To get cited by ChatGPT and Perplexity, publish pages that answer your buyers’ real questions in the first two sentences and are structured so a machine can lift the passage without edits. Then make the rest of the web agree with you: keep the company’s facts identical everywhere they appear, and earn mentions on sites you do not own. An llms.txt file helps at the margin, and a fixed set of test questions run monthly tells you whether any of it is moving. That is the playbook. The rest of this piece works through each move.
How do ChatGPT and Perplexity decide what to cite?
Both engines answer by retrieving pages and composing from what they find. Someone asks a question; the engine runs searches behind the scenes, pulls a shortlist of candidate pages, reads them, and assembles an answer. Citations go to the pages that supplied usable passages. Usable is the load-bearing word. A usable passage states a fact cleanly, in a self-contained block, near a heading that matches the question. Pages that bury the answer under three screens of brand story rarely make the cut. Pages the crawler cannot reach never do; retrieval happens before generation, and a page missing from the shortlist has no shot at the citation, whatever its prose.
So the old hygiene still matters: a crawlable site, fast pages, a sitemap that works, no login wall in front of the useful content. Answer engine optimization, AEO in the trade, starts where search engine optimization leaves off. What changes is the unit of competition. Search engines ranked whole pages; answer engines lift passages out of them, and the passage is what you now write for.
What should your pages actually say?
Your pages should answer the questions your buyers actually ask, in the words they ask them. Collect those questions from sales calls and support threads. Give each one a page or a heading of its own, and put the answer in the first sentence with the support directly under it. Background, for the readers who want it, goes last. Most marketing pages run that order in reverse, which is why so few of them get quoted by machines.
Picture a procurement team asking an engine how long a workflow automation build takes. The answer will be some vendor’s paragraph, read back with confidence. The vendor who wrote a direct paragraph on that exact question is in the running. The vendor whose page is titled “Reimagining Operational Excellence” is not.
Structure carries the rest. Phrase headings as questions. Keep answer blocks short enough to stand alone, so a reader dropped into the middle of the page loses nothing. State checkable facts as facts: who an offer is for, what shape the pricing takes, what is included, how support works. Add FAQ schema so the question and answer pairs reach the engine already labeled. Put a named author and a visible date on the page; an undated page competes against every fresher copy of the same claim.
Why do your facts need to match everywhere?
Because the engines read your whole footprint and notice when it disagrees with itself. Models pull from your site, your LinkedIn company page, founder profiles, directories, podcast show notes, and conference bios that have outlived their conferences. When those sources describe the company differently (a stale service name here, a tagline from two rebrands ago there), the machine hedges, or worse, quotes the stale version with full confidence. A model reading five different bios has no way to know which one you still stand behind. Agreement across sources is one of the few trust signals a machine can check on its own.
The remedy is clerical, and it sits at the filing-cabinet end of AEO and AI search visibility work. Bios drift on their own, so nothing stays current without upkeep. Write one fact sheet: what the company does, for whom, where it is based, who runs it, and what each offering is called. Reconcile every public copy against it, then set a quarterly reminder to run the same pass again and fix whatever has slipped.
Is llms.txt worth publishing?
Yes, on cost grounds alone. llms.txt is a Markdown file at your site root, written for language models rather than for browsers. It holds a short description of the organization and annotated links to the pages you most want a machine to read. We publish one on this site, alongside the spec sheets and FAQ schema this article keeps recommending. Engine support for the convention is uneven; no engine owes your file a read. It is also the cheapest item on this list, and it hands the machine a clean map it would otherwise have to infer from your menus.
How do you earn mentions on sites you do not own?
Earning off-site mentions is slow, deliberate work. A company described only by its own website is a company taking its own word for it, and machines discount that roughly the way buyers do. For comparison questions especially, engines lean on third-party pages: directories, industry roundups, trade press, community threads, event listings. Ask either engine who runs corporate AI training in your city and notice how much of the answer traces back to pages that list several providers at once.
The work here is recognizable PR with a narrower aim. Find the roundups and directories the engines already cite for your questions, and get listed in them accurately. Put practitioners’ names on bylines and conference programs; an event page describes you in a third party’s words, and those words persist for years. Podcasts count as well, because show notes are text, and text is what the engines read. Answer questions in the communities your buyers frequent, under your own name, with no pitch attached.
What should you measure each month?
You measure whether you show up in the answer, and whose words carry it. No engine sends a report when it cites you, so measurement is manual and cheap. Build a fixed set of the questions you want to win, phrased the way buyers ask them, and run the set through ChatGPT and Perplexity every month. A spreadsheet with a row per question and a column per month is the entire tooling requirement. Log the same things every time:
- Presence: whether you appear in the answer, and whether a link comes with it
- Sources: whose pages framed the answer, yours or someone else’s
- Accuracy: when you are named, whether what the engine says about you is true
- Referrals: visits arriving from AI surfaces in your analytics
Keep the question set stable so the months compare cleanly. Read the losing answers closely too; the citations in them are the engine showing you its favorite sources, free of charge.
One more line belongs in the plan, in writing: nobody can promise you a citation. Retrieval gets rebuilt on the engines’ schedule, never yours, and answers shift with phrasing and session. Anyone guaranteeing placement in an AI answer has confused a forecast with a contract. What the playbook controls is the odds. The clearest liftable answer, the most consistent public record, and the best corroborated name in your category make you the likeliest paragraph for the engine to pick up, and likeliest is the strongest word available to anyone doing this work. We would be slow to trust a vendor holding a stronger one.