What an AI test adds.
A good search position does not show whether an AI agent understands your offer and can perform a customer task. Check the answer, its sources and the steps actually performed separately.
Our DAX 40 Index tests defined tasks through different agent access modes [S1]. The prompts below adapt those questions into a simple self-test. They replace neither that measurement nor a technical SEO audit.
Prepare the test.
Choose a product page and a concrete customer task. Record the date, assistant, model, available tools and complete answer. For prompts 3, 4 and 6, you can supply this context first.
Test the brand [brand]. Website: [URL]. Category: [category]. Market: [market].
The customer task: [buy / request a quote / book / apply].
Main competitors: [A], [B], [C].
After each test, report: what could you demonstrably find or do? What could you not? Which other source or website did you use? Separate observation from inference. Start prompts 1, 5 and 7 in separate fresh chats without this brand context. Otherwise you supply the answer you are looking for. Prompt 2 also starts fresh and names only the brand. Temporary chats can reduce the influence of previous conversations; record the settings used.
Prompt 1 · Is your brand found?
Begin with a customer question that contains no brand name. Turn web search on and provide only the category and market.
Search the web: which [category] providers in [market] would you currently recommend to a customer, and why? List every source used. Is your brand named? Is one of your pages linked as a source? Check the link and the claim it supports. A mention without your own source does not establish where the model got its knowledge. Absence is a finding for this question, not a verdict on your entire AI visibility.
In our published citation log, one system failed to associate an existing dataset correctly [S4]. Missing mentions can have several causes. Start by checking the sources the answer actually uses.
Prompt 2 · Is the brand picture current?
Model knowledge can become stale. Comparing it with a search shows which claims current sources add or correct [S7].
What do you know about [brand]? Answer without web search. Identify claims you are unsure about. Do not claim a freshness date you cannot substantiate. Open a second fresh chat and ask the same question with web search. Compare products, locations and services with your current facts. The model's self-report proves neither its training cutoff nor the origin of individual claims.
An outdated brand picture becomes visible when you compare it with current facts.
Prompt 3 · What can the AI read?
A text fetch and a browser can receive different content from the same page. Test the access actually available. An assistant without a retrieval tool cannot perform this check [S5].
Retrieve [product page URL] with your available tool. Name the tool used. Which products, prices, specifications, availability and next steps can you substantiate directly from the page? List missing or unreadable content. If you have no web access, say so explicitly. In the published BMW test, a coding agent found 227 interactive elements but could not read their content [S2]. This is a historical finding for that configuration. It shows why a page visible in a browser should also be tested through the intended agent access mode.
Visible in a browser does not mean readable through every AI access mode.
Prompt 4 · How far does the agent get?
Now test a concrete task. Use an assistant with browser control. A description of clicks without executed steps is not a usability test.
On [URL], check how a customer can [buy a product / request a quote / prepare a booking]. Perform only read-only navigation. Stop before submitting personal data, an inquiry, order, booking, payment or signature. Log the steps actually performed, facts read and exact stopping point. Identify required access and approvals. Do not invent actions performed. Where does the journey end: a cookie dialog, configurator, form or login? Is this a technical barrier, missing permission or the deliberate test boundary? Record those separately.
The automotive research shows how much results vary by agent access mode [S2]. Your self-test demonstrates only the part tested. Full transaction evidence requires a suitable test environment and agreed permissions.
Prompt 5 · Where does the recommendation lead?
AI can name your brand while directing customers to a dealer or comparison site. Check whether the recommended route fits your sales model.
I want to [buy / insure / book] [product category] in [market]. Where should I do that? Name a first and second destination with a link, reasons and sources for each. Do not perform a transaction. Compare manufacturers, dealers and portals. An intermediary is not automatically a failure: in some markets it belongs to the intended route. The reinsurance research shows that familiar brands can be named while the broker needed for the next step is missing [S3].
What matters is whether the recommendation leads to the right next step.
Prompt 6 · Which evidence supports the answer?
Check how AI supports its recommendation. A source must substantiate the specific claim, not merely mention the same brand.
Why should a customer choose [brand] over [competitor]? Support every verifiable claim with a directly linked source. Flag uncertainty and conflicting information. State explicitly if you cannot retrieve a source. Open the links. Are they accessible, current and relevant to the claim? Your own evidence and independent sources can complement each other. Without a source, the claim remains unsupported in this test; its origin cannot be established reliably.
Our citation log shows different sources for differently phrased questions [S4]. Record the exact question and cited pages together.
Prompt 7 · Are your terms associated correctly?
If you use a distinctive product or technical term, check which provider is associated with it. Do not name your brand.
Explain [product or technical term]. Which providers or originators do you associate with it? Which sources support that association? Separate documented origins from guesses. Your page is cited, another provider is named or the association remains unclear. All three outcomes are clues. One attempt proves neither ownership of a term nor an uncontested market, and does not justify renaming it.
In our log, questions containing distinctive terms produced different results from plain-language versions [S4]. Repeat a question without your technical term as well. That tests your customers' language.
Put the seven answers in context.
Group observations into four questions: found, read, used and evidenced. The diagram is a diagnostic aid. The seven prompts do not calculate an Agent Success Score.
In the published Agent Success Score, AI usability carries a weight of 70 percent [S6]. It covers more than a good answer. The methodology explains the other weights and measurement scope.
| Check | Question | Possible warning signal |
|---|---|---|
| 1 · Find | Is your brand named? | The answer names only other providers. |
| 2 · Brand picture | Are the facts correct? | The answer describes an old offer. |
| 3 · Read | Are important details readable? | The fetch returns only scripts. |
| 4 · Act | Where does the task stop? | The agent cannot proceed past the configurator. |
| 5 · Next step | Does the route fit? | The required sales partner is missing. |
| 6 · Evidence | Do sources support the claim? | The cited source does not contain the claim. |
| 7 · Terms | Is the association correct? | The term is attributed to another provider. |
What the self-test leaves open.
One attempt captures one assistant, question and point in time. Answers can change. A reliable audit fixes tasks, repeats attempts and records conditions. Our public log also records missing citations [S4].
The free website check is the starting point if you want to discuss a finding with us. We show where AI finds your offer and where customer tasks get stuck. An audit and agreed implementation follow as a separate project. We agree scope, access and timing before work starts.
Start with prompt 1. Save the question, answer and sources. This gives marketing and engineering a concrete basis for discussion.