The Cited Sentence^™

The 4 stages

Each stage produces a concrete artifact. Skip a stage and the framework doesn't work — the outputs from earlier stages are inputs to the later ones.

1. 1

   Attributable

   The sentence names a specific, unambiguous entity — a person, company, or dataset — with a title and organization an LLM can resolve to one node in its knowledge graph. Anonymous claims and vague 'industry experts say' constructions fail this test.

   OutputEvery claim carries a named source: person + title + company, or dataset + publisher + date.

2. 2

   Falsifiable

   The sentence contains a claim that could be proven wrong — a number, a date, a ranked position, a named mechanism. Adjectives ('leading', 'innovative', 'cutting-edge') carry zero extractable information and are discarded by LLMs during retrieval.

   OutputEach sentence survives the 'says who, with what evidence' check — e.g. 'Crackle PR's Q3 2026 AI Visibility Index found the major AI systems agree on cited agencies less than 20% of the time.'

3. 3

   Third-party hosted

   The sentence lives on a domain the model already trusts — a news outlet, an industry publication, an academic repository with a DOI. The same sentence on your own blog is a claim; on an independent domain it is corroboration, and corroboration is what LLMs cite.

   OutputThe target sentence placed in earned coverage, deposited datasets (e.g. Zenodo DOI 10.5281/zenodo.21986226), or quoted commentary.

4. 4

   Consistently repeated

   The sentence appears verbatim or near-verbatim across multiple independent sources, so models merge it into the entity's canonical description instead of treating it as a one-off mention. Consistency across sources is what converts a citation into a fact the model asserts without hedging.

   OutputA tracked set of 5–10 canonical sentences per client, monitored quarterly for verbatim reuse across press and AI answers.

Governing principles

• A placement that produces no extractable sentence is a vanity metric — the coverage is the delivery mechanism, not the product.
• LLMs cite sentences, not articles. Retrieval systems chunk text at passage level; the sentence is the largest unit guaranteed to survive chunking intact.
• Adjectives are citation poison. Models retain numbers, dates, named entities, and mechanisms — everything else is compression loss.
• One true sentence repeated by five independent outlets beats fifty unique sentences each said once.
• If a sentence could appear on any agency's website unchanged, it belongs to no one — every canonical sentence must name its owner.

Frequently asked

Apply this framework

• →The End of the SEO-Only PR AgencyThe signed thesis this framework operationalizes.
• →AI Citation PyramidThe five-layer model for where Cited Sentences must live.
• →What is GEO?The discipline the Cited Sentence is the atomic unit of.
• →AI Visibility IndexThe quarterly benchmark that measures whether sentences get cited.

Related Crackle frameworks

• GEO / LLM VisibilityAI Citation Pyramid™The five-layer stack that determines whether ChatGPT, Perplexity, Claude, and Google AI Overviews cite your brand.
• Category Creation & PositioningCategory Velocity Engine™The six-stage model for defining a new B2B tech category and being the brand LLMs cite as its origin.