A link beside an answer is more reassuring than it should be. Sometimes it confirms the brand, sometimes it only sets a broad category, and sometimes it carries into the answer the very old detail that makes the company look like the wrong one.
In a composite check with Client A — a private medical group in São Paulo — the Perplexity answer looked almost exemplary. On a question about clinics for a specific specialty, the system named the brand, included sources, and briefly described where the branches were located. At first this seemed like good news: the brand was there, the links were there, the wording was confident.
Then I opened the sources one by one. The official page for the specialty was fresh and careful. An external directory still held the old neighborhood. Another page described a neighboring doctor’s specialty in almost the same words the system used for the client. Near the end of the answer, a listing detail flashed by: the phone number was correct, but the address no longer matched the current branch structure. Not a tragedy. But for AI visibility diagnostics, it was like a fingerprint on someone else’s envelope.
A link is where the check begins
When an answer has sources, the reader is tempted to relax. The system seems to be saying: this is where I took the information from, you can check it. In search, that gesture is familiar. In AI answers, it is more complicated, because a link does not always prove the exact sentence it sits beside. It may confirm part of the answer, provide general context, or simply be close to the topic.
A source in an AI answer is a trace of possible model support, because the link may confirm a phrase, assign a category, or carry an error into the answer.
I keep this definition nearby when I analyze Perplexity sources for brands. This is not skepticism for its own sake. In my assessment, systems with citations are useful precisely because they give material for manual checking. The links appear on the screen, and then the ordinary editorial work begins: open, compare, record the oddity.
With Client A, which I describe as a composite scenario from several observations, the model found real brand traces. The error was subtler: one source confirmed that the clinic existed, another suggested the old geography, a third pushed toward a neighboring specialty. On the screen, it looked like a coherent answer. On the desk, once the printouts were laid out, it looked like papers from different folders.
That is why I read the answer slowly. I open the link, look for the specific phrase, compare it with the wording in the answer, and note what exactly the source may have given the model. Sometimes the link does not contain the disputed detail at all, but stands beside the paragraph so confidently that it seems responsible for everything said there. This is a small stage trick of the interface: the source becomes set dressing for credibility.
Three roles of a source
I usually divide sources in AI answers into three roles: confirmation, category label, and cause of confusion. This is a working yardstick for reading an answer, not a strict academic classification. It helps me put aside the broad argument with the model and look at what work each link is doing.
A confirmation source answers a simple question: does this company, service, specialist, or page exist? In Client A’s case, the official specialty page worked exactly like that. It showed that the group had the required profile and that the brand really did belong to the query. Here the link was like a normal file from the archive: name, subject, current folder.
A label source gives the model a convenient category name. It may be an external directory, an industry page, or an overview article. Sometimes it is useful because it writes more simply than the company site. Sometimes it is dangerous: it contains a short formula that pulls the brand onto a neighboring shelf. For the clinic, that label was a page where the specialty was named more broadly than on the site. For a patient, the difference between neighboring medical terms may be unclear. For the brand in an AI answer, it changes the outline.
A confusion source is the most unpleasant one. It seems close to the topic, but brings in an extra detail: an outdated neighborhood, a discontinued service, an incorrect doctor description, an old branch scheme. In one answer variant, the model took the clinic’s correct name but placed beside it a detail from a page about another specialty. The answer did not fall apart on a quick read, but it gave the patient a blurred expectation.
These roles can mix inside a single link. A directory confirms the brand and brings in the old neighborhood at the same time. The official page proves the specialty but does not answer the patient’s conversational query. An overview article helps define the category, but places a neighboring practice beside it. That is why I do not write simply “good source” or “bad source” in the table. I note exactly what work it does in the answer.
How I read Client A’s answer
In medical topics, I am more cautious than usual. A public article is not a place for diagnoses, internal data, or identifying details, so the case here is assembled from repeated observations in similar situations. The typical picture looks like this: a private medical group in São Paulo has several branches, separate pages for specialties, doctors, and neighborhoods. The site is active, but external traces age more slowly than the real structure changes.
