AI Visibility Checker

How this AI visibility checker works

Type your brand, type the questions your buyers actually ask, and the tool sends each question to ChatGPT, Claude and Gemini several times over. It then counts how many of those answers mention you. That count is the whole product. There is no score to interpret and nothing is weighted behind your back.

Before any of that runs, it checks whether the AI crawlers can reach your site at all. Both layers matter and they fail for different reasons, so the tool reports them separately instead of folding them into one number.

Why a rank number is the wrong answer

Most AI visibility tools hand you a position. Rank 3 for ChatGPT, rank 7 for Perplexity. Rand Fishkin and Patrick O’Donnell measured that in January 2026: 2,961 runs of twelve brand-recommendation prompts across ChatGPT, Claude and Google’s AI Overview, each prompt repeated 60 to 100 times per platform. The same list of brands came back twice fewer than once in a hundred tries. The same list in the same order, fewer than once in a thousand. A single position is one sample dressed up as a measurement.

Run the same prompt ten times and the honest answer looks different. You were named in six of ten runs. That is 60 percent, and with ten runs the true rate could plausibly sit anywhere between 31 and 83 percent. Both halves of that sentence are useful. The rank was never useful. That is where the same study landed: a share measured over repeated runs survives as a comparison, a position does not.

So the headline here is a frequency and a Wilson 95 percent confidence interval, shown as a pale band behind the solid arc. The band shrinks when you run more samples, which is exactly the lesson most tools hide.

Training memory and live retrieval are two different problems

A model can name you for two completely unrelated reasons. Either you are in its weights from training, or it searched the web while answering and found you. Those have different failure modes and different fixes, so the tool asks each question twice: once with web search off, once with it forced on.

The retrieval half costs roughly fifteen times what the training half costs, because the search itself is what you pay for, not the tokens. Both halves are free to use, they just draw on separate allowances: three training runs a day and one retrieval run. Your own DataForSEO key removes both. That is the point where real money starts, said plainly rather than buried in a pricing page.

Perplexity sits in only one column, on purpose

Perplexity has no offline mode. It is a search product and it searches every single time, so there is no training-memory reading to take from it. Rather than show you an empty cell, the tool leaves Perplexity out of the training check and brings it into the retrieval pass, where its answers carry more citations than any other engine.

Where the AI actually looks

The runs above ask the models a question. That tells you what they say. It doesn't tell you which sources they lean on while answering, and those are two different problems with two different fixes.

So the tool also reads the archive. For a category keyword it pulls the domains that actually get cited in real Google AI Overview answers in your market, ranked by how often, with the AI search volume behind each. Run it on "seo tool" in Germany and YouTube comes first with over three thousand citations, ahead of every SEO vendor. Nobody would have guessed that, and you only find it by looking.

Your own domain is marked in the list when it's there. When it isn't, the list is still the answer: those are the pages answering your category in your place. The check costs a credit of its own, because the call behind it runs about sixty times the price of a model call. And it reads archived answers rather than running a live test. ChatGPT's mentions database takes United States and English only, which we measured rather than read, so everywhere else this is Google AI Overview.

What the crawler gate actually gates

Blocking an AI crawler in robots.txt does something, but not the thing most articles claim. The bots split into two jobs and only one of them changes today's answer.

GPTBot, ClaudeBot and Google-Extended collect training data. Block them and you slowly fall out of future model releases, which is a problem measured in quarters. OAI-SearchBot, Claude-SearchBot, PerplexityBot and Google-Agent fetch pages while the model is answering a live question. Block those and you disappear from the retrieval column immediately.

So the gate reports both, tells you which side each bot is on, and checks the real server response as well as the robots rule. A bot that is allowed in robots.txt and gets a 403 from your firewall is still blocked, and that mismatch is common enough to be worth a separate column.

More tools

FAQ