You've probably read that you should "add data" to earn AI citations. That advice is true and almost useless, because it never tells you which numbers work or how to write them. This guide fixes that: you'll learn how to add stats to get cited by AI, which data points AI cites most often, and how to format statistics for AI citations so an engine can lift your sentence as the sourced fact.
Here's the good news. You already have more evidence than you think. It's just buried in vague sentences.
The short version
- AI engines cite passages that stand alone. A number without a population, a date, or a named source is not a citable unit.
- Original, first-party data is the strongest asset you can build, because nobody else can be the source for it.
- One precise quote from a named expert beats three decorative ones.
- Schema helps engines read your page. It does not cause citations. Evidence does.
Why specific evidence gets picked up
Let's start with the mechanism, because it makes every step below obvious.
An AI engine reusing your sentence is taking a risk. It's asserting your claim to someone who might check it. A specific, attributed, recent claim is a safe thing to repeat. A vague one is not.
So the engine has two options with your page. It can preserve your wording, or it can paraphrase you into a generic statement and cite nobody. Evidence is what tips it toward the first.
The controlled evidence here is the GEO research from Aggarwal and colleagues, which tested optimization methods across a large query benchmark. Adding statistics, citations from relevant sources, and quotations were among the strongest-performing methods, with relative visibility improvements in the range of 30% to 40% and some cases above 40%. That study ran on Perplexity and Bing Chat, and those are relative results from an experiment, not a promise for your page.
There's a second reason this works. Engines pull from source mixes you don't control. Yext's analysis of 17.2 million citations from Q4 2025 found Wikipedia made up 7.8% of ChatGPT's citations and 47.9% of its top-10 source share. You're not going to out-Wikipedia Wikipedia. What you can do is match the property that makes those pages easy to cite: facts that are referenceable, definitions that are clear, sources that are named.
Use this the way it's meant to be used: as a reason to fix your evidence before you fix your adjectives.
Step 1: List every claim that's asking for trust
Read your page and mark each sentence that asks the reader to believe something. Market claims, product claims, behavior claims, comparisons, definitions. All of them.
Next to each one, write what it needs: original data, a third-party source, expert interpretation, or an honest label like "estimate" or "our opinion."
You'll know it's done when every marked claim has an evidence path assigned, including the ones you'd rather not think about.
Where people go wrong: they drop one impressive statistic into the intro and leave the other twenty assertions naked. Engines don't cite intros. They cite passages.
Tedious? Yes. It's also the step that pays. Most pages don't lose citations because their evidence is weak. They lose because nobody ever asked which sentences needed evidence at all.
Knowing the question helps here too. If you track the prompts your buyers actually ask AI engines, you can map each prompt to the claim your page needs to make. DeepSmith's AI Visibility module exists for exactly this: define the questions, watch mention rate and citation rate across engines, and see which pages get cited. Tracking tells you where to aim. It isn't evidence by itself.
Step 2: Choose the evidence, and know which data points AI cites
Every claim gets one of three paths: an authoritative external study, a first-party measurement you collect, or a labeled estimate with its assumptions shown.
Not sure which data points AI cites most reliably? Four kinds do the heavy lifting.
Original first-party data. Surveys of a defined audience, product telemetry, benchmarks against a repeatable test set, year-over-year comparisons using the same definition. HubSpot identifies original research as the leading content type for earning AI citations, and the logic holds up: other pages can summarize your study, but you stay the attribution point.
Specific third-party statistics. Name the original study, not the aggregator that repeated it. Include the year, the collection period, the sample, and the denominator.
Comparisons and deltas. These work when both sides are defined. Current period versus prior period. One segment versus another. A score against a fixed test set.
Definitions and thresholds. A definition is a citable unit even with no percentage in it. Define the term in one sentence, then give the measurement rule.
Here's a definition doing real work. Citation rate is the percentage of tracked answers that link to one of your pages. Mention rate is whether the engine names your brand at all, linked or not. Two different things, and the gap between them is wide: Semrush's Ghost Citations study reported that 62% of AI citations produced no brand mention. Write a definition that specifies what counts, what doesn't, the denominator, and the observation period, and you've built something an engine can lift whole.
