Read AI visibility with the evidence in view.

Start with the questions being measured.

Two kinds of questions test different things. A discovery question asks for a provider without naming one, so it shows whether you earn a place in the answer. A brand question names you, so it shows whether the service knows you and how it describes you.

Keep the two apart. A high mention rate on questions that already contain your name says nothing about whether new customers will find you. Write the questions in your customers’ language and for their market, and keep the wording stable once you track it: a reworded question starts a new comparison.

Read the three signals separately.

A mention means the answer names you. A recommendation means it presents you as an option to choose. A citation means it links a page as a source. They often appear together, but none of them implies the others.

Read the sentence around the name. A business can be named in a comparison, a warning or a qualified suggestion, and that context matters more than the count. That is why Opmerk shows the original answer next to each reading and lets you mark a reading that is wrong.

Check the coverage before the percentage.

A failed request is missing evidence, not an answer that left you out. A rate is only fair when it is calculated over the answers that actually came back, and when you can see how many did.

Compare two periods only when they cover similar questions, services and completion. In Opmerk the headline rates cover the successful answers of the last four weeks, so one unusual run does not decide the trend, and each run shows what it completed and what failed.

Open the cited page.

A citation shows where the answer looked. It does not prove that every statement is correct or that the source supports the claim, so read the page and the answer side by side.

For local services, the businesses in an answer can come from map listings, directories and review platforms as much as from websites. When the sources are third-party pages, the next step may be your business listing or a directory entry rather than a new page on your site.

• Your own page: check that it states the facts the answer needed, such as the service, the area and how to book.
• A directory, map listing or review site: check that your entry is current and matches your website.
• A competitor’s page: note what it covers that yours does not.

Tell the measurement every name you go by.

Answers do not always use your full name. A short name, a name without the legal suffix or a spelling without accents can all refer to you. If the reading only knows one form, those answers look like misses.

Add the names customers and answers actually use as brand aliases. Leave out a common word on its own when it would also match other businesses: a missed answer is a gap, but a wrong match is a false result.

Compare like with like over time.

Before reading a change as progress, check what else changed: the questions, the services, the models behind them, your aliases and your competitor list. Any of these can move a percentage without anything changing on your website.

Generated answers also vary on their own. That is why a question can be asked more than once in a run, and why a trend across several runs says more than a single reply. The results describe the questions you track, not every conversation on an AI service, and no ranking is guaranteed.

Turn a finding into one next step.

A report is worth reading when it ends in something to do. Start with the pattern that repeats across questions and runs, pick one step for it, and give the change a few runs before judging it.

• Missing from discovery answers while competitors are named: read which sources those answers cite, and start there.
• Named but not recommended: read how you are described; the wording often points to a missing fact.
• Not recognised in brand questions: check that your website says who you are, where you work and what you offer, in plain words.
• Described wrongly: correct the source the answer relied on, then judge the result over the next runs.