Short answer. The largest published study of the question found that what moves citation rates is evidence, not repetition. Across 10,000 queries, authoritative quotations raised citation visibility by up to 40%, statistics by roughly 30%, and improved fluency by 15–30%, while keyword stuffing offered little to no gain (about 10% worse than baseline on one Perplexity.ai metric in the same paper). The practical takeaway: put a sourced number in the passage, quote a named authority, and write an answer that survives being lifted out of the page with nothing around it.
Most advice about getting cited by AI answer engines is asserted rather than measured. This piece separates the two: what has been benchmarked, what is mechanically necessary, and what is folklore worth ignoring — followed by a page-level rubric you can score your own content against this afternoon.
Key takeaways
- The largest controlled study of AI citation (10,000 queries, ACM SIGKDD 2024) found authoritative quotations lift citation visibility by up to 40% and statistics by roughly 30%, while keyword stuffing offered little to no gain (about 10% worse than baseline on one Perplexity.ai metric in the same paper).
- A page is invisible to AI crawlers when its content only appears after JavaScript hydration, even though it displays correctly in a browser.
- The most liftable passages open with the answer in the first two sentences, stay self-contained, and include one plain, one-sentence definition.
- A workable target is eight sourced statistics per 1,000 words; publishing volume does not substitute for that density.
- Content changes typically show up on the browsing/retrieval surface within weeks, but on a model's trained memory only after the next training cycle, measured in months.
What did the GEO study actually find?
A 2024 benchmark measured how specific content changes moved citation visibility across 10,000 queries, and evidence-based changes won by a wide margin. Generative Engine Optimization (GEO) is the practice of structuring content so generative AI systems retrieve and cite it, and this study is its founding measurement.
Aggarwal et al., "GEO: Generative Engine Optimization" (ACM SIGKDD 2024), benchmarked content modifications and measured the change in citation visibility.
| Modification | Measured effect |
|---|---|
| Adding authoritative quotations | Up to +40% |
| Adding statistics | Roughly +30% |
| Improving fluency and clarity | +15% to +30% |
| Adding citations to sources | Positive |
| Keyword density optimisation | Minimal influence |
| Keyword stuffing | little to no gain (−10% on one Perplexity.ai metric) |
Two things are worth drawing out.
The winners are all forms of evidence. Statistics, quotations, sources. A model constructing an answer is assembling verifiable-looking claims; a passage that supplies one is more useful to it than a passage that asserts quality.
The losers are all forms of classical keyword optimisation. Not merely neutral — keyword stuffing measured negative. Tactics carried over from a decade of SEO practice are, in this specific measurement, actively harmful.
The study is not the last word — it predates the current generation of models and does not cover every engine — but it is the largest controlled measurement available, and it points consistently in one direction.
What is mechanically necessary before content quality matters?
Content quality is irrelevant if the content never reaches the engine. Three checks fail in this order, and each blocks everything after it.
The first check governs structuring a page so AI engines can crawl it at all: the answer must exist in the served HTML. Load your key page with JavaScript disabled and read what is left. Content that only appears after hydration arrives empty at crawlers that execute no JavaScript. This is the single most common reason a well-written page is invisible, and it is invisible in turn to everyone testing in a browser.
Second, AI crawlers must actually be served the full page. Name the relevant user agents explicitly in robots.txt rather than relying on a wildcard, then verify by fetching as each agent and comparing byte counts against a browser fetch. Rate limiters and bot walls that quietly return a shorter page are common, and they cannot be detected from inside the site.
Third, attribution must be unambiguous: self-consistent canonicals, single-hop redirects, structured data that matches the visible page, and real dates. An engine that cannot decide which URL owns a claim is less likely to attribute it to you.
Watch for content that renders as a placeholder in the served HTML — counters that animate up from zero, figures injected at runtime, content behind tabs. A crawler reading "0+ languages" is reading a stated fact.
What makes a passage liftable by an AI engine?
The unit an engine uses is a self-contained span of text, not a page, so each section needs to work when read completely alone. A liftable passage is one that answers a single question, in full, without needing anything before or after it to make sense.
Six properties make a passage liftable, each testable on its own:
- Question as a literal heading, phrased the way buyers ask it — not an internal product name.
- The answer in the first two sentences, with elaboration after. A narrative that arrives at the answer in paragraph nine competes badly against one that opens with it.
