What an AI assistant says about you

Anyone asking about an organization already gets an answer written by an AI. A market of advice and guarantees has grown up around that answer. We go through seven claims that keep coming up, with what has been published on each.

Use is no longer marginal. In 2025, 32.7% of Europeans aged 16 to 74 used generative AI tools, according to Eurostat. In February 2026 OpenAI reported more than 900 million weekly users of ChatGPT. For following the news, use is lower: 10% turn to AI assistants for news, 16% among those under 35, and Spain is one of the countries where that figure doubled in a year, according to the Reuters Institute’s Digital News Report 2026.

Two caveats before we start. Several of the studies below come from companies that sell tools for this very problem, and we say so each time. Kuantia works on this too, so readers would do well to apply the same caution to us. That is why every figure comes with its source.

“The assistant only knows what it learned in training”

Half true

An assistant has two ways of knowing something. One is what it learned during training, which stops at a cutoff date: anything that happened later is not there. The other is to search the web at the moment it answers and write from what it finds, citing the pages. Google describes this as connecting the model to real-time web content so it can answer beyond that date.

But the search is not guaranteed. OpenAI explains that ChatGPT searches automatically when it judges that the question calls for it. If it does not search, it answers from what it remembers. That is why the same organization can be described with old data or with this morning’s news. Knowing which of the two routes an answer came from changes what can be done about it.

“The first thing it checks is the official website”

No evidence

We have found no independent study that measures how often an assistant cites the organization’s own website. What has been measured points elsewhere. The Pew Research Center analyzed 68,879 Google searches made by 900 U.S. adults in March 2025. The sources cited most often in AI summaries were Wikipedia, YouTube, and Reddit, which together made up 15% of all the sources linked.

Semrush, which sells search-visibility tools, analyzed 230,000 prompts over thirteen weeks in 2025 and found Reddit and LinkedIn among the five most-cited domains on ChatGPT, Google’s AI Mode, and Perplexity. We have found no equivalent study carried out in Spain or in Spanish.

What these figures suggest is that much of what an assistant says about an organization was written by a third party.

“There is a fixed list of sources that AI trusts”

False

In the same Semrush study, Reddit appeared in close to 60% of ChatGPT’s responses in early August 2025, and in around 10% by mid-September. Wikipedia went from roughly 55% to under 20%. Any figure of this kind is a snapshot of one moment.

The answer does not repeat either. SparkToro and Gumshoe, two companies in the field, asked 600 volunteers to put the same twelve brand-recommendation questions to ChatGPT, Claude, and Google’s AI, 2,961 times in all. On ChatGPT and Google, the chance of getting the same list of brands in two responses was less than one in a hundred.

A screenshot of what “the AI says” is not a measurement. To know how an assistant describes an organization, you have to ask many times, regularly, and model by model.

“What an assistant answers is reliable”

Half true

In October 2025 the European Broadcasting Union and the BBC published a study in which journalists from 22 public service media organizations in 18 countries evaluated more than 3,000 responses from ChatGPT, Copilot, Gemini, and Perplexity to questions about the news, in 14 languages. Of those, 45% had at least one significant issue. Sourcing was a problem in 31%, with missing, misleading, or incorrect attributions. And 20% contained major accuracy issues, such as invented details or outdated information.

A few months earlier, the Tow Center at Columbia University had asked eight AI search engines to identify which article a news excerpt came from. Across 1,600 queries, more than 60% of the answers were incorrect.

Read the other way, more than half of the responses in the first study had no significant issue. And both studies are about news, not organizations: we have found no study on the accuracy of what assistants say about a company or an institution. It is reasonable to assume they get that wrong too. How often has to be measured case by case.

“With robots.txt you decide what AI reads”

Half true

The robots.txt file is the note a website leaves for the programs that crawl it, telling them what they may read. It works, but each company runs several different programs. OpenAI separates the one that collects content to train its models from the one that feeds its search-based answers, and warns that a site that blocks the second will not appear in those answers. Anthropic draws a similar distinction.

There are two limits. When it is a user who asks the assistant to open a specific page, OpenAI and Perplexity warn that those rules may not apply. And the option Google offers for opting out of Gemini training does not affect whether a site appears in its search engine.

Closing the door without telling the programs apart can leave an organization out of the answers without stopping the assistant from talking about it using what others have published.

“There are techniques that raise your visibility by 40%”

Half true

The figure exists, and it comes from an academic paper by researchers at Princeton and other institutions, presented in 2024. They tested nine ways of rewriting a page across 10,000 queries. Adding sources, quotations, and statistics raised the page’s presence in the answer by up to 40%. Repeating keywords did not help.

The fine print matters. It is “up to” 40%, on an engine built by the authors themselves; on Perplexity, the maximum was 37%. What is measured is how much of the answer is attributed to a page, not whether the answer is accurate or favorable. And the authors warn that the engines change. Google, for its part, says in its documentation that there are no additional requirements or special optimizations needed to appear in its AI answers.

Something similar applies to the llms.txt file, a 2024 proposal for summarizing a website for assistants. Google states that its search engine ignores it, and we have found no statement from OpenAI or Anthropic saying they use it. Chrome, however, has started checking for it experimentally in its audit tool. This website has one: it costs nothing, and we do not expect much from it.

“What an assistant says can be manipulated”

Half true

In December 2024, The Guardian found that ChatGPT’s search tool could be fooled with hidden text on a page: with concealed instructions, the summary of a product always came out positive even when the page carried negative reviews. Researchers at Harvard showed something similar in the lab, with a fictitious catalog of coffee machines.

These are tests under controlled conditions, on versions from almost two years ago. We do not know how much of that works today. Google warns in its guide that seeking mentions that are not authentic helps less than it seems. For an organization that looks after its reputation, the risk of being found out outweighs the gain.

What does hold up

First, know what they say: ask each assistant many times, regularly, and note which sources it cites. Second, look at those sources, because many of them belong to third parties and that is where the error or the gap is. Third, publish your own information in a form that can be checked, with its sources and figures: it is what worked in the academic study, and it does not depend on any trick. Fourth, be wary of anyone who guarantees a result. On what has been published so far, no one can be sure what an assistant will answer tomorrow.

This is how we work at Kuantia. The platform audits every search result and every assistant answer, fills information gaps with original sources, and measures regularly, model by model, what the assistants answer. How the platform works.

All linked sources were accessed on October 6, 2026.

Blog