TL;DR
- Generative engines extract sentences that contain a specific claim plus an attributable source. Marketing adjectives get skipped.
- Every paragraph should answer: who said this, when, measured how, on what sample. If three of those are missing, rewrite it.
- Replace hedges ("industry-leading", "many companies") with named entities, dated numbers, and concrete units.
- Build a house style around "one claim per sentence" and "source within 15 words of the claim".
- Audit existing pages by highlighting every unsourced quantitative statement — that's your rewrite queue.
Evidence-dense writing is the practice of packing prose with verifiable, attributable claims so that retrieval systems — ChatGPT, Claude, Perplexity, Gemini, Google's AI Overviews — can lift a clean sentence and cite you as the origin. It is the opposite of brand copy. Brand copy maximizes feeling per word; evidence-dense copy maximizes checkable facts per paragraph. For GEO, the second style wins because LLMs are trained and grounded on text that looks like reference material, not landing pages.
Why LLMs prefer verifiable claims
Retrieval-augmented generation works by chunking documents, embedding them, and pulling the chunks most likely to answer a query. The model then synthesizes an answer and — in citing engines like Perplexity and AI Overviews — surfaces the source. A chunk that reads "Our platform dramatically improves conversion" has no extractable claim. A chunk that reads "In a 2024 test across 1,200 ecommerce sites, sites using server-side tagging saw a 12% lift in tracked conversions (Source: Acme Analytics)" gives the model something to quote and attribute.
Google's own Search Quality Rater Guidelines repeatedly emphasize that high-quality pages demonstrate first-hand experience, named expertise, and supporting evidence. Anthropic and OpenAI have both published model spec documents stating that models should prefer accurate, sourced information and reduce confident assertions without backing. The implication for writers: sentences that look like footnotes will be retrieved more than sentences that look like taglines.
The anatomy of a citable sentence
A citable sentence usually contains four elements:
- A subject — a named entity (company, study, person, dataset), not "we" or "many users".
- A measurement — a number, percentage, date range, or named outcome.
- A method or context — sample size, geography, time period.
- An attribution — the source, inline or adjacent.
When all four are present, the sentence stands alone outside your page. That portability is what makes it citable. If a model needs to drag three surrounding paragraphs to make the claim coherent, it usually picks a competitor's sentence instead.
Before and after rewrites
Before:
Our customers see incredible results with our email platform — open rates skyrocket and revenue grows fast.
This has zero extractable facts. No engine will cite it because there is nothing to verify.
After:
Across 340 B2B SaaS accounts that migrated to Postmark between January and December 2023, median open rates rose from 22% to 31% within 90 days, per Postmark's 2024 deliverability report.
Named sample, dated window, two measurements, named source. A model can lift this verbatim.
Before:
Most marketers struggle with attribution in a cookieless world.
"Most" is unfalsifiable. "Struggle" is editorializing.
After:
In Gartner's 2024 CMO Spend Survey of 395 marketing leaders, 63% reported that third-party cookie deprecation has materially reduced measurement confidence (Gartner).
Before:
AI is transforming customer support at lightning speed.
After:
Zendesk's 2024 CX Trends study, based on responses from 5,620 support leaders across 20 countries, found that 70% of organizations had deployed at least one generative AI feature in production within 12 months of GPT-4's release (Zendesk CX Trends 2024).
The pattern is consistent: strip the adjective, name the study, anchor the number.
A practical rewrite workflow
- Highlight every quantitative or qualitative claim in a draft. If you can't follow it with "(Source: …)", flag it.
- Find a primary source. Industry reports, peer-reviewed papers, government datasets, vendor benchmarks with disclosed methodology. Avoid citing other blog posts that themselves cite nothing — Wikipedia editors call this citogenesis.
- Quote the number, not the spin. If a report says "grew from 14% to 19% between 2022 and 2024", use those four data points. Don't compress it to "grew sharply".
- Put the citation within 15 words of the claim. RAG chunkers often split on paragraphs or fixed token windows; if source and claim drift apart, they get separated.
- Date everything. "Recent" ages badly. "Q2 2024" doesn't.
- One claim per sentence. Compound sentences with three statistics dilute extractability — the model picks one and drops the others.
For internal data, publish the methodology. A sentence like "we analyzed 1.2M support tickets from 412 SaaS companies between March and August 2024" is more citable than a third-party stat with no methodology, because the engine treats it as primary research.
What to cut
- Superlatives without comparison ("best-in-class", "leading").
- Vague quantifiers ("many", "most", "a lot of") without a percentage.
- Future-tense predictions without a source ("AI will replace 50% of…").
- Self-referential testimonials without a customer name and outcome.
- Stock phrases ("game-changer", "revolutionize", "unlock") — they signal marketing copy to both readers and rankers. Search Engine Land's coverage of helpful content signals repeatedly flags these as indicators of low-utility content.
A simple test: print the page, cross out every sentence that lacks a number or a named entity. If more than half the page disappears, you have a brand brochure, not a reference document.
FAQ
Does evidence-dense writing hurt readability?
No, if you write tightly. The discipline forces shorter sentences, concrete nouns, and active verbs — all hallmarks of readable prose. What it removes is filler, not clarity.
What if my industry doesn't have public data?
Generate it. Survey your customers, publish anonymized usage benchmarks, run controlled tests. Primary data is the most citable kind because it has no upstream source competing for the citation.
How many citations per 1,000 words is enough?
There is no official threshold, but reference-grade pages on Wikipedia and major research outlets typically carry 8–15 inline citations per 1,000 words. Aim for that range when the topic is empirical.
