Layer 3 · Quotability
Generative engine optimization (GEO)
Retrieval gets you into the candidate set. GEO decides whether the model reaches for your passage when it writes the sentence that carries a citation.
Generative engine optimization is the practice of making a passage worth quoting once it has already been retrieved.
The foundational study — Aggarwal et al., presented at KDD 2024 by researchers from Princeton, Georgia Tech, the Allen Institute for AI and IIT Delhi — tested content changes across roughly 10,000 queries and found that adding quotations, statistics and citations drove the largest visibility gains, in the region of 30–41%.
01What the research found
The study's value isn't the exact percentages, which will drift as models change. It's the direction: the things that increase citation are the things that make a passage evidentially stronger, not the things that make it more keyword-dense.
| Strategy | Reported lift | What it means in practice |
|---|---|---|
| Add quotations | ~41% | Quote a named expert or primary document directly, with attribution |
| Add statistics | ~32% | Replace vague quantifiers with specific, sourced numbers |
| Add citations | ~30% | Link claims to the original source, not to another blog post |
| Improve fluency | ~28% | Clearer sentences, better transitions, less hedging |
| Keyword stuffing | No gain | The tactic that defined a decade of SEO does nothing here |
Read the last row twice. The mechanism has changed. A generative engine is not counting term frequency; it is selecting passages that let it write a confident, attributable sentence.
02Turning that into edits
Replace every vague quantifier
“Many companies” is unquotable. “61% of the 340 B2B marketers we surveyed in March 2026” is quotable, checkable and specific enough that a model can use it without hedging. If you don't have your own number, cite someone who does — by name, with a date and a link.
Quote humans, not paraphrases
A direct quotation from a named person with a stated role gives an engine an attributable fragment. “According to Dr. Lena Ortiz, head of retrieval at …” carries more weight than the same idea expressed anonymously.
Cite upstream
Link to the primary source: the paper, the filing, the official documentation. Citing another blog that cites the paper puts you a hop further from the fact and makes you the weaker candidate.
Date everything
Freshness is a genuine ranking signal in retrieval, and undated content is risky for an engine
to assert. Put a visible last reviewed date on the page and mirror it in
dateModified. Update it only when you actually change something.
03Citation is not recommendation
This distinction decides a lot of budgets. An engine may pull a fact from your page and then name a larger competitor as the recommended option in the same answer. Citation is a content outcome. Recommendation is an entity outcome — it depends on how strongly the model associates your brand with the category, which is built from third-party mentions, reference-grade sources and consistent identity data across the web.
One measured example of how this goes wrong: a 2026 analysis by Lily Ray found that brands publishing their own “best in category” lists were excluded from the recommendation about 69% of the time. Self-interested comparison content appears to be discounted. Building the entity is the work that fixes this — see get cited faster.
04Engines disagree with each other
Do not optimize for one assistant. A 2026 Otterly.ai study found Claude and ChatGPT citing the same domains only around 13% of the time — and barely 4% of identical URLs — and Profound's citation analysis reported that ChatGPT leans heavily on encyclopedic sources — with Wikipedia its single most-cited domain. Each engine reads a different slice of the web through a different index.
The practical response is not to build five strategies. It's to build one source of truth that is structurally excellent, and then to make sure the reference-grade surfaces engines lean on — Wikipedia, Wikidata, industry databases, well-moderated communities — contain accurate information about you.
Where this site's numbers come fromThe KDD 2024 figures are from the published GEO benchmark study, “GEO: Generative Engine Optimization” (Aggarwal et al., arxiv.org/abs/2311.09735). The overlap, self-ranking and engine-disagreement figures come from 2026 vendor and industry analyses summarised in the trade press; they are directionally consistent across sources but have not been independently replicated. Treat percentages as signposts, not constants.
Next, the machine-readable layer: structured data for AI.
Questions engines askFrequently asked questions
What did the Princeton GEO study actually test?
Researchers built a benchmark of roughly 10,000 queries across 25 domains, applied nine content modifications to source websites, and measured the change in visibility inside generated answers using a position-adjusted word count metric. The strategies that helped most were adding quotations from credible sources, adding statistics, adding citations, and improving fluency. Keyword stuffing did not help. Results were validated against a live generative engine, not only in the lab.
Is GEO different from AEO or LLMO?
Largely a labelling question. AEO usually refers to winning direct answers to specific questions; GEO covers any inclusion in generated content including longer synthesis and comparisons; LLMO and ‘AI SEO’ are used interchangeably with both. Day to day the tactics overlap almost completely. Use whichever term your organisation understands and don't let the taxonomy debate consume time better spent on the work.
Do I need original research to compete?
Not strictly, but it's the most durable advantage available. Anyone can restate the same twelve facts; nobody else can publish your support-ticket analysis, your pricing benchmark, or your survey of 200 customers. Original data is also the thing most likely to be cited by other sites, which compounds into the entity signals that drive recommendations rather than mere citations.