You'll know it's done when every claim has a named source, a date, and enough context to interpret it.
Where people go wrong: repeating an aggregator's number without checking the original denominator. If you can't find the original study, you don't have a statistic. You have a rumor with a percent sign.
Pro tip: Always say whether a change is relative or in percentage points. A move from 20% to 30% is a 10 percentage-point increase and a 50% relative increase. Mixing those up is how good writers publish wrong numbers.
Step 3: How to add stats to get cited by AI, one unit at a time
This is the step that changes your results. Well-formed statistics for AI citations all share the same anatomy, and it's short. Every statistic you want cited needs five parts in one place:
- Claim. One declarative sentence that answers a real question.
- Evidence. The number, comparison, or exact finding.
- Source. A named study, organization, person, or first-party measurement.
- Context. Population, sample size, denominator, timeframe, geography, or baseline.
- Freshness. The publication date or collection period.
Miss two of these and a competing passage that has all five will win.
Here's the pattern to write against:
[Named source] found that [X]% of [defined population] [did something] during [timeframe], based on [sample or measurement basis].
Put your most important statistic in the first 100 to 180 words of its section. Then run the copy test: cut the sentence out of the paragraph and paste it somewhere else. Does it still answer who, what, how much, and when? If yes, it's citable. If it needs the paragraph above it, it isn't.
You'll know it's done when every key number survives that copy test on its own.
Where people go wrong: three ways, all common. Burying the number under context. Using a percentage with no denominator. Leaving the only copy of a figure inside a chart image where no crawler can read it.
Tables are fine for genuinely parallel values. Just repeat the essential figure in prose too, because prose is what gets quoted.
Step 4: Original data AEO starts with a real methodology
Original data AEO advice usually stops at "publish research," which skips the part that makes it work. Data isn't original because it's on your site. It's original because your team collected or calculated it through a defined process, and someone else could audit how.
So write the methodology down. Sample size, collection dates, population and eligibility, geography, recruitment method, your instrument or telemetry definition, cleaning and exclusion rules, and the limitations you'd rather nobody noticed.
That last one matters more than it feels like it should. Stating what your data can't prove is what makes the rest of it believable.
Let's make this concrete with our own situation. DeepSmith tracks how AI engines answer questions about a brand across ChatGPT, Gemini, Perplexity, Claude, and Google AI Mode, with four defined metrics: Mention Rate (how often an answer names the brand), Citation Rate (how often an answer links to one of your pages), Share of Voice (visibility against tracked competitors), and Visibility Trend (change across comparable periods).
If we published our own citation benchmark from that data, here's what it would have to document: the exact prompt set and how we chose it, the engines tested, the collection start and end dates, the number of runs per prompt, what counts as a mention, what counts as a citation, the denominator for each rate, the pages cited, and the results we excluded and why.
Notice what's missing from that list: a headline number. That's deliberate. If you don't have the collected result yet, publish the methodology and leave the number out. A documented method with no figure is honest. A figure with no method is the thing engines and readers learn to distrust.
The shape of a clean first-party result sentence looks like this:
From [date] to [date], we tested [number] defined prompts across [engines]. [X]% of collected answers mentioned the brand and [Y]% cited one of our pages, using the measurement rules in the methodology.
Every bracket is a decision you make before you collect, not after you see the results. That ordering is the whole discipline. Deciding what counts as a citation once you already know which number you'd prefer isn't research, it's decoration.
And you don't need a research team for this. If you're tracking prompts already, you have a dataset. Most teams sit on measurable first-party data for months without noticing it's publishable.
You'll know it's done when a skeptical reader could reproduce your study from your own page.
Where people go wrong: publishing the headline percentage while quietly omitting that the sample was self-selected, the data was self-reported, or the finding only applies to one segment.
Step 5: Add quotes that get cited
A quote earns its place when it adds evidence or interpretation. If it restates your heading, delete it.
Quotes that get cited carry four identity elements every time: full name, job title, organization, and a statement that contributes something new. Add the date, and a link to a verifiable profile where it helps.
Keep it to one to three sentences, ideally under 50 words. Short declarative statements extract cleanly. Long ones get chopped in ways you won't like.
