Statistical web content analysis
A browser extension that surfaces the structural and linguistic patterns behind web content — without sending a single character of text anywhere.
The problem
Template-generated articles, affiliate farms, and AI-expanded text all share structural properties that create friction — without ever triggering a clear signal. Existing tools respond by asserting authorship claims they cannot actually support. Signal Layer takes a different position: observe the patterns; report what is there; do not overclaim.
Content built for ranking rather than reading tends to repeat the same three-word clusters, rely on heavy bullet structure, and saturate pages with affiliate links.
Certain transitional phrase patterns, sentence rhythm uniformity, and low vocabulary variation surface in expanded or model-generated writing — measurably, across languages.
Confident-sounding percentages are the wrong response to this problem. They create liability without adding accuracy, and have already caused real harm in academic settings.
Output
Every output is computed locally, in your browser, against the current page. Nothing is inferred about authorship.
A single number summarising how organically the text is structured. Higher values indicate lower template and synthetic signal density. It is a structural observation, not an authorship determination.
Measures the density of structured transitional phrases, the uniformity of sentence-length rhythm (Coefficient of Variation), and the breadth of vocabulary relative to text length (Type-Token Ratio).
Detects repeated three-word phrase clusters relative to text length, differentiates genuine affiliate commission links from standard analytics trackers, and measures bullet-heavy structural layouts.
The current engine — v0.6
Each step runs inside your browser. The page is never transmitted.
The page's lang attribute is read to select the correct phrase list and calibrated thresholds. Full linguistic analysis is applied for English, Turkish, Spanish, French, German, and Italian. Structural analysis is applied for other detected languages.
The visible text content is read from the page — exactly what a reader sees. No network request is made. For PDFs, the raw file is fetched locally and parsed with a bundled copy of PDF.js.
If the page is from a scholarly domain or contains markers such as abstract, doi, references, or keywords, thresholds are adjusted to prevent false positives on formal academic language — which is naturally uniform and terminology-dense without being synthetic.
Thresholds for TTR and trigram density are adjusted by text length. A 100-word page naturally has high vocabulary diversity and few phrase repetitions; penalising it against the same scale as a 5,000-word article would be statistically unsound.
The two sub-signals and the composite Human Trace Score are computed, and a plain-language explanation is assembled from the specific patterns that were observed. Results are cached locally for 24 hours by full URL — not by domain.
Privacy by design
The zero-data architecture is not a feature that can be disabled — it is the foundational constraint around which everything else is built. There is no server to send data to, because there is no server.
No page content is ever transmitted — not to us, not to any API, not to any third party.
No account. No login. No persistent identifier of any kind.
The only optional data flow is the voluntary feedback button: domain name + score + your rating. Nothing else, and never automatically.
All analysis results are cached locally on your device with a 24-hour expiry, indexed by full URL — not domain.
Observed patterns — representative, not exhaustive
These are structural tendencies, not verdicts on any specific site. Calibration is ongoing.
Low vocabulary variation is expected and shielded from penalty — scientific terminology repeats by methodological necessity, not by template. Academic shield thresholds apply automatically.
High phrase density of structured transitional expressions, elevated trigram repetition ratio, and heavy bullet-list architecture tend to surface together in high density.
Irregular sentence-length rhythm, natural vocabulary variation, and low phrase-cluster repetition are characteristic of a single author's voice written under no template constraint.
Commission-based links (identified by known affiliate network patterns) are separated from standard analytics trackers, and their density is reported as a distinct signal.
Scope
Signal Layer does not determine authorship. It surfaces statistical patterns; it makes no claim about who wrote a page.
Signal Layer does not assert that content is AI-generated. It reports that certain structural and linguistic signals are present, and explains what they are.
Signal Layer does not upload, transmit, or log the text of any page you visit.
Signal Layer does not track your browsing activity or build a profile of your reading behaviour.
Signal Layer scores must not be used as sole evidence in academic, disciplinary, or legal decisions. They are statistical observations — one input among many for a qualified human reader.
Direction
Development follows the principle established in the lab: calibrate the current layer completely before adding the next one.
Robust multilingual structural analysis; length-aware and academic-context-aware thresholds; real-world calibration via voluntary feedback.
Understanding a page in the context of the domain it lives on — building a local, private record of observed signal patterns across multiple visits.
A personal, fully local map of the quality of the sources you read — never shared, never synced, never leaving your device.
Batch analysis, institutional licensing, and a methodology-first API for researchers who need repeatable, explainable content quality signals at scale.
Questions
No. Signal Layer does not attempt to determine whether content was written by a human or a language model. It measures structural and statistical properties of text — repetition, rhythm, phrase density — and reports them as observations, not as authorship claims. The distinction is intentional and legally significant.
Because that level of precision is not supportable by the underlying mathematics. No text-analysis tool can reliably prove authorship from statistical patterns alone, and confident-sounding percentages have already caused documented harm in academic contexts. Signal Layer reports what it can actually defend.
No. It reads only the page you are currently viewing, and only when you explicitly click "Analyze This Page." No background activity. No passive monitoring.
By default, nothing leaves your browser. The only optional exception is the feedback button: if you tap it, the page's domain name, the computed score, and your accurate/inaccurate rating are sent for calibration. This is always voluntary and clearly labelled.
Full linguistic analysis — including calibrated phrase-density detection and lexical diversity scoring — is available for English, Turkish, Spanish, French, German, and Italian. Structural analysis is applied for a broader range of additional languages. Coverage is expanding as calibration data matures.
Yes. Signal Layer can extract and analyze text from PDF documents opened in the browser, using a locally-bundled copy of PDF.js. The file is never uploaded. Scanned PDFs that contain no machine-readable text cannot currently be processed.
Signal Layer was built by an academic chemist — not a software company. The methodological approach (calibration before claims, explicit thresholds, documented limitations) reflects that background. The full methodology will be published separately.
Available now
Free. Local. No account required. Works on Chrome, Edge, Brave, and all Chromium-based browsers.
Signal Layer