How a machine chooses the brand it recommends

When somebody asks an assistant which brand to choose, they do not get ten links to compare. They get one name, sometimes two. That choice is not a ranking: it is a decision made from what the machine was able to find, understand and verify.

What has changed in the question

A search engine answers with a list and leaves the judgement to the person. An assistant makes the judgement itself, then announces it. A results page was a shelf; an assistant’s answer is a recommendation.

For a brand the consequence is blunt: you are named, or you do not exist. There is no second page.

What a machine can actually read of you

An assistant does not visit your site the way a customer does. It retrieves the text, the structure, and what other sources say about you. Three things decide what it takes away:

• Access. The assistants’ crawlers are not the search engines’ crawlers. A site can be perfectly indexed by Google and still be closed to ChatGPT, Perplexity or Claude, because their agents are blocked without anyone having decided it.
• The text actually there. Whatever appears only after a script has run is often invisible. A page whose content is assembled by the browser can be empty to the machine reading it.
• Structure. A title, a hierarchy, structured data saying what you are and where you operate. Without a usable category, you cannot be the answer to a question about a category.

Why it cites one brand rather than another

In our view three criteria weigh more than the rest, and none of them can be bought:

• Verifiability. A precise, checkable statement is worth more than a superlative. “Workshops here, this range, these retailers” can be quoted; “market leader” cannot.
• Corroboration. An assistant cross-checks. A brand that exists only on its own site is a weak candidate, however good that site may be. Being mentioned elsewhere counts as much as being well written at home.
• Specificity. Pages that answer a precise question are picked up more readily than pages that discuss a sector in general.

This requirement is not peculiar to machines. When a buyer asks someone whose judgement they trust, they too get a name, not a list. And whoever names you is putting their own word on the line: they will say of you only what they could defend. Give them nothing verifiable and they improvise, or they stay silent. What is written for the machine also holds up in a conversation.

What does not work

• Repeating keywords. The text is read by a language model, not counted by a machine.
• Publishing a great deal for its own sake. One weak page dilutes the strong ones.
• Entrusting your visibility to a tool that promises a score. A score is not a citation.

What we do

We open access to the agents that produce the citations, we write and structure what has to be legible to a machine, we make your claims verifiable. We then record what the assistants say about you and about your market. We do not measure: we observe. An answer is not stable, and this record does not convert into a score or a position. It contains the question asked, the date and the answer obtained, along with what is needed to repeat it. This site is built to the same rules: it is our first demonstration.