How to get your company cited by ChatGPT and Perplexity

A buyer used to open ten tabs. Now they ask an assistant for a shortlist, and three names come back. If yours is not one of them, you never find out, because that decision leaves no trace in your analytics.

What does it mean to be cited by ChatGPT?

A citation is when an answer engine names your company or links one of your pages inside an answer it generates for a user. It comes in two forms, and they behave differently.

The first is a linked source. Perplexity, search-connected ChatGPT and Google AI answers pull live pages, quote from them and footnote the URL. The second is an unlinked mention, where the model names you from what it already knows without fetching anything. The linked kind is faster to earn and easier to test. The unlinked kind is worth more, because a model naming you unprompted reads to the buyer as a recommendation rather than a search result.

Both come down to the same thing: your company has to be the clearest available answer to a question someone actually asks.

How do answer engines choose what to cite?

A model does not rank your homepage. It breaks the user's question into several narrower ones, retrieves candidate passages that look like answers to each, judges them on relevance and apparent reliability, then writes one response and attributes the parts it leaned on.

The practical consequence is that the unit of competition is the passage, not the page. A 2,000-word article with the answer buried in paragraph fourteen loses to a 300-word page that states the answer in the first sentence. This is where generative engine optimization departs from classic SEO: you are not competing for a click, you are competing to be the sentence that gets quoted.

Answer the question in the first forty words

Write the answer before the context. Every page should have a question-shaped heading, a direct answer immediately underneath it in plain declarative language, and the evidence after that. If a reader could screenshot your first two sentences and have a complete answer, a retrieval system can lift them cleanly.

Three habits make passages extractable:

• Define before you discuss. Open with what the thing is, in one sentence, using the words the buyer typed.
• Keep answers self-contained. A passage that starts with "as we saw above" is useless once it is lifted out of the page.
• One question per section. Headings that bundle three ideas produce passages that answer none of them well.

Hedging is the most common failure. A page that spends four paragraphs on "it depends" gives a model nothing to quote. Commit to an answer first, then qualify it.

Publish numbers nobody else has

Models cite what only one source says. If your page repeats the same statistic as forty other pages, there is no reason to pick yours. If it reports something from your own work, you become the origin.

For an outbound agency, that means publishing what we actually see: that cold email reply rates across B2B typically land between 1% and 5% of delivered emails, what pushes a campaign to the top of that band, and what drags it under. Where a number is uncertain, give the range and say what moves it. A stated range with an explanation is more citable than a false precision, and it survives being quoted back at you.

You do not need a research budget. You need to write down what you already know from doing the work, and to be specific about the conditions under which it holds. Our benchmarks page exists for exactly this reason.

Does FAQ schema actually help?

Yes, though not as a switch that turns citations on. FAQPage schema hands a crawler an explicit question and answer pair with no parsing to guess at. That removes ambiguity, and ambiguity is what gets a source skipped.

Two rules matter. The schema must match the visible text on the page word for word, because a mismatch reads as manipulation. And the questions must be ones people ask, phrased the way they ask them, not marketing headings dressed up with question marks. For B2B sites this is the highest-leverage markup available, since most buying research is literally a sequence of questions.

What about llms.txt?

llms.txt is a plain text file at the root of your domain that lists your most important pages with a one-line description of each, so an AI crawler can find the substance without crawling your whole site. It is a proposed convention rather than a standard, and adoption across the major crawlers is still uneven.

Ship one anyway. It costs an hour, it cannot hurt, and it forces you to decide which twenty pages actually represent your company. Ripe Leads publishes one at /llms.txt as a worked example.

Get named where you cannot edit

The part you cannot write yourself does most of the work. Models weigh corroboration: a company described consistently across several independent sources looks more real than one that only describes itself.

That means getting into third-party comparisons, category roundups, directories, review sites, podcast episodes and other people's articles. Ask satisfied clients to leave a public review. Answer questions in communities under your own name. Appear as a guest where your buyers already read and listen. This is slow, it does not scale on demand, and it is the single strongest signal you can build. Sustained content work is what makes those mentions possible in the first place.

Keep your facts identical everywhere

Contradictions make a model hedge or drop you. If your pricing appears three different ways across your site, your LinkedIn page and a directory listing, none of the three is safe to quote.

Pick the canonical version of every fact and repeat it exactly: what you sell, where you are based, which languages you operate in, who founded the company, what it costs. Ours reads the same everywhere, including on our pricing section: a flat EUR 3,750 for the first month covering setup and launch, then EUR 2,850 per month, cancel anytime. Same wording, same numbers, every page.

What you cannot control

Be honest with yourself about the limits, because plenty of people will sell you certainty here.

• There is no paid placement. You cannot bid your way into a generated answer. Anyone guaranteeing AI citations is guaranteeing something outside their control.
• Answers are not deterministic. The same question, asked twice, can return different companies. One test proves nothing.
• Phrasing changes everything. "Best lead generation agency in Europe" and "who can run outbound for my Lithuanian software company" retrieve almost different worlds.
• Model knowledge lags. Unlinked mentions depend on training cycles you have no visibility into, so brand-level recall moves in quarters, not weeks.
• There is no dashboard. No Search Console equivalent exists for ChatGPT. Manual prompt testing and watching referral traffic from AI sources is currently the whole toolkit.

A checklist for one page

Before you publish anything you want quoted, run it against this:

• The page targets one question, and the title is close to how a person would ask it.
• The answer appears in the first forty words, stated plainly, with no hedging.
• At least one number or detail on the page originates with you.
• Headings are questions or definitions, and each section answers exactly one.
• Every passage stands alone if lifted out of context.
• FAQPage schema is present and matches the visible questions word for word.
• Company facts on the page match every other page and every off-site profile.
• The page is linked from somewhere crawlable, and ideally mentioned somewhere you do not own.

None of this is a trick. It is the same discipline that made a page worth reading before answer engines existed, applied with the knowledge that the reader is now sometimes a machine summarising you for someone who will never see your site. If you would rather spend that effort on pipeline instead, we run outbound as a done-for-you service and you keep writing the pages only you can write. For the wider shift this sits inside, see what actually changed in B2B marketing this year.