What Makes a Fact Citable? The Anatomy of a Statistic an AI Assistant Will Repeat

Every answer an assistant composes is built from facts it did not invent. A language model can produce fluent general prose on its own; it cannot produce a verifiable, specific number about the world. When an answer needs one, the model retrieves it from a page — and that page is what gets cited. The question for a publisher is therefore not “how do I rank for this query?” but “when the answer needs a fact, is mine the fact it can use?”

This report is about what “usable” means. We draw on the peer-reviewed GEO benchmark, the Tow Center’s citation-accuracy study, Google’s information-gain patent, Ahrefs’ freshness and citation data, Pew’s source analysis — and on a decade of publishing statistics that hundreds of institutions and journalists have cited. The conclusion is practical: a citable fact is a specific artifact with eight parts, and most published statistics are missing half of them.

1. The evidence that facts drive citations

In 2024, researchers from Princeton, Georgia Tech, the Allen Institute, and IIT Delhi published GEO: Generative Engine Optimization at KDD. They built a benchmark of 10,000 queries, generated answers with a retrieval-backed engine, and tested nine ways of modifying a source page to see which changed how much of the answer drew on it. The visibility metric was position-adjusted word count — how many of the answer’s words came from the page, weighted toward the top.

Source: Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024; GEO-bench, 10,000 queries. Keyword stuffing: −9%.

Three of the top four are the same idea from different angles: give the engine something specific and attributable to work with. The authors also found statistics addition was especially effective for fact-heavy domains such as law and government — the domains where a number is the answer.

Google’s own thinking points the same way. Its 2022 patent on “information gain” describes scoring a document by how much new information it adds beyond what a searcher has already seen, and using that score in ranking. An answer engine has the identical need in sharper form: it is assembling a response, and a page that contributes a fact the model did not already have is a page it must draw on.

2. Why provenance is not optional

The Tow Center for Digital Journalism’s March 2025 study is usually read as a story about AI unreliability. It is also a story about what happens to a fact that arrives without its source attached. Given 1,600 excerpts from real articles across eight AI search engines, the engines identified the correct source less than 40% of the time; ChatGPT Search was wrong 67% of the time and Grok 3 94%, with 154 of its 200 citations leading to error pages.

Source: Columbia Journalism Review / Tow Center, March 2025; 200 queries per engine, 1,600 total.

A fact that lives on your page as a bare number — “graduation rates rose to 64%” — can be lifted, repeated, and attributed to someone else, or to no one. A fact that carries its entity, date, source, and method in the same sentence is far harder to detach from you, because the attribution is part of the fact. Provenance is not a courtesy to the reader. It is how you stay attached to the number when it travels.

3. The eight parts of a citable fact

Take a statistic from the first report in this series and write it two ways.

Written to be ignored: Studies show people click less when AI summaries appear in search results.

Written to be cited: In March 2025, U.S. Google users clicked a search result on 8% of searches that showed an AI summary, versus 15% of searches that did not, according to a Pew Research Center analysis of 68,879 searches by a panel of 900 adults.

Both sentences are true. Only the second can be used by an engine that has to justify what it says. Pulling it apart gives the eight parts:

1. Entity — U.S. Google users. What, precisely, the fact is about. Assistants organize knowledge around entities; a fact about nothing in particular attaches to nothing.
2. Value — 8% and 15%. The number itself, stated, not implied by an adjective.
3. Unit — of searches. Percent of what. Half of all ambiguity in published statistics is a missing denominator.
4. Date or period — March 2025. When the fact was true. Ahrefs found AI-cited pages average 25.7% fresher than organic results; a dated fact can be judged current, an undated one cannot.
5. Source — Pew Research Center. Who produced it. The Pew data shows .gov sources cited in AI summaries at three times their share of ordinary results — the canonical source of a number gets chosen.
6. Method — a panel of 900 adults, 68,879 searches. How it was produced, in a clause. Enough for a reader, or a model, to weigh it.
7. Comparison — versus 15% without. A number alone is trivia; a number against a baseline is an insight. The GEO authors’ finding that statistics help most in fact-heavy domains is a finding about comparisons: the answer to “is that a lot?” is what people are asking.
8. Permanent address — the fact lives at one stable URL that will exist next year. Citations, links, and the trust that accrues to them attach to addresses. A fact that moves is a fact that starts over.

