How to get your company cited by ChatGPT, Gemini and Perplexity

What you get from this article

• AI extracts claims, it does not copy pages: what has to be impeccable is the sentence, not the keyword.
• Any number with no source and no date is a number the model would rather not cite.
• Explicit authorship is not vanity: it is what lets the model know who to credit.
• What people say about you off your site weighs as much as what you say on it.
• There is no official dashboard for AI mentions: whoever measures, measures by hand, with a fixed list of questions.
• Blocking AI crawlers in robots.txt is the fastest way to disappear from those answers.

The question moved, and the answer changed shape

When somebody wanted an accountant in their city, they typed "accountant city" and picked one of ten links. Today a growing share of those people writes something quite different: "I need to change accountants, my company is small and the current one takes too long to reply, what should I look at before choosing?". And they get four paragraphs of text that sometimes names names.

The difference is not only format. In the list of links, ten companies appeared and the person chose. In generated text, three appear and the other seven do not exist. Search stopped distributing attention and started concentrating it. That is bad if you are outside and disproportionately good if you are inside.

The good news is that the selection logic is neither mysterious nor for sale. It is technical, and in large part it is the same thing a human editor would do: pick the source that states something clear, that shows where it got it, and that lets you know who it is. The bad news is that almost no small or mid-sized company website does those three things.

That work has a name, GEO, and it coexists with SEO rather than replacing it. If you have not separated the two in your head yet, it is worth reading the difference between SEO, AEO and GEO first, because what follows assumes the SEO foundation already exists.

The two routes by which a model reaches you

There are two different paths, and they call for different work. Confusing them is why most GEO attempts go nowhere.

The first route is live search. When you ask about something current, the assistant does not answer from memory: it fires off its own searches, often several at once, reads the pages that come back and writes new text based on them. This route is fast to influence, because it depends on what is published now and on the crawler being able to read it. This is where a new, well-built page can break in within weeks.

The second route is training memory. The model was trained on an enormous amount of public text, and whatever appeared many times there became part of what it knows without needing to search. This route is slow, you do not control it and there is no shortcut: it is the consequence of years of presence, mentions and repetition by other people.

In practice you work both at once, but with different expectations. The first route is where this month effort shows up. The second is where this year effort shows up two years from now. Anyone promising fast entry into the model memory is selling something they do not control.

Live search is what you change this month. The model memory is what you change in two years. Selling the second as if it were the first is the most common con going right now.

Write claims, not pretty paragraphs

The unit the model carries into the answer is not your page: it is a claim. A sentence that says something specific and can be verified. If your text has thirty sentences and none of them states anything concrete, there is nothing to extract, however pleasant it is to read.

Compare. "Our team is highly qualified and always seeks the best results for the client" is a sentence that states nothing: it is not false, it is empty, and no model will carry it anywhere. Whereas "the average response time to a support ticket at ROO3 is four hours on a business day" states something, is verifiable and is citable.

The same logic applies to definitions. If your field has a term that confuses people, write the definition in two sentences, on its own, in its own paragraph, right after the heading that asks the question. Models and search engines cut that format easily because it is already shaped to be cut.

And every time a number appears, put the source and the date next to it in the text itself, not only in a footer. "According to a Seer Interactive study published in 2026" costs five words and turns a suspect claim into a citable one. A number with no provenance is the first thing a system trained to avoid error discards.

Say who wrote it, and prove that person exists

Here is the point almost every company website gets wrong. The text has no author. No date. No revision. And the company that published it has no page confirming that it genuinely exists, with an address, a phone number and a history.

For a model, that is an attribution problem. It read a claim, found it useful, and has nobody to credit. Citing "a website" is not citing. So it prefers the source that has a name, a role and a date, even when that text is worse than yours.

The fix is cheap and almost nobody does it: a byline with name and role, publication and last-review dates, an about page that is specific rather than generic, and structured data markup saying all of that in code for whoever does not read the layout. Every article on this site carries that markup, and it is not decoration.

One important caveat: signing as a person who did not write it, or inventing credentials, is the shortest path to destroying the trust you are trying to build. A real author with a modest record is worth more than a fictional expert with an impeccable biography.

What people say about you off your site

A model that only finds you on your own site has a single, unconfirmed source. A model that finds the same information in three independent places has confirmation. That shifts the odds of citation disproportionately.

The places that weigh most are the most tedious to get and the most durable: profiles in recognized directories for your sector, a complete and verified Google business profile, taking part in third-party content such as an interview or a guest article, and mentions in outlets that cover your market.

Reinforce consistency. Company name, address, phone and description have to be written exactly the same way everywhere. Two different spellings of the same name become two different entities to an automatic system, and each one gets half the weight.

There is a practical example here too: if you have products or published work, a projects page with a real description of each one gives the model concrete material to associate with your brand. A generic portfolio does not generate citations, a specific one does.

The technical basics nobody can skip

Before any content strategy, three checks. They take an hour and they knock out most GEO attempts that go nowhere.

