Opened eight times, linked twice, never named

Most people who start measuring AI visibility watch one number: the share of answers in which the model names their brand. It is the figure that ends up in every screenshot and every meeting.

The trouble starts when that number reads zero, which is the normal state early on and especially if you are small. A zero confirms you have a problem and leaves you exactly as lost as you were before you bought the tool.

This piece is about the other three levels, which do tell you what to do. The data is ours, 30 days of it, with the limits set out at the end.

Opened, cited and named are three different things

Most people work with two states, you appear or you do not. There are really three levels, and separating them matters because the fix is different in each one.

• Opened. The search step puts your page in front of the model while it writes. The user never sees this.
• Cited. Your page is left as a link in the text the user reads and can click.
• Named. The model states your brand, with or without a link.

Over that month our visibility was 0%. In the same period the engines opened five of our URLs, eight times in total, and of all that only two were left as links in the text.

That is not a contradiction, it is three different questions. You can feed an answer that ends up recommending somebody else. That is what happened to us.

Our own URLs detected in real engine answers over 30 days, to 18 August 2026. "Linked in the text" comes from the attribution each engine returns between a fragment of its answer and the source behind it.

Two details we took to heart. First, none of the five is the pricing page: the model opens us to explain the subject and then recommends a competitor. Second, and more useful, the only two that made it into the visible text are the same study built on original data, in Spanish and in English, both linked by Gemini. Not a "best tools" listicle, but research with numbers that existed nowhere else.

Where the AI gets the answer from

The second level is the external domains the engine opens, and it is the one that changes strategy most, because it shows where your category is being decided without you.

Domains the engines opened most while answering our monitored questions, 30 days to 18 August 2026. These are retrievals, not necessarily visible links.

Reddit alone is opened more than arXiv, TechRadar and Ahrefs combined, which add up to 3,767. If your plan was to publish more articles on your own blog, that number should give you pause: a good part of the conversation feeding these answers happens somewhere you do not control and cannot edit.

This is not an invitation to go and post about your brand on Reddit, which usually ends badly and breaks the rules of most communities anyway. It is an argument for knowing what is said there, and for accepting that reputation work now weighs as much as content work. We go deeper in how LLMs choose their sources.

The searches the engine invents before answering you

The third record is the least known and the one that helped us most. Before writing, the engine generates its own queries to research the topic. A single user question can turn into several distinct searches.

In 30 days we logged 2,158 distinct sub-searches across our monitored questions. That is where you see how the model rewrites a question, and therefore what you would actually need to cover, which rarely matches the phrase the user typed.

A pattern turned up that we did not expect. Queries using the site: operator came from ChatGPT, which went straight to each vendor's official site and very often to its pricing page:

• site:peec.ai pricing
• site:otterly.ai pricing
• site:semrush.com AI Visibility Toolkit
• site:ahrefs.com Brand Radar AI visibility

The natural language ones, along the lines of "best AI visibility tools for agencies", came from Gemini.

The practical consequence is uncomfortable and quite specific. When somebody asks which tool to buy, the model does not read your blog post, it goes to your price. If your pricing is not on a public page, you are not in that answer. We do have a pricing page and the model still never opened it, which points at a different problem, one of authority and links rather than existence.

What to do with the three records

• Mentions. Good for knowing whether there is a problem and for watching it move. No use at all for deciding what to write.
• External domains opened. They tell you where to invest off your own site: which publications, which forums, which third-party comparisons.
• Your own URLs, opened versus linked. If it opens you and does not link you, it read you and preferred somebody else, so the problem is the page. If it never opens you, the problem is that it cannot find you. Two different fixes.
• Sub-searches. They give you the real syllabus, in the engine's words rather than yours.

Without the URLs our reading would have been "this is not working, keep publishing". With them the diagnosis was different, and a good deal more uncomfortable.

The limits, which are worth stating

These numbers come from one project, ours, in a narrow category, AI visibility tools, over 30 days, across a set of questions we chose ourselves. That shapes the results in three ways worth keeping in mind.

First, arXiv's weight is almost certainly a feature of our sector, which is full of recent research; in a consumer category it would not show up. Second, our questions include buying queries, which is probably what pushes the model towards pricing pages. Third, eight retrievals and two links are small numbers, so the split by page type is a signal rather than a demonstration.

What we think does hold beyond our case is the underlying distinction: opened, cited and named are three different things, and watching only the last leaves out nearly all the actionable information.

Questions people ask

Can you have 0% visibility and still be opened?

Yes. Opened is a page the search step puts in front of the model, cited is one left as a link in the text, and named is when it states your brand. You can be the material for an answer that recommends somebody else.

What are fan-out queries?

The searches the engine issues by itself to research before answering. We logged 2,158 distinct ones in 30 days.

Do ChatGPT and Gemini search the same way?

Not in our data. site: queries appeared in ChatGPT and natural language ones in Gemini. One dataset, not a general law.

Where do I start if my visibility is zero?

With the sub-searches, to learn what the engine is really asking, and with the domains it opens, to learn where your category is being decided. Publishing more articles without looking at either is working blind.

Read next: how LLMs choose their sources · AI cites deep pages, visitors land on the homepage · query fan-out in ChatGPT and Gemini.