Audit methodology
Inspect what LLMScore checks, how the score is calculated and how suggested fixes are ordered. Use the evidence and limitations to decide whether a finding warrants a change.
What the audit measures
The audit applies eleven heuristic readiness checks to sampled website pages and supporting files. It returns bounded observations and manual recommendations, with experimental legacy numbers retained separately. It does not measure real AI answers, citations, rankings or conversions. Successful sampled pages do not represent a full-site audit.
Browse all eleven check categoriesEvidence-first measurement
Confirmed issue requires an explicit supported atomic evaluator and observed input receipts. Advisory reports optional conventions and observations needing purpose review. Unknown means required evidence or evaluator is unavailable. Not applicable means a rule prerequisite is observed absent, not repaired. Observed pass is an explicit successful bounded check, not a category verdict. Empty failure lists never create passes. Coverage and applicability are separate counts with no renormalization, and this measurement has no numeric unknown value.
Experimental legacy calculation
Experimental legacy numeric score, not readiness certification or a risk indicator. Overall = round(sum(category score × fixed weight)); categories = round(clamp(heuristic points, 0, 100)). Unknowns and optional conventions can affect legacy values. No unknown renormalization or calibrated AI citation outcome. See methodology for every category calculation and historical limitations. Calculation and limitations
Each category scores out of 100. The overall score is the rounded sum of each category score multiplied by its configured weight. These weights are heuristic choices, not calibrated predictions of citations or business outcomes.
| Check group | Each |
|---|---|
| llms.txt, JSON-LD, extractability, brand and entity metadata | 12% |
| Robots, sitemap, canonicals, social cards, internal links, content gaps | 8% |
| llms-full.txt | 4% |
4 × 12% + 6 × 8% + 4% = 100%
Actual category calculations and prerequisites
The following formulae describe the current method, not historical equivalence. Historical saved values are not rewritten. Unknown-input return values and optional legacy deductions remain limitations, not evidence of defects.
llms.txt: inspect a voluntary page reference
Start 100. Retrieval unknown returns 0; absent/empty returns 100; leading HTML pattern returns 10. Subtract min(40, 10 × sampled 404/410 count) and min(15, 5 × redirect count). Other observations do not deduct.
Observed checks and manual interpretationllms-full.txt: inspect expanded text, not inferred coverage
Start 100. Retrieval unknown returns 0; absent/empty returns 100; leading HTML pattern returns 10. Heading counts, repeated names and encoding-like patterns do not deduct.
Observed checks and manual interpretationJSON-LD: distinguish JSON syntax from schema meaning
No blocks now returns 100 (optional nonpenalty, changed from v7). Otherwise start 100 and subtract 10 per invalid JSON block. Missing types and sampled block coverage do not deduct. Clamp to 0..100 and round.
Observed checks and manual interpretationCrawler robots rules: inspect purpose before changing access
Start 100; transport/access/non-200/non-404/non-410 retrieval returns 40; observed HTML response returns 0; missing/empty returns 100. Deliberate policy patterns and missing Sitemap directives do not deduct.
Observed checks and manual interpretationStatic content: review observed structure and rendered completeness
Start 100. Current purpose-dependent heading, summary, paragraph and text-length observations do not deduct. A direct auditor call without a homepage returns 0; a public audit instead fails when homepage retrieval fails.
Observed checks and manual interpretationSitemap discovery: inspect sampled URL responses
Start 100; no discovered page URLs returns 0 (legacy unknown/coverage behavior). Subtract min(40, 12 × 404/410 count), min(40, 12 × unknown-status count), and min(15, 5 × redirect count). Homepage set differences do not deduct. These legacy unknown deductions are not evidence of defects.
Observed checks and manual interpretationCanonical declarations: inspect syntax and intended consolidation
Start 100. Separately subtract min(50, 15 × absent-page count), min(50, 15 × missing-href count), and min(50, 15 × unparseable-href count). Exact differences and parameter patterns do not deduct. Absent declarations still affect legacy points despite advisory public applicability.
Observed checks and manual interpretationSocial-card metadata: inspect declarations and image status
Start 100. Missing title deducts min(30, 10 × pages); description and image each deduct min(25, 8 × pages). Any sampled image 404/410 deducts 20 once. Unknown image checks and title divergence do not deduct. Optional omissions remain legacy point deductions, not calibrated defects.
Observed checks and manual interpretationInternal links: inspect a sampled union, not click depth
Start 100. No observed links, missing nav elements and sitemap/union differences are advisory without deductions. Missing homepage returns 0 only for a direct auditor call; public audit retrieval failure is an error.
Observed checks and manual interpretationBrand and entity metadata: inspect self-declarations
Start 0. If identity is inferred, add round(25 × title-hit ratio), round(10 × description-hit ratio), round(10 × H1-hit ratio). Organization name adds 10. Declared sameAs URLs add 10 for at least three, otherwise 3 per URL. llms.txt brand mention adds 5. Identity/help URL pattern adds 5 once. Actual attainable ceiling is 75/100, not 85: 15 external-identity points and 10 unused allowance points are not evaluated.
Observed checks and manual interpretationContent patterns: interpret purpose before alleging a gap
Start 100. Meta-description contribution is round(15 × fraction with nonempty description) - 15. Word counts, FAQ markers, page-type patterns, heading patterns and homepage length do not deduct. Missing descriptions still affect legacy points although their purpose is unverified.
Observed checks and manual interpretationHow suggested fixes are ordered
New-method priority fixes include only supported Confirmed issues. They are ordered by legacy issue severity, then the number of affected sampled URLs supported by the issue's evidence. Ties retain their input order. Optional conventions are excluded from priority fixes. This order does not establish goal fit, implementation effort or usefulness for your business.
Public file examples
These illustrative excerpts show crawler-rule and page-reference formats. They are not current file snapshots, customer results or evidence that an AI system used the files.
Crawler access
User-agent: GPTBot
Allow: /
User-agent: ClaudeBot
Allow: /
User-agent: PerplexityBot
Allow: /Page references
From llmscore.io/llms.txt
# LLMScore
## Pages
- [Home](https://llmscore.io/):
What LLMScore audits, how it works,
and a sample report. Get your
AI search readiness score.Versions, evidence and uncertainty
Current audit method: heuristic-readiness-v8-bounded-public-measurement-optional-jsonld
Use the report's methodology version, source version, audit time and page receipts to interpret the findings. Unknown or failed retrievals must not be treated as evidence that content is absent. A new methodology version is a new measurement, not proof that earlier scores are equivalent.
Matching the sampled HTML pages alone does not establish equivalent category or overall scores. Auxiliary inputs and coverage can differ between audits. Before/after tracking requires separate activation and evidence checks. Comparisons remain exploratory unless their specific comparability requirements are satisfied; they do not establish causation or business outcomes.
Readiness and AI-answer measurement
The experimental legacy number is not readiness certification or a risk indicator. Website observations are distinct from AI answers. Private provider panels measure recorded answers separately and require sign-in, configured storage, provider activation and an approved request budget. These API panels do not monitor consumer ChatGPT, Claude UI or Google AI Overviews. Point-in-time answer evidence does not guarantee future citations.
Preview availability
The public readiness preview is free and rate-limited. Running an audit does not change your website or repository. PR execution is disabled. History and schedules require authentication and configured storage and cron; email alerts are not enabled.