Most published statistics have parts 2 and 5 and one or two of the rest. The difference between the two sentences above is not writing skill. It is a decision about what a fact is.

4. Structure: putting the fact where the machine looks

The eight parts make a fact citable. Structure makes it findable. Extraction systems look in predictable places — the first sentence under a heading that matches the question, a table with labeled columns, a definition list, an FAQ pair, and the page’s structured data. A fact buried in the fourth paragraph of a narrative section may be citable and never cited.

The reliable pattern is redundancy without contradiction: state the fact in prose, show it in a table or chart with the same value and units, and emit it in the page’s schema markup. When those three agree, the machine-readable layer confirms the human-readable one. When they disagree — and on hand-maintained pages they disagree constantly, because someone updated the paragraph and not the table — the engine has a reason to distrust the page.

5. Why templated facts are citable by construction

This is where a data-driven publisher has an advantage that is not about effort. When a page is rendered from a knowledge graph, each of the eight parts is a field, not a writing decision:

• The entity is the record the page is about, with a consistent name and identifier across every page that mentions it.
• The value and unit come from the data, and the same field renders identically in the sentence, the table, and the JSON-LD.
• The date is the data release the value came from, stamped automatically.
• The source and method are properties of the dataset — a named public source and a published methodology — inherited by every fact derived from it.
• The comparison is computed: the entity’s value against its cohort, its state, its peers, last year.
• The permanent address is a stable URL keyed to the entity, which regenerates in place when the data updates rather than spawning a new page.

A writer producing one article can do all of this once, carefully. A template does it for every entity in the category, every time the data changes. That is why the statistics on a data-driven page are cited at a rate hand-written pages struggle to match: not because the facts are better, but because they are complete.

Methodology and caveats

Sources and their limits. The GEO benchmark measured visibility inside answers generated by a retrieval-backed engine in a controlled setting; it did not measure clicks, and generative engines have changed since 2024. The Tow Center study tested source identification from excerpts, a harder task than citing a fact in an answer, so its error rates are an upper bound on attribution failure for that task, not a citation rate for ordinary queries. Pew’s source shares are for Google AI summaries in March 2025. Ahrefs’ freshness figure comes from 16.975 million citations across five platforms and uses publication dates where available. The eight-part framework is ours; it is a synthesis of what these studies reward, not a finding of any one of them.

What to do about it

1. Audit your top pages for the eight parts. For every statistic on a page that matters, check for entity, value, unit, date, source, method, comparison, and a stable URL. Most audits find three or four; the missing ones are the work.
2. Write the source into the sentence, not the footnote. “According to X’s analysis of N in [period]” travels with the fact. A footnote does not.
3. Add the comparison. Every number should answer “compared to what?” in the same breath. That is the difference between a fact and an insight, and between a page that informs and a page that gets cited.
4. Make the prose, the table, and the schema agree. If you cannot guarantee it by hand, generate all three from the same data.
5. Date everything and refresh on the data’s schedule. The freshness gap is real and measurable; close it by design.
6. Give facts permanent homes. One entity, one URL, updated in place. Citations compound only on addresses that persist.

Sources

1. Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande — GEO: Generative Engine Optimization (KDD 2024) — GEO-bench, 10,000 queries; Table 1 visibility changes by method
2. Columbia Journalism Review / Tow Center — AI Search Has a Citation Problem (Jaźwińska & Chandrasekar, March 6, 2025) — 8 engines, 1,600 queries
3. Search Engine Journal — Google's Information Gain Patent (analysis of US patent granted 2022)
4. Pew Research Center — Google users are less likely to click on links when an AI summary appears in the results (July 22, 2025) — Source shares in AI summaries vs standard results
5. Ahrefs — Do AI assistants prefer to cite fresh content? (Law, July 28, 2025) — 16.975 million citations