The Island Test, popularised by researchers studying generative engine optimisation, asks one question of every paragraph in your content: can it be cited as a standalone unit, like an island, with no bridge to surrounding context?
A passing paragraph names its subject explicitly within the first six words, avoids anaphoric references like "this", "it", "they", or "as mentioned above", stays under 80 words, and reads as factual rather than narrative or rhetorical.
Why AI engines need this
When ChatGPT, Claude, or Perplexity answers a question, the model retrieves chunks of text — usually a paragraph or two — and either summarises them or quotes them verbatim. If a paragraph depends on the previous one to make sense, it cannot be lifted cleanly. The model either skips it or hallucinates the missing context.
Standalone paragraphs are the units of currency in AI search. The more of them you have, the more citation surface area your page exposes.
Before and after
Fails the Island Test:
"This is the most important reason. It comes down to how the compiler resolves ambiguous types. As mentioned above, the resolver walks up the lexical scope chain."
Passes the Island Test:
"TypeScript's compiler resolves ambiguous types by walking up the lexical scope chain — the same algorithm used by JavaScript variable resolution. This means a local type declaration always shadows an outer one, even if the outer one was imported from a global module."
The second version names the subject ("TypeScript's compiler"), avoids back-references, fits in ~50 words, and reads factually. An AI engine can quote it verbatim with no broken context.
You do not need 100% pass rate
The goal is to maximise citation surface area, not to write robotic prose. A 1,200-word article with 15 paragraphs and 10 standalone-passing ones gives an AI engine 10 candidate citations — more than enough to dominate AI answers in your niche.
Narrative paragraphs that flow into each other are fine and often necessary. The Island Test is a tool to apply selectively: load the first paragraph of each section with citation-worthy content, then let the rest flow naturally.
How we score it
Our auditor scores every paragraph on a 0-4 scale across the four criteria above (subject named, no back-references, under 80 words, factual tone). Pages scoring above 2.5 average tend to appear in Perplexity and ChatGPT citations within a few weeks of publication based on our benchmarks across 200 sites tracked since launch.