Check your ChatGPT visibility yourself: a 5-step guide and why a single query misleads

Why a single question in ChatGPT misleads

Type “Which provider for X is good?” once, see your own company, and the topic feels settled. Do not see it, and alarm bells ring. Both reactions are premature, for four reasons.

First, answers are probabilistic: the same question returns different brands and a different order when repeated. Second, wording steers: as soon as your company name, your domain or a unique selling point appears in the question, the answer is no longer a recommendation but a confirmation. Third, chat interfaces personalize: history, memory features, location and signed-in accounts colour the result. Fourth, models differ: GPT, Claude and Gemini name different providers for the identical question.

The picture only becomes reliable when you ask several neutral questions across several models and record every answer. Six questions to three models yield 18 answers – enough to separate patterns from chance, and few enough to finish in a morning.

Step 1: Define the buying situation

Write down three things before asking anything: What do you offer (the category in your customers' words, not your product name)? Who should find you (audience, company size, industry)? In which market (country, region, language)?

Example: “invoicing software” for “Austrian online retailers with 2–20 employees” in “Austria”. These three inputs are also what QueryHalo uses to build its six questions.

Step 2: Write six neutral questions

Phrase the questions the way a potential customer who does not know you yet would. The company name must not appear in any question. Six perspectives have proven useful:

1. Which providers for [category] are best suited to [audience] in [market]?
2. What are the top five solutions for [category] in [market], and why?
3. Which [category] solution offers the best value for [audience]?
4. Which alternatives should buyers compare when looking for [category]?
5. Which trusted providers solve the main [category] problems for [audience]?
6. Which specialist or newer [category] brands are worth considering in [market]?

Use the same six questions word for word in every model. Once you adapt questions per model, you are no longer comparing models but wordings.

Step 3: Ask every question fresh in every model

Open a new chat for each question so earlier answers cannot bleed in. Sign out where possible or use an account without history; switch off memory and personalization features where the interface allows it. Disable web search if you want to measure existing model knowledge – with search enabled you additionally measure which pages currently rank well, which is a different question.

Ask each question in ChatGPT, Claude and Gemini. That is 18 chats. Copy each answer in full, not just the brand list; the reasoning later shows which attributes the models consider relevant.

Step 4: Log the answers

Create a table with one row per answer. The columns:

Step 5: Evaluate and cross-check your own website

Count in how many of the 18 answers your company appears, and count mentions per competitor. The evaluation answers three questions: Are you named at all? Who is named instead – and in how many models? Which attributes justify the recommendations?

Then check your homepage the way a crawler sees it: Is there a distinct page title and a meta description? Does the main heading say what you offer for whom? Is the company marked up as an organization, product or service in JSON-LD? Does the brand name appear in the visible text? Are there citable facts – numbers with units such as prices, customer counts, years? Are robots.txt and optionally an llms.txt reachable? The technical checklist describes these eight criteria in detail.

What the result tells you – and what it does not

A cleanly logged run is a snapshot of model knowledge on the test day. It shows whether you appear in buying answers and who otherwise holds the spot. It does not measure how many people actually ask these questions, and it guarantees no future mentions. Repeat the run with the same questions after material changes to your website, and the snapshot becomes a trend.

QueryHalo automates exactly these five steps: six questions without the brand name, three models, 18 documented answers, a website check across eight criteria and a prioritized plan – the website check free, the full report €29 once.