- Self-containment. No unresolved "as described above". Read each passage alone and ask whether it still makes sense, an approach covered further in question headings and answer-first writing.
- A definition stated plainly, in one sentence, quotable verbatim.
- Specific and dated claims. "Measured 10 August 2026" beats "recently".
- A formula or a method where one exists. Rarely used, and unusually effective — a passage that defines how something is calculated is far more quotable than one asserting that quality is high. Most vendor sites in most categories carry no formulas at all in body copy.
How do you score a page for AI citability?
A twenty-point rubric scores structure and evidence together, and anything below twelve will struggle regardless of how much traffic the page gets. Evidence density is the count of sourced statistics per 1,000 words, and it is the single heaviest-weighted signal in the rubric below.
| Signal | Points | Pass condition |
|---|---|---|
| Statistics per 1,000 words | 4 | ≥ 8, each with a source |
| Named authoritative quotations or references | 3 | ≥ 2, resolvable |
| Question-shaped headings | 3 | Every major section |
| Answer-first structure | 2 | Direct answer within two sentences of each heading |
| Self-contained passages | 2 | No cross-references needed to understand a section |
| Plain definitions | 2 | At least one one-sentence, quotable definition |
| Formula or defined method | 1 | Present where the subject allows |
| Correct, non-contradictory schema | 1 | Matches the visible page |
| Real publication and modification dates | 1 | Reflect actual changes |
| Readable with JavaScript disabled | 1 | Full text present in served HTML |
Evidence density = Statistics with a source ÷ (Word count ÷ 1000)
Eight per thousand words is a workable target. It sounds high until you count what already qualifies: dates, counts, thresholds, versions, durations, percentages — provided each has a source or is directly verifiable.
What tactics do not work, despite persisting?
Four widely-sold tactics have weak or negative evidence behind them, and each persists because it is easy to sell rather than because it measurably works.
Publishing volume is the first: sixty thin articles a quarter adds crawl cost and little else. The measured lever is evidence density, not surface area. Volume persists as advice because it is easy to sell and easy to deliver.
Keyword density targets are the second, measured as minimal influence, with stuffing measuring negative — carried over from a practice where it did once matter.
A special file or schema type that grants inclusion is the third. There is no "AI Overviews schema". llms.txt may help some AI systems and is not used by Google Search; treat it as a low-cost addition after the fundamentals, never as a route into a specific engine's answers.
Guaranteed placement is the fourth: nobody controls the output of a model they do not operate, and answers vary between runs. A close relative is optimising for one engine only — the content work is largely shared across engines; what differs is measurement and time-to-effect, so building separately per engine duplicates the expensive part and skips the cheap part.
How do you know whether any of it worked?
Measurement has to happen before you change anything, or nothing afterwards is attributable. Share of answer is the proportion of runs on a fixed prompt set in which a brand is mentioned at all, and it is the baseline every later comparison depends on.
A fixed prompt set of 20 to 40 questions per market, held constant, should cover category, comparison and brand questions, tracked the way share of answer is measured over time. Both surfaces need reporting separately: a model answering from its training weights moves on model-release timescales, while the same model with browsing enabled moves in weeks. Blended into one number, a genuine retrieval win stays hidden for months. Multiple runs per prompt matter too, because generative answers vary between runs, and raw answers should be retained — the number says something moved, the text says why.
Share of answer = Answers mentioning the brand ÷ Total answers for the prompt set
Cited share = Answers linking or attributing the brand ÷ Answers mentioning the brand
Expect the retrieval surface to move first. If it does not, check delivery before rewriting content — most "the content didn't work" outcomes are pages an engine never received in full.
How does Lifewood approach this?
Lifewood runs this as a measured programme on client sites and its own, using fixed prompt sets, a pre-work baseline, and memory and retrieval reported separately with raw run files retained and readable.
The rubric above is applied as a pre-ship gate rather than an audit performed afterwards — a page that fails on evidence density or self-containment is rewritten before publication, because retrofitting evidence into finished copy is far more expensive than writing it in. For multi-market programmes the constraint is authorship: 100+ languages, 40+ delivery centres across 30+ countries, and 56,000+ registered contributors mean content is written in-market rather than translated, which matters because the question a buyer asks in Vietnamese is rarely the English question translated into Vietnamese. See GEO vs AEO vs traditional SEO for how the three disciplines divide, alongside Lifewood's AEO and GEO services.