First: is your robots.txt letting AI crawlers in? A lot of people copied a restrictive configuration from some tutorial and blocked them without knowing. If you block an assistant crawler, you do not appear in its answers, full stop. And there is a specific trap: in robots.txt, a named group does not inherit the rules of the generic group, so a block written for one crawler can release what you thought was blocked, and the reverse.

Second: does the page work without JavaScript? Several AI crawlers do not execute script. If your content only appears after the browser runs a JS bundle, the crawler receives an empty page. Test it by turning JavaScript off in the browser and looking at what is left.

Third: is there any real public content? A site with five institutional pages of shop-window copy has nothing to be cited. It is not about volume, it is about there being any claim at all on the subject you want to be remembered for.

• Robots.txt allowing the AI crawlers you want to reach, with attention to named groups.
• Content present in the served HTML, without depending on JavaScript to exist.
• Its own title, description and canonical on every page, with no duplication.
• Structured data for article, author, organization and frequently asked questions.
• Publication and revision dates visible to the reader and to the machine.

About llms.txt, and what is myth in this story

The idea going around is that simply publishing an llms.txt file at the root of your site makes AI systems start citing you. It is worth separating fact from hope.

llms.txt is a proposed text file that describes the site in an organized way for a model, like an annotated index. It is easy to make and costs nothing. What none of the big search engines or labs has publicly confirmed so far is using it as a selection factor. In other words: publishing it is reasonable, promising results because of it is irresponsible.

The honest recommendation is this: publish it, because the cost is an hour and the risk is zero, and treat it as internal organization and a low-cost bet, not as strategy. If it becomes a standard one day, you are ready. If it does not, you lost an hour.

The same scepticism applies to any tool promising to "put your brand in ChatGPT" for a monthly fee. You cannot buy a position in an AI answer. There is content that deserves to be cited and infrastructure that makes citing possible.

How to measure, and what cannot be measured

There is no Search Console for AI. Anyone claiming to measure assistant mentions precisely is estimating from a sample and calling it measurement. That does not mean you cannot track it: it means the method is manual and honest rather than automatic and false.

The method that works: build a fixed list of twenty to thirty questions a real customer would ask, written the way they would say them. Run that list once a month through the assistants that matter to your audience. Note in how many you were cited, in how many the competitor was, and which of your pages was used. Keep it in a simple file. In three months you have a real timeline.

Add two indirect signals to that. On the server, watch visits from AI crawlers: if they are not coming, nothing else matters. And on your contact form, ask how the person found you, with an explicit option for an AI assistant. That is the most reliable attribution data in this channel today.

If you want that survey done in a structured way, with the question list built around your sector and a comparison against whoever is already being cited in your place, that is what ROO3 AI consulting does in the first stage. And if the SEO base is not standing yet, that comes first: see how the SEO work runs.

Frequently asked questions

Can you pay to appear in ChatGPT answers?

There is no way to buy a position inside the generated answer the way there is advertising in search. What exists is content that deserves to be cited and technical infrastructure that lets the model find and credit it. Tools promising to put your brand in the answer for a monthly fee are selling something they do not control.

How long before an AI starts citing me?

Through the live-search route, a new, well-built page can be found within weeks, because the assistant searches at the moment of the question. Through the training-memory route it is years and you do not control the timing. Most short-term results come from the first route.

Do I need to allow AI crawlers on my site?

If your goal is to be cited, yes. Blocking an assistant crawler in robots.txt is the most direct way to guarantee you do not appear in its answers. The decision is legitimate either way, but it has to be deliberate: many people blocked them without knowing, by copying a tutorial configuration.

Does an llms.txt file make AI cite me?

No major search engine or lab has publicly confirmed that llms.txt is used as a selection factor. It is cheap to publish and does no harm, so it works as a low-cost bet and internal organization. It is not a strategy, and anyone promising results because of it is promising what they cannot deliver.

Does AI-written content hurt your chances of being cited?

The problem is not the tool, it is the result. Generic text, with no verifiable claim, no source and no author, has little chance of citation whether a person or a machine wrote it. Text with data, provenance and a byline has a chance, also in both cases.

My site is small. Is it still worth it?

It is, and sometimes it is worth more. On a specific or regional subject, the competition for citation is far smaller than on a generic term. A small company that publishes five genuinely good pieces about the problem it solves competes better in this game than it would for position in a contested search.

Sources

• Google Search Central: AI features and your website
• Google Search Central: succeeding in AI experiences in Search
• Seer Interactive: AIO Impact on Google CTR, 2026 update

Rodrigo Fávaro

Founder of ROO3, a marketing and technology agency in São José do Rio Preto, Brazil. Builds AI products running in production (Tobia, gerar.app, Pense Mercado) and maintains the AI Benchmark, a public ranking of AI models. See ROO3 AI consulting.

X @rodmf LinkedIn rodrigofavaro