How to get cited by AI engines

Generative engines cite a small number of sources per answer, and the research on what earns those citations is unusually clear. This page turns it into a working playbook. If you want the background first, start with what GEO is and where the evidence comes from.

1. Put citable material in the copy itself

The GEO research paper measured which page changes increase visibility in generated answers. Adding quotations lifted visibility by around 41 percent, adding statistics by around 30 percent, and citing sources by around 27 percent. Structural changes alone did not produce those gains.

In practice this means:

• Quote named people. A sentence attributed to a practitioner, researcher or named document is material an engine can lift and attribute. Paraphrased common knowledge is not.
• Use real numbers. Give the figure, the timeframe and where it comes from. A number without a source is a liability, not an asset.
• Link primary sources inline. Cite the study, the standard, the filing or the documentation directly, at the point where you rely on it.

2. Publish first-hand data no competitor can publish

Google's guidance on AI features says unique, non-commodity content and first-hand perspective is what gets selected. The strongest form of that is data only you hold: a benchmark you ran, aggregate numbers from your own operations, a survey of your own customers, observations from your own support queue. If a fact exists only on your page, any engine that wants that fact has to cite you. Summaries of other people's articles have the opposite property: the engine can always go to the original instead.

3. Write answerable claims

Engines extract sentences, so write sentences that survive extraction. A good claim is self-contained: it names the subject, the figure, the timeframe and the source, and it still makes sense outside the surrounding paragraph. "Results improved significantly this year" dies when lifted out of context. A sentence that states who, what, how much and when does not. Open each section by answering its question directly, then elaborate. Burying the answer beneath the preamble makes the page harder to cite.

4. Keep content fresh, and refresh rather than only republish

A study across 7,683 pages and 47,097 citations measured how often engines cite content updated within the past year:

Yet only 42 percent of the cited content was recently published. The gap between those figures is the actionable part: engines reward content that is kept current, not only content that is new. An older page with updated figures, dates and examples can outperform a brand new one. Put your best existing pieces on a refresh schedule instead of letting them decay while you chase net-new topics.

5. Verify facts so they survive checking

Citable content is only an asset if it is correct, and single-model drafting produces more errors than most workflows assume. In controlled testing on hard factual benchmarks, reviewing a draft with a model from a different vendor than the writer reached 52.15 percent accuracy, against 44.74 percent for the best single model: a 7.41 percentage point gain, significant at p < 0.0001. Models from the same family tend to share blind spots; a reviewer from a different vendor catches errors the writer's family repeats.

The consequence for citation is direct. A page whose numbers do not hold up is a risk for any engine that grounds its answers, and a correction issued later does not restore the citation. This is why Vupie's article pipeline reviews every draft on a model from a different vendor than the one that wrote it, before anything is published.

The short version

1. Add quotations, statistics and inline source citations to the copy itself.
2. Publish at least some data only you can publish.
3. Write self-contained claims that answer the question early.
4. Refresh your best existing pages on a schedule.
5. Verify facts with a second model from a different vendor before publishing.