Information density is how many specific, checkable facts a page delivers for the words it uses. It matters for AI search because answer engines lift short passages to support a claim, and a passage with no concrete claim in it has nothing to lend. “Write shorter” misses the point. The research behind AI citations measured two different things, coverage of a topic’s key facts and the presence of proof such as numbers, quotes and sources, and neither rewards brevity for its own sake. Below: what the studies measured, a way to count facts, and a before/after rewrite.
What Information Density Actually Means
There is no official formula, and Google publishes no such metric. A useful working definition is a ratio: checkable facts per 100 words, read section by section rather than for the whole page.
“Prices vary” is not a fact. “Google says there are no additional technical requirements to appear in AI Overviews” is one: a reader can verify it, and it changes what they do next. Density is the share of your text doing that work, as opposed to framing, transitions and filler.
Two related ideas get mixed up with it constantly:
- Content depth (or coverage): how many of a topic’s important facts the page includes at all. A 4,000-word page can have high coverage and low density.
- Information gain: facts the page has that competing pages don’t. We cover that separately in our guide to information gain SEO.
You need all three, and they fail in different ways.
Why Fact-Rich Pages Get AI Citations
AI answers work one level below the page: the system retrieves candidates, then picks passages that support specific statements in its answer. Google’s documentation on AI features and your website says AI Overviews and AI Mode may use a “query fan-out” technique, issuing multiple related searches across subtopics, and so show a wider, more diverse set of supporting links than classic results.
A supporting link sits next to a sentence in the answer as evidence. A passage that says “there are many factors to consider” cannot support anything; one that names the requirement, the number or the step can. Fan-out sharpens this: each sub-query is narrow, so the passage that answers it directly wins. More on selection in how LLMs choose sources.
One caveat Google states plainly: to be a supporting link, a page must be indexed and eligible to show in Search with a snippet, and there are no extra technical requirements or special optimisations. Density does not replace eligibility; it helps decide which eligible passage gets used.
What the Research Measured (and What It Didn’t)
The GEO paper (Princeton and collaborators, KDD 2024)
“GEO: Generative Engine Optimization” by Aggarwal et al. rewrote source pages using nine methods and measured how much of each source showed up in generative answers. The authors report that their top methods (Cite Sources, Quotation Addition and Statistics Addition) achieved a relative improvement of 30–40% on their position-adjusted word count metric. Keyword stuffing, by contrast, offered “little to no improvement.” They also found lower-ranked sources gained far more than top-ranked ones: Cite Sources raised visibility 115.1% for sites ranked fifth, while the top-ranked site’s visibility fell by 30.3%.
What it measured: controlled rewrites on a research benchmark plus a Perplexity test. Not Google AI Overviews, and it tested adding proof, not cutting words.
Surfer’s key-facts study (AI Overviews)
Surfer analysed 1,591 keywords and 57,253 URLs: the top 30 organic results for each keyword plus every URL cited in its Google AI Overview, comparing cited pages with non-cited ones. Its “Fact Coverage” metric is the share of a topic’s key facts a page includes. At the median, cited pages covered 27% of key facts against 17% for non-cited pages, which Surfer describes as 62% more facts. It also found 28% of pages covered almost none of the key facts.
What it measured: coverage, not density per word. It is a correlation (Surfer notes cited pages may simply rank higher), and Surfer sells a facts feature, so treat it as a sensible vendor study, not settled science.
Together, the evidence supports a modest claim: pages that contain the topic’s key facts and back them with sources, numbers and quotes get cited more. Neither study supports thresholds like “one fact every N words”.
Information Density vs Keyword Density, Word Count and Depth
| Concept | What it counts | What raising it does | Failure mode |
|---|---|---|---|
| Keyword density | Repeats of a term per words | Little on its own; the GEO paper found keyword stuffing did not help | Unreadable, spammy copy |
| Word count | Length | Nothing directly; Google says it has no preferred word count | Padding to hit a target |
| Content depth (coverage) | Key facts present on the page | Correlates with AI Overview citations in Surfer’s data | Long pages with the facts buried |
| Information density | Checkable facts per 100 words, per section | Makes each passage quotable and liftable | Fact dumps with no explanation |
| Information gain | Facts competitors don’t have | Gives the engine a reason to cite you, not the next page | Novelty that isn’t true or useful |
Our Keyword Density Checker is useful for spotting over-optimisation, but it is not a proxy for fact density; the two often move in opposite directions.
The Fact-Unit Test: What Counts as a Key Fact
A sentence carries a fact unit if it is:
- Specific: a number, name, date, requirement, step, condition or comparison, not a quality word like “powerful” or “important”.
- Checkable: a reader could confirm it from a source or by doing it.
- Attributable: where it isn’t common knowledge, you say whose claim it is (Google’s docs, a named study, your own test).
- Self-contained: it still makes sense if lifted out of the page. “This makes it eligible” fails; “A noindexed page cannot be a supporting link in AI Overviews” passes.
