Everybody's asking "how do I rank in AI search?" It's the wrong question. AI doesn't rank your page and send a click — it lifts a passage out of some source and stitches it into an answer. So the real question is narrower and more useful: what makes a model pull a sentence from your page instead of someone else's?
I build FixAEO, a tool for tracking whether AI engines cite and recommend brands, so I spend a lot of time looking at what actually gets pulled into answers versus what gets ignored. The good news: there's real research here, and it mostly contradicts the tidy "AEO checklist" everyone's copying off each other. Here are the content changes with evidence behind them — and, because I'd rather save you time than pad this out, the two popular ones the data says are a waste.
Quick honesty note: I'll cite sources as I go, and where something is correlation rather than proven cause, I'll say so. That's kind of the whole point of this post.
First, what "getting cited" actually is
When an engine answers "what's the best X for Y," it's not ranking a list of pages. It's composing prose and reaching into a handful of sources for facts, phrases, and names. To be in the answer, you need a chunk of your content that is easy to find, easy to trust, and easy to lift cleanly. Every change below serves one of those three.
That reframe matters because it kills a lot of busywork. You're not optimizing a page to rank. You're engineering quotable passages.
The changes that actually work
1. Put citations, quotations, and statistics into your content
This is the single most evidence-backed change, so it goes first. A GEO study out of Princeton and IIT Delhi (KDD 2024) tested this directly and found that adding citations, direct quotations, and statistics lifted a page's visibility in AI answers by up to 40%. Not vague "authority signals" — literally these three moves.
Concretely: when you make a claim, back it with a real number and name the primary source in the sentence. Quote an expert or a study directly. Swap "many companies struggle with X" for "68% of them did X in 2026, per [source]." A model composing an answer loves a clean, sourced, quotable stat because it makes its answer look credible. You're doing its homework for it.
2. Lead with a direct, self-contained answer
Models lift passages that stand on their own. If the answer to a question is smeared across three paragraphs and depends on the two above it, it's hard to extract cleanly — so it doesn't get extracted.
So: phrase your headings as the questions people actually ask, and put a crisp, complete answer in the first sentence or two underneath. Then elaborate. It's the inverted pyramid your journalism teacher nagged you about, and it happens to be exactly what an extraction step wants. Bonus: it's better for human skimmers too.
3. Be specific and first-hand
Generic content is interchangeable, and interchangeable content doesn't get chosen — there's no reason to pick yours. Original data, a concrete example, a real number from your own experience, an actual opinion: these are things only your page has, which makes your page the only place to get them.
This is also why "we asked an AI to write 50 blog posts" rarely moves AI visibility. If it's the average of everything already online, it adds nothing an engine couldn't already say. Say something the internet doesn't already know.
4. Get mentioned off your own site
Here's the uncomfortable one for people who only optimize their own pages. When Ahrefs ran correlations across 75,000 brands, web mentions (0.66) and especially YouTube mentions (0.73) tracked AI visibility far more closely than backlinks (0.22). That's correlation, not proof of cause — but the direction is clear, and it lines up with where the engines actually source. Profound's analysis of ~30 million citations found Google's AI Overviews lean heavily on Reddit and YouTube, not polished vendor pages.
So a real chunk of "content work" happens off your domain: being genuinely useful in the community threads where your buyers hang out, getting your product discussed on YouTube, earning mentions in third-party roundups. You can't fake this with a schema tag. You have to be worth talking about, in the places the models are reading.
5. Structure it so a machine can lift a clean chunk
Format is doing quiet work. Short paragraphs. One idea per section. Lists and tables where the content is actually a list or a comparison. Clear, descriptive subheads. All of it makes a self-contained, liftable unit that a model can drop into an answer without dragging in noise. Walls of text hide your best sentence inside a paragraph no engine will bother to untangle.
6. Keep it current
Freshness is a real signal, especially for the more search-like engines like Perplexity and for Google's AI Overviews. A page dated three years ago with 2023 stats is a weaker citation than the same page updated with 2026 numbers and a recent "last updated." Refresh your statistics, re-date genuinely revised pages, and prune claims that have gone stale. It's low-effort and it compounds.
The two changes that don't (sorry)
Now the myth-busting, because an honest guide has to tell you what to stop doing.
Schema markup is not a citation lever. This one's everywhere in AEO advice, and the data doesn't support it. Ahrefs added JSON-LD schema to 1,885 pages and measured roughly zero citation lift — AI Overview citations actually dipped 4.6%. Schema helps machines parse your page, which is table-stakes hygiene, but it is not a weight that earns you a recommendation. Add it because it's harmless and good practice; don't add it expecting to get cited.
llms.txt won't get you cited either. It's the shiny new checkbox, and it's unproven. Google has said outright it doesn't use it, and roughly 97% of llms.txt files receive zero AI requests. Its real use today is dev and IDE agents, not answer engines. Publish one if you want — zero downside — but crossing it off your list does not move your AI visibility.
While I'm here: a backlink-only push is also weaker than most people assume (that 0.22 correlation versus 0.66–0.73 for mentions). Links still matter for classic SEO; they're just not the AEO lever they're sold as.
How to know if any of it worked
Here's the trap with everything above: AI answers aren't repeatable, so you can't eyeball whether your changes worked. SparkToro and Gumshoe ran the same "best brand" prompt 2,961 times and got the identical list twice less than 1 in 100 times. One screenshot of ChatGPT quoting you after your rewrite proves nothing. You need a real before-and-after hit rate, sampled across many prompts and across engines.
That's the gap I built FixAEO to close. It runs your buyer questions across six engines — ChatGPT, Gemini, Perplexity, Copilot, Google AI Overviews, and Google AI Mode — and tracks your citation rate over time, so you can actually see whether "I added stats and rewrote the intro" moved the needle, and on which engines. It also shows the sources behind each answer, so you can see what the model is pulling instead of you and aim your next change at it.
Honest limits: on the entry plan those scans refresh every 72 hours (not daily), it's one seat and 15 tracked prompts, and it's early software. On price, the free scan is $0 and the full six-engine tracking is $29/mo, where a lot of tools in this space start around $95–$300. That's the pitch; the six changes above work whether or not you ever use it.
The takeaway
Getting cited by AI isn't a dark art, and it definitely isn't a schema tag. It's a short list of real content changes: cite your claims with quotes and stats, answer the question directly, be specific and first-hand, get mentioned where the engines read, structure for clean extraction, and keep it current. Skip the schema-and-llms.txt theater. Then measure, because you can't trust a single screenshot.
If you want a starting point, run a free scan on your own domain, see which pages (if any) AI already pulls from, and make the six changes to the ones it doesn't.
And if you make these changes and your citations don't move, tell me — I'd genuinely like to know where this breaks.
Top comments (0)