Methodology
Insight does not invent its own thresholds or hide assumptions. This page describes how data turns into a score, where thresholds come from, what the scan levels are and where automated analysis ends — everything you need to defend a report in front of a client.
/ 01 · Scoring
Every score in a report follows the same path — from raw measurement to a 0–100 number. No manual adjustment and no after-the-fact “expert opinion”.
Score path
The same path for every module and every scan
Source data
Lab measurement, server response, rendered DOM, crawl
Threshold
Compared against a public standard or reference configuration
Finding
pass · warn · fail — or N/A when data is missing
Severity & confidence
Finding impact × automated detection confidence
Priority
A shared P1–P4 scale across all modules
Score 0–100
Weighted aggregation of a module’s checks
Fixed scoring rules
Checks do not count equally: the weight depends on impact (high > medium > low). A passed check earns its full weight, a warning part of the points, a failure zero. A module score is earned points over maximum points, scaled to 0–100.
A module or check without data is reported as N/A and excluded from the denominator — it neither lowers nor inflates the score. A deeper scan level adds modules and scope, but does not change the verdicts of checks already measured.
A module score comes from available, completed checks. N/A checks are neither a pass nor a zero and do not enter the denominator.
/ 02 · Data sources
Insight does not install a script on the site. It reads what the browser shows and what the server returns — in one scan session.
Two analysis inputs
Rendered page
A real browser, JS, DOM and what the user actually sees.
Server response
Statuses, headers, redirects and cookies at scan time.
No script on the client site
Source catalog
Lab performance metrics
Lighthouse / PageSpeed API
Results may vary between runs
HTTP checks
Statuses, headers, redirects, cookies
Server state at scan time
WCAG tests
Automated, on the rendered DOM
Some criteria require manual review
CO₂e estimate
The SWDM v4 model based on data transfer
Not an emissions measurement or a certificate
Crawl
Links, sitemap.xml and a manual list — the page is rendered with JS
Page limit per plan; recurring issues are grouped into one finding with its scope
/ 03 · Standards
Score thresholds come from standards you can cite in a report — not from our whim.
What we cite in a report
Chrome UX / web.dev
W3C
OWASP
The Insight scale
How we map onto the Insight scale
The Insight score is not a Lighthouse score, a compliance certificate or a legal opinion — it is our interpretation of source thresholds on a shared 0–100 scale.
/ 04 · Priorities
P1–P4 is implementation order, not a synonym for severity. Severity comes from the threshold; priority combines severity, scope and confidence. P1 is not always a critical defect — sometimes it is a wide, high-confidence issue. Implementation cost is only an auxiliary recommendation.
P1–P4 scale
What moves priority
Severity
Comes from the threshold — the same problem gets the same severity.
Scope
Affected page count moves a finding along the P1–P4 scale.
Confidence
Automated detection confidence affects position on the scale.
Implementation cost
An auxiliary recommendation — it does not lower severity.
N/A — not enough data. Never zero or success.
From signal to score
Security
Accessibility
/ 05 · Scan levels
Each level runs a defined module set. Credit costs match the pricing page.
A deeper level adds scope — it does not change verdicts of checks already measured.
What gets added
No Lighthouse
Best when
Quick screening of a single URL
What gets added
Adds performance
Best when
Regular site quality control
What gets added
Security cluster (ZAP)
Best when
Audit before a release or acceptance
What gets added
Competitors + AI visibility
Best when
Extended scope for organisations
/ 06 · The role of AI
Insight uses generative models in the description layer, not in the measurement layer.
Measurement layer
Measurement, thresholds and severity always come from source data.
Description layer
Descriptions, grouping and recommendations — labelled, for review.
Rules
Measurement without a model
Scores, thresholds and severity never come from a generative model.
AI on raw scan data
It groups findings, describes impact and drafts recommendations.
Review only
AI recommendations are labelled — nothing is applied automatically.
Complete without AI too
The score and findings list do not depend on the generative layer.
/ 07 · Limits
An automated scan detects part of the problems. Honest limits are part of the methodology.
Honest limits
WCAG / EAA
Does not confirm compliance
Replaces neither manual review nor legal opinion.
CO₂e
Not a measurement or certificate
An SWDM model estimate — not a lab result.
ZAP
Not a pentest
An automated vulnerability scan, not a manual penetration test.
Crawl
Not a full enterprise crawl
No guarantee of finding every URL on the site.
Lab vs CrUX
Not the same source
Shown separately — without treating lab data as field data.
/ 08 · Comparability
Page content, third parties and the network change between scans. The report records conditions — comparisons use the same scale and thresholds.
What we lock for comparison
Time and version
Scan date and methodology version
Scope
Scan level, URLs and device profile
Completeness
Modules completed and marked N/A
Method changes
Communicated — never a silent backfill
Illustrative example · same thresholds, same scale — no fabricated scores
/ 09 · Sources
The documentation behind the thresholds and models in the report — cite it yourself.
Public sources
LCP, INP and CLS thresholds
web.devHow Lighthouse computes category scores
developer.chrome.comW3C success criteria, including AA contrast
w3.orgRecommended HTTP security headers
owasp.orgVulnerability scanner in the security cluster
zaproxy.orgEmission estimate from data transfer
sustainablewebdesign.orgChrome UX Report field data
developer.chrome.comExternal links · open in a new tab · access checked Jul 2026
Table of contents, aggregation and N/A rules, full signal→score example, explicit AI role, versioned sources.
First public methodology: scoring path, sources, thresholds, priorities, scan levels and limits.
The sample report shows the data sources, timing and limits of the analysis. Free then lets you scan one project without a card.
