service-page-audit

name	service-page-audit
description	Audit, rate and fix a marketing service page under /services/*. Use when asked to review, rate, compare against a competitor, or fix leaked/generic copy on any of the 62 service pages. Encodes the conventions established across PRs

Service Page Audit & Fix

62 service pages render from one template. This skill is the repeatable version of a review that was rebuilt by hand four times and got several measurements wrong in the process. Run the tool first, then work the checklist.

---

0 · Hard rules — read before touching anything

Do not undo what already shipped. Every item below was a deliberate decision. If a fix seems to require reversing one, stop and ask.

Platform patterns — reuse, do not rebuild

Pattern	Where	Do not
POLICY_ETHICS_STACK — exported from servicesData.js. Four layers: Policy & Regulatory Controls, Responsible AI & Ethics, Risk Governance & Assurance, Human Oversight & Accountability	Adopt with architectureNodes: POLICY_ETHICS_STACK	Re-write these four layers inline on another service. It is deliberately shared across Intelligent Automation, Agentic AI, AI Governance, Decision Intelligence and Kangqore View so the answer cannot drift between pages
DEPT_SECTION_COPY in getParityService — per-department industryHeading, midCta, closingCta, architectureEyebrow	Template	Move back to per-service data, or change a department default without checking every page in that department

Template hooks — all opt-in, all with defaults that preserve the other 61 pages

Hooks	Section
architectureEyebrow / architectureTitle / architectureTitleHighlight / architectureLede	Architecture
capabilitiesLabel / capabilitiesSectionTitle / capabilitiesSectionHighlight / capabilitiesLede	Capabilities
comparisonTable.eyebrow	Comparison
outcomesEyebrow / outcomesHeading / outcomesHeadingHighlight	Engagement outcomes (the h2 is opt-in; without it the block has no heading at all)
engagementEyebrow / engagementHeading / engagementHeadingHighlight / engagementLede	Five ways to start
faqEyebrow / faqHeading / faqHeadingHighlight	FAQ
midCta / midCtaLabel	Mid-page CTA
closingCta.primaryLabel / .secondaryLabel / .proofLabel	Closing CTA
conciergeChips / conciergeHeading / conciergeIntro	eQORE
practiceLabel / practiceHeading / practiceHeadingHighlight / practiceLede	Practice cluster
hidePartnershipModel / hideBadgeStrip	Section removal
heroTitle / heroBadge / heroStripItems	Hero

Never hardcode any of these back into the template.

Content decisions

Decision	Do not
8 capability areas, 51 sub-capabilities, no numbering (ai-cognitive-computing)	Re-add numbering or collapse the taxonomy
Titles 45-60 chars, department word dropped ("Cognition", "Foundry")	Re-add department prefixes
Bespoke hero SVGs on 7 slugs: agentic-ai-led-application-modernization, genai-business-services, ai-governance, ai-cognitive-computing, mlops, data-science-ai, intelligent-automation	Replace any with the shared agentic default
whatIsPara3/4 stay in the DOM when collapsed (grid-template-rows + inert)	Re-gate behind a mount condition
Real <table> semantics + explicit ARIA on the comparison section	Return it to a <div> grid
Contextual internal links in toolchain and comparison (row.link, item.link)	Strip them
Accessibility landmarks on EQoreChatbot, CookieConsent, both FloatingButtons stacks, and ServiceGlassCards scroll region	Remove any — every page is currently axe 0 and the right-hand floating stack is easy to miss

Gates and tooling

Decision	Do not
Copy gate Rule 2b — catches percentages split across literals (value: '94' + suffix: '%'), which Rule 2's regex cannot see	Remove or "simplify" it. Twelve services were publishing unsourced percentages this way while the gate reported clean
48 metric objects across 12 services carry illustrative: true	Unflag any without a real source
Generator emits comparison, architecture, industries, outcomes, packages, closing CTA, eQORE chips + heading + intro, heroBadge, midCta	Remove emitters — coverage was 6-30 per cent before them
Rubric excludes bands under 250px from the thin-section check (a CTA band is sparse by design)	Confuse this with lowering the 10 w/100px threshold
Rubric finds the FAQ by counting question-form h3s, not by matching heading text	Go back to text matching — rewriting a page's FAQ heading broke its own detection

