Tools · 05 · built here
The Statement Extractor.
Paste a paragraph. The extractor breaks it into individual statements and shows, sentence by sentence, which claims a model can lift intact — and which dissolve into filler. A Hemingway for GEO.
—
—
No statements detected — paste a sentence or two and analyze again.
Transparent heuristic preview — a legibility check, not a guarantee. It reads the shape of each sentence on your device, never the live model response. Method from the working paper on statement-level visibility. The full 14-model probe is the LLM Visibility Audit.
Five signals, one per sentence.
Each statement is scored on the structural features that separated surviving claims from filler across a controlled corpus — the same features the GEO Scorer applies to a whole page, here resolved to the sentence.
Quantification
A number, percentage, or date gives a model something concrete to anchor and reproduce. Quantified statements survived retrieval at 2.4× the rate of bare assertions.
Named entities
A capitalised proper noun or acronym — a company, product, standard, place — is a handle the model can hold onto. Named claims survived at 1.7×.
Source anchoring
A cue that the claim is measured, audited, certified, benchmarked, or cited signals it is defensible. Sourced statements were attributed far more often than unsourced ones.
Self-contained length
Between roughly eight and forty-five words a statement carries its own context into a retrieval chunk. Short fragments and long run-ons both get cut mid-argument.
No empty superlatives
“Best”, “leading”, “world-class”, “revolutionary”, “seamless” — unfalsifiable modifiers a model cannot verify and tends to strip. Their presence subtracts from the score.
Everything is on the surface.
There is no model call. Each sentence earns points for a number, a named entity, a source cue, and a self-contained length, and loses points for unfalsifiable superlatives — so a fully-loaded claim lands near the top and pure fluff falls to the floor. The weighting traces to the working paper on statement-level visibility; the bands are strong ≥ 66, medium 33–65, weak below 33. Read the highlights, then fix the weak sentences first.
This is the preview. The probe is the proof.
The extractor reads structure on your device. The LLM Visibility Audit runs your actual statements against fourteen frontier models over a thirty-day window — verbatim, paraphrase, contradiction, decay.
Run the 14-model probe →