Framing sentences are fine; readers need a reason to care and a link between ideas. The test is a warning, not a quota: if three sentences in a row carry no fact unit, at least one is probably filler.
How to Measure Information Density on a Page
- Pick one section, ideally the one that should answer the main query, and note its word count.
- Tag every sentence as Fact (passes the four checks), Frame (context or transition) or Filler (removable with no loss).
- Calculate fact units per 100 words, and the share of sentences tagged Filler.
- Build a key-facts list for the topic: read the top five ranking pages and the current AI answer for the query, and write down every fact that appears in more than one of them. Those are the table stakes.
- Check coverage: tick off which table-stakes facts your page states clearly. Then list facts only you have; that is your information gain.
Coverage matters more than the ratio. A short, dense page that misses half the table-stakes facts still loses to a thorough competitor. The ratio tells you whether your facts are findable; coverage tells you whether you have them at all.
Worked Example: A Low-Density Paragraph, Rewritten
A typical intro for “how to appear in AI Overviews”:
Getting into Google’s AI Overviews can seem complicated. There are many factors to consider, and every website is different. It’s important to follow best practices and create high-quality content that your users will love. With the right approach, you can improve your chances of being featured.
Four sentences, 46 words, zero fact units. Nothing in it could support a claim in an AI answer. The rewrite, using only what Google’s documentation states:
To appear as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to show in Google Search with a snippet; Google says there are no additional technical requirements. That means a noindex tag or a nosnippet directive takes the page out of consideration. Google also says you don’t need new machine-readable files, AI text files or special markup to appear.
Three sentences, 67 words, five fact units: the indexing requirement, the snippet requirement, no extra technical requirements, the effect of noindex and nosnippet, and no special files or markup. It is longer than the original and far denser, because density is about what each sentence carries, not how short the page is.
The rewrite added no statistic it could not source; every specific came from the primary document.
Where to Put the Facts: Answer First, One Idea per Section
Facts buried at the bottom of a section help less than they could; engines and readers both look for the answer near the start. Three placement rules:
- Lead with the answer. The first sentence under each heading should state the fact the heading promises. Our guide to BLUF writing covers the bottom-line-up-front pattern in detail.
- One question per section. A heading that matches a sub-query, followed by a passage that answers it, maps neatly to fan-out retrieval. See content chunking for LLMs for how to size those chunks.
- Use tables for comparisons. Three or more options on the same attributes belong in a table, not five paragraphs.
Google’s AI features guidance also asks that important content be available as text and that structured data match the visible text. A fact that lives only in an image is hard to quote.
Mistakes That Make Pages Less Citable
The failures we see most often:
- Invented or unsourced statistics. The GEO paper tested adding statistics, but real ones. A made-up percentage is a liability: engines can cross-check it, and readers who check it stop trusting the page. Our post on statistics and AI citations covers sourcing properly.
- Letting AI drafts supply the facts. A language model will happily write fact-shaped sentences that are wrong. Every fact in an AI-assisted draft needs a source you have opened; use an AI fact-checking workflow before publishing.
- Stripping out the explanation. Facts with no “why” give readers nothing to act on. Keep the mechanism sentence; make it specific.
- Only repeating table-stakes facts. Coverage gets you into the candidate set; without facts of your own you are interchangeable.
- Measuring the whole page instead of sections. A dense FAQ can hide a vague opening. Engines lift passages, so audit passages.
Making Fact-Rich Pages a Repeatable Process
For a whole site, start with pages that already rank for informational queries but below the top spot. In SEO Rocket, rank tracking and the Search Console integration show which URLs those are; the GEO paper’s finding that lower-ranked sources gain most suggests a better-sourced passage there can change who gets cited. The site audit runs a real crawler with JavaScript rendering and flags soft-404 candidates, which are often the near-empty pages with the worst coverage.
For new content, SEO Rocket’s article writer runs validation gates (minimum length, title and meta limits) with a repair loop. The gates check structure; the facts still come from your sources, data and experience. It is a playbook proven across 1,000,000+ ranking pages: let software enforce structure, and let people own the facts.
The free AI Search Readiness Checker scores how easily ChatGPT, Perplexity and AI Overviews can quote a page, which is a quick before/after check on a rewrite.
FAQ
Is fact density a Google ranking factor?
No. Google publishes no such metric, and its AI features documentation says no special optimisation is needed beyond normal SEO. It is a writing discipline that makes passages easier to quote, and it lines up with Google’s helpful-content question about whether a page gives “a substantial, complete, or comprehensive description of the topic.”
What is a good facts-per-100-words score?
There is no validated benchmark; the published studies don’t support an ideal facts-per-word ratio. Compare your sections with the pages currently cited for the query, and prioritise key-fact coverage.
Should I make my articles shorter?
Only if the length is filler. Google says it has no preferred word count. Cut sentences that carry no fact and no needed framing; keep length that adds key facts or information gain.
Do statistics help AI citations?
In the GEO paper, adding statistics, quotations and source citations were among the most effective methods tested. The numbers must be real and attributed. A fabricated statistic adds risk, not authority.