Standing constraints from the user

• Do not touch images. Not the mapping, not dimensions, not the generator manifest. Report findings, change nothing.
• The oversized logo (4500px at ~104px) and the 152KB PNG are knowingly ignored.
• "Shield" in breadcrumb schema is knowingly ignored.
• eQORE cannot be cut. Make it page-specific instead.
• /services/ai-governance is parked.
• Illustrative metrics stay. The user was told that disclaimed numbers read to a sophisticated buyer as "no clients", chose to keep them and fix the worst offenders, and that decision stands. Absolute claims ("100% operational reliability", "maximum operational yield") are the part to remove.
• "Robotic Operations Center" on intelligent-automation is Tech Mahindra's named offering. Flagged, kept as supplied, rename parked.
• E-E-A-T is blocked on a real name, title and LinkedIn. Never invent a person.

---

1 · Set up

One tree. /Users/maheshkumar/Kangqore, served on :3000.

The user instructed this on 2026-08-22: "ignore kq-acc [the second copy of this repo]". A git worktree at /Users/maheshkumar/kq-acc (:3010) still exists on disk and must not be used — do not build there, do not measure there, do not cd there.

cd /Users/maheshkumar/Kangqore
git status --porcelain            # must be clean before you start
git checkout -b <branch>          # branch in place; do not create a worktree

Why this rule exists. The worktree was meant to keep :3000 stable while a page was being rebuilt on a branch. What it actually produced was the user refreshing :3000, seeing the old page, and asking "where is the change??" — three separate times, across the analytics, big-data and DPA pages. The isolation was worth less than the confusion cost. Build where the user is looking.

--base is always http://localhost:3000. If a measurement is ever taken against :3010, the number is about a tree nobody is looking at.

Before trusting any measurement, confirm the server is serving your edit:

curl -s localhost:3000/src/data/servicesData.js | grep -c "<a string you just added>"

cd does not persist between Bash calls — use git -C <path> for any git command that must land in a specific tree.

2 · Measure

node scripts/audit-page-quality.mjs <slug>
node scripts/audit-page-quality.mjs <slug> --compare=<competitor-url>
node scripts/audit-page-quality.mjs --all --json     # takes over 10 min, run backgrounded

Everything reported comes from the rendered page, never the data file. Exits non-zero on render errors or inherited copy, so it can gate a build.

Scoring is rubric-anchored (thresholds in RUBRIC). Two runs agree. Do not hand out a score that is not from the tool — an unanchored number cannot be defended when challenged, and cannot be compared across pages or over time. A second "commercial readiness" score was invented once and had to be withdrawn for exactly this reason.

Known limits — do not quote these numbers without the caveat:

• Cross-tree audits are safe now, but know why. The snapshot is fetched from ${BASE}/prerender/services/<slug>.html — the same server as the page — so the two cannot come from different trees. Before that fix it read snapshots off the local filesystem and once reported 93.5 per cent coverage that was really 29. If the server does not serve that path the tool falls back to the local file, prints a WARNING naming the base, and marks the report line [LOCAL FILE — not verified against --base]. Treat any run showing that marker as unverified.
• Coverage reads about 3 points low and is noisy. The denominator counts FAQ text twice — once in the FAQ section, once in the animated glass card that types it out — while the snapshot emits it once. Page word count swings roughly 77 words run to run for the same reason.
• Boilerplate is measured against rendered innerText, so it includes UI chrome. Comparable across pages, not to per-data-block figures from older sessions.

3 · Work the checklist

Cheapest lever first.

