whether a company enters the consideration set by itself. Branded questions remain useful for checking accuracy, reputation and description, but their mentions should be excluded from visibility scoring because the person asking placed the company into the conversation.

Why does ChatGPT give a different answer when asked about a company by name?

Naming a company changes the task. Instead of deciding which businesses fit the customer’s need, the assistant is now being asked to retrieve, summarise or judge information about a specific entity. That makes a mention much more likely and can create a misleading impression that the company is visible for broader recommendation questions.

“Is Company A suitable for small manufacturers?” invites discussion of Company A. “Which software suits a small manufacturer that needs inventory and accounting in one system?” asks the assistant to build a shortlist. The first tests what the model can say about the named company; the second tests whether the company is selected.

Both checks matter, but they answer different questions. Branded checking can uncover factual errors or weak positioning. It cannot demonstrate discovery by customers who did not already know the name.

How do you check if ChatGPT recommends a company without being told to?

Ask realistic buying questions that do not contain the company’s name, then check whether it appears in the answer, shortlist or recommendation. Repeat the same question because one response can change on the next run. The test should reflect a real customer need rather than a generic request designed to force a long list.

“Which payroll platforms suit a UK company with employees in three countries?” creates a decision context. “List 50 payroll companies” may reward list length rather than relevance.

This page stops at defining the question set. The repeatable workflow and tooling belong in how to track AI search visibility, which should be linked when published. An AI visibility audit, also pending publication, can then test the chosen questions across assistants, repeated runs and competitors.

What happens when someone asks AI to compare two companies?

The assistant usually organises the answer around contrasts such as price, fit, features, service model, proof, limitations or customer type. It may choose one company overall, choose different winners for different situations or avoid a firm verdict. The answer reveals which distinctions the model believes matter between the two named companies.

This does not test whether either company would have entered the answer unaided; both names were supplied. It tests how the assistant frames the contest after they are present.

The useful evidence is not only the final choice. Look at which criteria are introduced, which claims are repeated, whether one company receives clearer language and which sources support the answer. Ongoing competitor AI tracking belongs on its own page and should be linked when published.

Can a business change which company ChatGPT picks in a comparison?

A business can improve the evidence that shapes the comparison, but it cannot guarantee the same verdict on every run. Clear positioning, direct comparison information, specific proof, consistent third-party descriptions, current reviews and pages that answer the buyer’s actual criteria can make the company easier to understand and justify.

First identify the criteria the assistant repeatedly uses. If answers keep contrasting implementation time, local support, contract flexibility and integrations, the company needs clear, verifiable information on those points. Generic claims such as “best-in-class service” give the model little to work with.

Then correct the evidence gaps. That may mean improving the company’s pages, earning credible third-party coverage or making distinctions more explicit across trusted sources. The goal is not to manipulate one answer; it is to strengthen the information environment that shapes recommendations.

Why does a business show up in AI answers sometimes but not others?

Two different causes are often confused. Run-to-run variation happens when the same question produces different companies on separate attempts; that is normal noise. Question-to-question variation happens when a company appears for one customer need but disappears for another; that is useful evidence about where its relevance, proof or positioning is weak.

A company may appear for “best accounting software for freelancers” but not for “accounting software for a growing retailer with inventory.” That difference may show that available evidence supports one use case and barely supports the other. It should not be averaged away as random inconsistency.

Run-to-run variation needs repeated sampling and is explained in the next section. Question-to-question variation should be mapped back to the question set: which needs, stages, locations, customer types or criteria consistently exclude the company?

Finding questions nobody has claimed is a separate task covered by tracking prompt ownership, which should be linked when that page is published.

Why do AI answers change every time you ask?

AI assistants are non-deterministic, so the same question asked twice can return different wording, evidence and companies. Inconsistency is normal. That is why one check is a sample, not a result, and why a company should not declare a win or loss from a single screenshot.

Variation can come from response generation, changes in retrieved sources, platform updates, prompt context and differences in how the assistant interprets an ambiguous need. Even when the same companies appear, their order or descriptions may change.

The practical response is to repeat important questions and look for patterns. A company that appears once in ten runs has a different position from one that appears nine times in ten, even though both can produce a screenshot showing a mention. This page defines the questions worth checking; the tracking page defines the cadence, tooling and reporting.

First decide what real customers ask and which questions genuinely test discovery. Then evaluate performance against that set without treating a named-company answer or a single run as proof.