Placement matters too. Put the quote next to the statistic it explains, and inside the first 300 words of that section when you can. Practitioner research from Allmond points the same direction: pages with three or more expert quotes saw roughly 47% more AI citations, and quotes in the first 300 words about 24% more. Those are directional benchmarks from one analysis, not laws. Test them, don't quote them as gospel.
You'll know it's done when each quote stands alone without losing who said it or why they're qualified.
Where people go wrong: "industry experts say." That phrase gives an engine nothing to attribute. An anonymous authority is not an authority. One named person with a title beats a chorus of nobody.
A few more quote patterns that reliably fail:
- Pull quotes lifted from an interview with no date or context.
- Quotes that repeat your thesis word for word.
- Long blocks carrying several unrelated ideas at once.
- Speakers whose expertise can't be verified anywhere.
- Quotes edited so heavily the meaning shifted.
If you're staring at a quote and can't say what new thing it adds, that's your answer. Cut it and keep the one that earns its space.
Step 6: Build the structure around your evidence
You've got the units. Now make them easy to find.
Open each major section with the direct answer, not a warm-up. Use headings shaped like the questions people actually ask. Keep paragraphs to two to four sentences, one claim each. Use lists for sequences and criteria.
Keep the attribution in the same sentence as the number, or the very next one. When a source sits three paragraphs from its figure, the engine can't connect them and neither can your reader.
Question-shaped headings do something quiet but useful. "How long should an expert quote be?" marks that section as a discrete answer unit, which is easier to identify than "Quote considerations." They won't rescue thin evidence. They will help good evidence get found.
If your guide runs past 1,500 words, put a short key-takeaways block near the top. Three to five bullets, each one understandable on its own, none of them contradicting a number in the body.
Worth remembering: an audit found ChatGPT cites only about 15% of the pages it retrieves. Being found is not being cited. Structure is what closes that gap.
You'll know it's done when your page's central answer is available in its first 200 words.
Where people go wrong: opening with scene-setting about how fast AI is changing everything. Your reader knows. Answer the question.
Step 7: Add schema, then QA the facts and the crawl
Use structured data that matches what's visible: Article, Person, Organization, Dataset for a real data release, FAQPage only for Q&A that's actually on the page. Validate it.
Then be honest about what it does. Ahrefs tracked 1,885 pages and found cited pages were almost three times as likely to contain JSON-LD, while also finding that adding schema alone barely moved citation results. That's correlation living next to a null result. Accurate schema clarifies your page. Evidence is what gets you cited.
Now the unglamorous part. Check every number, denominator, source, quote, date, and link. Point each source link at the specific study or methodology page, never the organization's homepage. Confirm crawlers can reach the page and your robots rules don't block the engines you care about. Record a last-updated date and decide how often you'll refresh.
Technical access is a real failure mode, not a footnote. Restricted robots rules, crawl errors, and slow server responses are commonly reported reasons an engine skips a site it would otherwise cite. A page you can see in your browser is not automatically a page an engine can read.
While you're in there, resolve conflicting numbers across your own site. If the same statistic appears with two different values on two pages, you've given every engine a reason to trust neither.
You'll know it's done when every citable unit can be verified in under a minute and the page is genuinely reachable.
Where people go wrong: keeping a stale statistic because the page still ranks. Ranking and citation are different games. One audit reported that 80% of AI-cited sources had not ranked in Google's top 10.
What to do next
Don't rebuild your library this week. Take one page.
Find its vaguest claim and replace it with one fully attributed statistic. Add one named expert quote. Link the source. Validate the markup. Then watch what the engines do with it over the next month.
That's it. One page, four fixes. Momentum matters more than a perfect backlog.
When you're ready to see whether it worked, you need to measure per engine, because their source mixes differ. One audit found only 11% overlap between the sources ChatGPT and Perplexity cite. A page that lands on one may be invisible on another. DeepSmith can define your prompt set, track mention and citation rates across engines, show which of your pages earn citations, and produce content against the gaps you find. It tracks what happens; it can't guarantee a citation.
Want to see where you stand right now? Start a free trial and check your first prompts against real answers.
