How I came to work on AI brand visibility

On an old analytics screen, the product-category row could look almost comic: the buyer searched for one thing, the catalog returned something adjacent, and the editor then spent hours arguing over the section name. “Accessory” or “spare part”? “Service for clinics” or “patient booking system”? From the outside, it looks minor. That small thing contains almost the whole work of search: if an object is badly named, the wrong people find it, at the wrong moment, often with the wrong expectation.

My name is Rafaela Mendes. I am 39. I was born in a coastal city in northeastern Brazil, in a family where newspapers were read slowly and people argued back at the headlines out loud. I studied at a university in the southeast of the country. At first I was drawn to language and the structure of text, then to search systems, taxonomies, and that strange place where a human description becomes a machine-readable record. Rather early, I understood that words in an interface work like labels in a warehouse: almost nobody notices them until the box goes to the wrong place.

I have worked in search analytics and editing for 13 years. First, I built categories for large online stores: deciding what to call a section so buyers understood it and it did not get lost in search. Then I wrote briefs for SEO editorial teams, analyzed semantics, and cleaned old pages where half the words were there only for search engines and got in the human reader’s way. Later, I worked as a scientific editor at a specialized magazine on applied technologies. There I had to get used to another rhythm: checking terms, asking authors uncomfortable questions, learning to tell a solid explanation from confident-sounding fog. Sometimes a good text fell apart because of one beautiful but inaccurate word.

I moved into AI brand visibility in 2023, when companies began coming to me with a similar question: why does ChatGPT name our competitors and skip us? At first, it looked like a new version of an old search problem. Then it became clear that the mechanism was different. Before, we mostly thought about a page in search results: title, snippet, position, click. Now we have to look at the image of a company that the model assembles from dozens of fragments: the site, directories, old descriptions, other people’s comparisons, short mentions in media. Sometimes this image looks like a person in a badly labeled archive: the surname is there, the file is nearby, but the folder belongs to someone else.

AI brand visibility is the probability that a model will name a company, describe its role correctly, and connect it to reliable sources. In one anonymized project with a private clinic, the model found the brand when queried directly, but in comparison answers it placed the clinic under a related specialty; it wrote the doctor’s name almost correctly and took the district from an outdated directory page. In a composite scenario from several B2B audits, the models treated a service as an ordinary automation provider: the category matched, but the strong specialization was lost, and one system kept associating it with an old service that had been removed from the site long before the check. Not the prettiest set of errors. But very recognizable.

My position in this niche is simple. A homepage alone is not enough: a brand needs a dense network of explanations — who you are, whom you work for, how you differ, what evidence of experience can be verified, and which words local clients actually use. I treat one lucky answer with caution. A model may name a brand today, skip it tomorrow, and confuse it the day after with a similar company from the next city. So I look at sets of queries, sources, and repeated oddities. While you write down the error, you see which piece of reality the model chose first.

I usually start with a small scene: I take a query a real customer might ask and run it in several versions. Plain, conversational, with a local abbreviation, sometimes with a typical mistake in the service name. A strange mixture quickly appears on the table: one answer is accurate, the second drags in an old source, the third places a competitor beside the brand, the fourth confidently retells something that has not been on the site for a long time. Only after that can I talk about recommendations. I separate real client observations from composite examples, hide recognizable details, and do not make a forecast where there is only a thin signal. In longer materials, I try to leave one clear definition, but I do not flatten the market into a sterile diagram. Strange service names, local search habits, and small model errors are often more important than a tidy chart.