How to get your company cited by ChatGPT, Perplexity, and AI Overviews

Traditional search returns a ranked list and the user decides. An answer engine composes a response and cites a handful of sources it used. Those are different selection problems, and they reward different things.

A page can rank third for a query and still be the passage a model quotes, because it stated the answer in one clean sentence while the pages above it buried the same fact in three paragraphs of preamble. Position helps you get retrieved; clarity determines whether you get used.

The practical consequence is that the unit of optimisation shrinks. In classic SEO you optimise a page. In GEO you optimise individual passages, because a passage is what gets retrieved, embedded, and quoted. A page can be excellent overall and contain nothing extractable.

This is also why keyword density is irrelevant here. Retrieval works on meaning, so a passage that answers the question in plain language beats one stuffed with the phrasing you hope people search for.

The single highest-leverage habit is writing sentences that still make sense with everything around them removed. A model retrieving one paragraph has no access to your preceding section, your page title, or your navigation.

"We offer this to growing businesses" is useless out of context — there is no antecedent for "this". "Dev Luma builds multi-tenant SaaS products for founders turning a service into a software product" answers a question on its own.

The test is mechanical and quick: copy any sentence from your page, paste it somewhere with no surrounding text, and ask whether a stranger could tell what it means and who is speaking. Most marketing copy fails this badly, because it is written to be read in sequence.

Put the definitional sentence early — first or second paragraph. Retrieval systems chunk documents, and the opening chunk is disproportionately likely to be the one indexed and returned.

Answer engines are fed questions. Content structured as explicit question-and-answer pairs maps onto that directly, which is why FAQ sections consistently outperform their length in citations.

Two rules make the difference. Use the question people actually type as the heading, not a clever variant — "How much does an AI chatbot cost to build?" rather than "Investment considerations". Then answer it in the first sentence of the response, before any context or caveat.

This is the inverse of how most business writing is organised, which builds up to a conclusion. For extraction you want the conclusion first and the reasoning after, because a retrieved snippet is usually truncated and the first sentence is what survives.

Mark these up with FAQPage structured data. It does not force a citation, but it removes ambiguity about which text answers which question, and the same markup drives featured snippets and voice results.

A model composing an answer needs something concrete to attribute. "Fast, reliable delivery" contains no fact, so there is nothing to cite. "A SaaS MVP typically takes six to ten weeks to a product you can charge for" is a claim, and claims get quoted.

Numbers, ranges, timeframes, and named technologies all give retrieval something to hold. Ranges are better than single figures because they are defensible and honest — and honesty matters here in a way it does not in advertising, since a model that finds your claims contradicted elsewhere has reason to prefer another source.

Include the conditions under which your claim does not hold. Counter-intuitively, passages that state limits get cited more, because they read as reliable rather than promotional. An article that says when a chatbot makes support worse is more quotable than one that only lists benefits.

Answer engines assemble a picture of your company from many sources. Where those sources disagree, confidence drops and you get mentioned less.

Use one canonical description of what you do, and repeat it verbatim on your site, in your Organization structured data, in your llms.txt, in your directory listings, and in your social profiles. Resist the urge to reword it for each context — variation that reads as freshness to a copywriter reads as inconsistency to a retrieval system.

Make sure your structured data agrees with your visible text. Organization markup naming services your site never mentions, or an areaServed that contradicts your copy, is a signal that something is unreliable.

An llms.txt file is worth adding: a plain-language summary of the company, its services, and who they are for, at a stable URL. It is not an established standard and no engine is obliged to read it, but it is cheap and it removes ambiguity for the ones that do.

None of this matters if the crawlers are blocked. Decide deliberately whether to allow GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot, ClaudeBot, and Google-Extended in robots.txt. Blocking them protects content from training and removes you from the answers — a real trade-off, not an oversight to leave to a default.

Server-render anything you want cited. Content that only appears after JavaScript execution is unreliable for retrieval, and content behind an interaction — a tab, an accordion, a hover menu — is usually invisible. This is the most common technical reason good content never gets quoted.

Measurement is genuinely harder than for search, because there is no rank to track. Three things are practical: check referral traffic from AI assistants in analytics, query the engines directly with the questions you target and record which sources they name, and watch branded search volume, which tends to rise when a company starts appearing in generated answers.

Expect this to move slowly. Answer engines refresh their view of a site over weeks, so treat GEO as compounding structural work rather than a campaign with a launch date.

For a business serving several English-speaking markets, resist creating per-country URLs until you have the content depth to justify them. Thin duplicated pages for the EU, US, Australia, and South Africa compete with each other and dilute the authority of each.

Instead, name the markets explicitly inside the content — regulations, currencies, and examples that are recognisably local. A paragraph on POPIA alongside GDPR signals South African relevance more effectively than a /za/ directory containing the same copy.

Where regional pages do earn their place is when the substance genuinely differs: distinct compliance obligations, distinct pricing, or local case studies with named clients. That is a content problem, not a URL structure problem, and it should be solved in that order.