METHOD 1.0.0 · CHECK SET 2026-08-30

AI Citability Analyzer methodology

A transparent description of what the analyzer measures, what it estimates, and what it cannot know.

What “citability” means here

Citability readiness is the degree to which a page contains passages that can be extracted, understood, verified and attributed when answering a question. It is not a promise of visibility, ranking, inclusion or citation by any AI product.

The analyzer deliberately separates three questions:

  1. Retrieval eligibility: can a search or answer system fetch and index the page?
  2. Passage readiness: does the page contain independently useful, supported answers?
  3. Observed citation: did a specific product cite the page for a specific prompt at a specific time?

This tool answers the first two. A future AI Citation Tracker will measure the third.

Analysis process

  1. 01

    Fetch safely

    The service accepts only public HTTP or HTTPS URLs, follows a limited number of public redirects and caps response size. Private, local and metadata addresses are rejected.

  2. 02

    Render when needed

    Direct HTML is analysed first. Browser rendering is attempted for JavaScript-thin pages and for this site’s own pages, then the richer usable response is selected.

  3. 03

    Isolate main content

    A semantic <main> region is preferred, then <article>, then the cleaned body. Navigation, header, footer, form, script and style regions are removed from passage analysis.

  4. 04

    Segment passages

    Heading-bounded sections become candidate passages. Meaningful paragraphs, list items, quotations, definitions and table cells are retained, with a practical cap to keep one interface-heavy section from swallowing the report.

  5. 05

    Inspect claims and evidence

    Deterministic signals identify numerical or research-like claims, links in the same section, authorship, schema, dates, content landmarks, abbreviations and time-sensitive language.

  6. 06

    Score applicable findings

    Unable and not-applicable findings are excluded. A target question adds an alignment overlay; it does not silently change the base category weights.

Scoring formula

CategoryWeightWhat it represents
Evidence and source support25%Whether factual claims expose a nearby, specific evidence trail.
Passage structure20%Whether sections can be extracted with their subject and necessary context intact.
Answer contribution20%Whether the page adds definitions, procedures, comparisons, findings or other usable answers.
Entity and terminology clarity15%Whether the subject and important shorthand remain unambiguous.
Attribution and provenance10%Whether responsibility and declared original work are visible.
Freshness integrity10%Whether dates and time-sensitive claims are accurate and bounded.

Each applicable finding receives Pass = 1, Opportunity = 0.55, or Issue = 0. Findings marked Unable or Not applicable have no denominator weight. A category score is the weighted mean of its applicable findings. The overall score is the weighted mean of applicable category scores.

category = Σ(finding value × finding weight) / Σ(applicable finding weight)
overall = Σ(category score × category weight) / Σ(applicable category weight)

Threshold labels are descriptive, not industry benchmarks: 82–100 Highly citable; 65–81 Citation ready; 45–64 Promising, with gaps; 0–44 Weak citation readiness.

Check definitions

Claims have nearby support

Deterministic measurement

Detects sentences with numbers, explicit research language or measurable result language. A claim is counted as locally supported when the same heading-bounded section contains an external reference. This is a proximity check, not validation that the source proves the claim.

References are specific

Deterministic measurement

Counts unique external links in analysed content. Direct source links are useful; homepage links, broken links and weak sources still need human evaluation. Version 1.0 does not fetch every reference to judge its quality.

Passages stand on their own

Experimental heuristic

Scores practical extract length, section boundary, sentence count, a named rather than pronoun-led opening, factual specificity, nearby references and promotional filler. It is intentionally a heuristic: no universal ideal passage length exists.

Main content is extractable

Deterministic measurement

Checks whether semantic main or article boundaries isolate unique content and whether substantial sections can be recovered. It does not reward heading quantity for its own sake.

Sections contribute usable answers

Experimental heuristic

Uses passage strength to estimate whether the page contains multiple independently useful answer candidates. A marketing section with many headings does not pass merely because it is scannable.

Question and task coverage

Deterministic or model-assisted

Without a target question, the check observes structures suited to definitions, procedures and comparisons. With a target question, the service asks a small language model to estimate relevance and completeness for each passage. If that call is unavailable, a disclosed deterministic term-overlap fallback is used.

Primary subject is explicit

Deterministic measurement

Infers a subject from suitable structured data or the page title, then checks whether meaningful parts of that name appear in analysed text. Schema is treated as a consistency signal, not a source of authority.

Terminology is defined

Deterministic measurement

Flags recurring all-caps abbreviations that do not appear near a parenthetical definition. Specialist terms that are not abbreviations are outside the version 1.0 detector.

Authorship is declared

Deterministic measurement

Reads visible metadata and structured data for a named author or responsible person. It records the declaration only; it does not verify identity, credentials or expertise.

Original work is declared

Experimental heuristic

Looks for explicit statements of testing, measurement, original data, methodology or case evidence. A pass means the page makes a first-hand declaration, not that the work has been independently verified.

Publication and update dates are clear

Deterministic measurement

Checks declared publication and modification dates for articles, documentation and comparisons. A date is not rewarded on page types where it adds little value, and a recent date is not assumed to mean accurate content.

Time-sensitive claims are bounded

Experimental heuristic

Flags words such as “currently”, “latest”, “today”, explicit years and “as of”. These signals invite review so the claim has a precise date and source; they are not automatically errors.

Confidence and evidence labels

Analysis confidence describes the stability of the extraction—not the probability of being cited. It rises when the page exposes a clear main-content region, enough substantive text, multiple sections and a language declaration.

Official guidance
A platform or standards owner directly documents the behaviour.
Observed research
A published study or measured observation informs the rationale.
Deterministic measurement
The same HTML produces the same result without a model judgement.
Model-assisted estimate
A named model estimates semantic fit; the report stores its model ID.
Experimental heuristic
A disclosed proxy supports prioritisation but is not an established standard.

Limits and safe interpretation

  • The report is a point-in-time view of one URL, not the whole domain.
  • It cannot observe private model indexes, selection systems or user-specific answer behaviour.
  • It does not verify that a linked source proves the adjacent claim.
  • It cannot prove authorship, expertise, first-hand experience or editorial review.
  • JavaScript rendering can fail, time out or differ from a user’s browser.
  • English-language patterns are the best-tested. Other languages may receive lower-confidence results.
  • Pages behind authentication, consent walls or interaction gates may be only partly analysable.
  • Scores should be used to choose edits and compare rescans—not as a guarantee or competitive benchmark.

External basis

The method follows platform guidance where it exists and avoids inventing a universal “AI optimisation” standard:

Method changes that alter scores will increment the methodology version. Every stored report records its methodology version, check-set date, model ID when used, page type and analysis timestamp.