How to Get Cited by AI Search: Measure First, Then Write

You get cited by AI answer engines by owning one specific, checkable claim per page and structuring that page so the claim lifts out cleanly — then measuring which engines actually cite you before you change anything else. Not by adding an llms.txt file. Not by rewriting your site in “AI-friendly” prose. The order matters more than the tactics, and almost everyone runs it backwards.

This sits across two layers of the Ordered Clarity system: Outcomes asks what worked, and what compounds? Weak Outcomes create fake productivity — which is exactly what happens when you spend a quarter on AI-search tactics you never measured.

Start with the uncomfortable fact

Ranking #1 in Google no longer buys you a seat in the answer. One 2026 analysis put the overlap between Google’s top rankings and AI-cited sources at under 20%, down from roughly 70% — the two systems have genuinely diverged. Treat your search rankings as evidence about search rankings and nothing more.

The engines also disagree with each other. A 2026 study of 34,234 responses found Perplexity named brands in around 13% of answers, against under 1% for ChatGPT, with only about 11% overlap in the domains they drew from. These are single studies on a young measurement category, not settled numbers — but the direction is consistent enough to plan around: winning on one engine tells you very little about the others.

The pattern underneath the data: ChatGPT leans on established reference and editorial sources. Perplexity rewards primary sources and recent, specific pages. Those are different jobs, and one page rarely does both.

Step 1 — Measure before you optimize

You cannot improve a number you have never seen. Before writing a word, find out which engines cite you today, for which prompts, against which competitors.

Profound holds this seat in our stack — dedicated answer-engine visibility tracking across ChatGPT, Perplexity, and AI Overviews. It’s honest to say this category is early: entry is $99/mo for a single engine and real multi-engine coverage starts at $399/mo, which is a serious line item for a solo operator. If that’s out of range, Surfer layered AI-search visibility tracking onto its content scoring in its May 2026 rebrand, starting at $49/mo, and Semrush has added AI-visibility reporting to its suite.

The free version of this step: pick your ten highest-intent prompts — the questions a buyer actually types — and run them by hand in ChatGPT and Perplexity once a month. Log who gets cited. Ten prompts in a spreadsheet beats $399/mo of dashboards you don’t read.

Step 2 — Own one checkable claim per page

Answer engines lift specifics. A sentence like “pricing varies by plan” is unliftable. “Consensus Pro is $20/month, or $144/year” is a fact an engine can quote and a reader can check — and if it’s wrong, you’ll hear about it, which is the point.

Three things make a claim liftable:

• It is specific. A number, a date, a named limit. Vague copy gets summarized away.
• It is dated. Say when you checked. Freshness is a citation signal, especially on Perplexity.
• It is yours. If you’re restating what five other sites say, the engine has no reason to prefer you. Primary observation — your own test, your own pricing table, your own count — is the only durable moat here.

Step 3 — Structure so it lifts

Format does real work. Put the answer in sentence one, under a heading that matches the question. Use real HTML tables for anything numeric — engines lift tables cleanly and mangle prose about numbers. Keep one question per heading, and answer it directly underneath.

Frase is useful here for the research half — turning a target question into a brief built from what actually ranks, so you’re not guessing at the sub-questions. Write the prose in your assistant seat; use Frase for the scoreboard.

Step 4 — Skip llms.txt (for now)

This is where most GEO advice will cost you a weekend for nothing. Google confirmed it does not use llms.txt for Search — Gary Illyes said so in July 2025, and John Mueller compared it to the keywords meta tag. As of 2026, no major AI company has publicly committed to reading it in a production search product. Adoption sits around 10% across one 300,000-domain sample, concentrated in developer-tool companies where coding agents genuinely consume it.

Ship one if your audience is developers using coding agents. It costs an hour. Just don’t file it under “AI search visibility,” because nothing verifiable connects it to citations yet.

Step 5 — Watch what the traffic does

AI-referred visitors behave differently: fewer of them, further along in the decision. Tag them properly in Microsoft Clarity or your analytics seat and check whether they convert, rather than counting sessions. If the traffic arrives and bounces, more citations won’t fix it — the page is the problem.

The call

Situation	Do this
No idea whether AI engines cite you	Ten prompts, by hand, monthly
Search is a real revenue channel	Profound or Surfer’s AI tracking
Writing a page you want cited	One checkable claim, dated, in a table
Tempted by llms.txt	Ship it for developer docs only
Citations rising, revenue flat	Fix the landing page, not the visibility

The workflow should choose the tool. Right now the workflow is measurement — you’re trying to find out whether a channel exists for you at all. Buy the cheapest honest answer to that question, and don’t fund the tactics until the measurement says there’s something to win.