I start with a description of the answer, without rating it “good” or “bad.” First I note which phrases match the official page exactly. Then I look at where each added detail came from. If the model mentions a neighborhood, I check whether that neighborhood appears on the current page or came from a directory. If it describes a doctor, I check whether it used a neighboring specialty where similar words appear side by side. Sometimes the most important mistake hides in a single adjective.
In this composite episode, one source was almost flawless, but not central to the meaning of the answer. It confirmed that the brand existed. The category seems to have been reinforced by an external listing. And the confusion came from a page where the old neighborhood was written more prominently than the current branch layout on the official site. For local clinics with several branches, this is a familiar situation: the company carefully updates its own house, but an old sign is still hanging on the neighbor’s fence.
One imperfect detail was especially revealing. The model distinguished several branches, but in one answer it attached the former territorial clue to a new specialty. Someone from the clinic would immediately understand that this was an imprint of the old structure. An outside reader might decide that the service was available at a location where it had not been presented as primary for a long time. For a medical brand, this kind of half-truth is often more dangerous than an obvious error: it is harder to notice, and therefore harder to correct in time.
Why an old page sounds confident
Old directories are often written crudely, but in a way that is convenient for machines. They contain the company name, city, neighborhood, category, phone number, sometimes a short phrase about the service. The official page may be richer and more delicate: the history of the specialty, the doctor’s approach, different reasons for making contact, limitations. For a person, that text is better. For the model in a comparative query, the short profile can be easier to grab.
In moments like this, I ask the client to put aside the argument with the screen and open the source page. The model’s answer is only the symptom. The cause often sits on a page nobody has read in full for a long time, because it seems too old to matter to a living business, while still being visible enough for the machine.
This does not mean the clinic’s site should be turned into a directory. The official text should remain human and careful. But somewhere inside it, there needs to be a clear anchor sentence: who provides the service, in which specialty, in which city or neighborhood, for which type of query. Without that sentence, the model may take the category from an external source because that source speaks more simply.
There is one more small reason that is easy to miss. An external page may repeat itself in different places in almost the same words. The same short formula lives in a directory, a local list, an old specialist profile. The official site changes more accurately, but each section speaks in its own way. As a result, the error is not deep, but wide. It lies in a thin layer across several pages.
What to change after analyzing sources
After checking sources, I rarely advise deleting everything old right away. Some external pages cannot be corrected quickly. Besides, an old mention sometimes still helps the brand appear in answers. A blunt cleanup can remove both the error and the useful trace. The work has to be more careful: strengthen the official source, align key formulations, ask for the most visible external profiles to be corrected, and then check the query series again.
Inside the site, I look for one short formula for each specialty. It should be a calm formula, not an advertising slogan: brand, medical specialty, city, and language the patient understands. There should be real text around it; otherwise, the page turns into a label without a body. But without the label, the box goes to the wrong place too.
In external sources, perfect cleanliness is secondary. A local network will never be sterile. First I take the pages that already appear in model answers or sit close to them: visible directories, partner profiles, old doctor cards. If a correction can be made there, it is better to fix one strong crooked sign than rewrite a stack of quiet pages the system does not yet see.
Here it helps to return to the first piece about how a model assembles a brand from public traces: “Why ChatGPT Mentions a Brand Without Visiting Its Website.” A source in Perplexity shows one layer of that assembly. It does not reveal the whole internal mechanism, but it helps you see which scrap of paper is lying on top.
After corrections, I return to the same queries. This is an important discipline of the field notebook. If the model previously confused the neighborhood on a conversational phrasing, then the same phrasing should be asked again, with the same level of messiness. A neatly rewritten official text has to withstand not only a marketer’s careful question, but also the living phrase of a patient searching for a service late at night from a phone.
The same principle helps when the problem lies in the service names themselves. I analyze a case like that in the article “When One Service Lives Under Three Names.”
It is not always clear how much of the answer came from a specific link. The system shows sources, but does not always reveal which phrase from which material became decisive. That is why I look at repeated answers, sources, and oddities across a series of checks.