1. Inherited copy — any hit is a hard fail.
2. Render errors — hard fail. Usually a wrong data key or prop name (§5).
3. Content fit — does the copy belong to this service? Read the capability names, the comparison argument and the toolchain against the service name. A toolchain naming none of the service's own platforms is the loudest signal.
4. FAQ depth — 8+ questions, 3 paragraphs, real sources. Usually the single biggest scoring lever (worth up to 4 points across two groups) and the best content.
5. Snapshot coverage — fix in scripts/generate-prerender.mjs, not the React page.
6. Boilerplate and second person — rewrite in the page's voice. Concrete nouns and "you", not "enterprise-grade frameworks".
7. Thin sections — under 10 w/100px and over 250px tall.
8. E-E-A-T — blocked on the user. Ask; never invent.

4 · Gate, in this order

node scripts/generate-prerender.mjs
npm run audit:all          # check the exit code, do not grep for "pass"
npm run check:prerender
npm run check:sitemap
cd frontend && npm run build

Then open the page in a browser. All four have passed on a page that fell to the error boundary rendering twelve words.

Re-run audit-page-quality.mjs <slug> and confirm each point moved for the reason you expected. Spot-check three unrelated services whenever you touched the template.

Revert shared/serviceIndex.json if only its generatedAt changed.

5 · Traps that have actually bitten

The template resolves content through getParityService(). Most services carry a few hundred bytes and the parity layer synthesizes the rest, keyed by deptSlug. Never conclude anything about what renders by reading servicesData.js. Counting raw keys produced two wrong answers in one session: "industryHeading leaks on zero pages" (56) and "EXECUTION LOOP is on 3 pages" (56).

Dual emission. Content added to the React page must also be added to generate-prerender.mjs or crawlers never see it. Grep the regenerated snapshot.

Verify what renders, not what you wrote. Check the collapsed default state.

Vite misses whole-file writes. After a scripted rewrite, touch and re-curl.

The copy gate scans code comments. US English only. unauthorised / authorisation are absent from its UK list; modelling and cataloguing are caught.

Prop names, not just data keys. GeminiComparisonSection takes comparisonTable, not data. Writing data.eyebrow threw and took the whole page to the error boundary while all four gates passed.

Data key contracts:

Key	Contract
architectureNodes[].features	Not keyCapabilities. Wrong name throws.
architectureNodes[].icon	Must exist in JOURNEY_ICON_MAP or it silently falls back to Target.
capabilityAreas[].items	'Name: prose' splits on the colon for the card front; a bare name renders fine on both front and overlay. Card front shows the first 6 only.
industryUseCases[]	items or agents; prefer items. Section renders only if the key exists.
closingCta	Whole-object override. Set title explicitly or the default fills it.
businessMetrics[]	Needs illustrative: true per metric for the disclaimer.
heroTitle	Must not end with a period. The template renders the last word as {titleHighlight}. with the period baked in — a trailing period gives "Defend..".
servicePackages[]	tier renders only inside the duration block, so tier without duration is invisible.

Editing a service block with a script. capabilityAreas is often the last key, and blocks are followed by section banner comments — so "replace from key to end of block" drops the object's closing brace, and rstrip() does not land on },. Locate blk.rfind('\n },') explicitly and assert the result still closes the object.

Contrast measurement. Walk up the DOM for the first non-transparent background and composite the alpha. Measuring against document.body invents failures — that produced a bogus "9 failing footer nodes" when the real count was 1.

6 · Ship

One branch, one commit, pushed from the main tree. gh is not installed — hand the user the compare URL, and lead with the URL where the change is actually visible, not the branch name. Saying "applied" about work on an unmerged branch the user cannot see has caused repeated confusion.

State plainly in the commit message which changes are template-wide and which are page-scoped.

7 · Open items, not yet fixed

• The parity-layer snapshot gap. generate-prerender.mjs reads raw service data while the page renders through getParityService(), so ~55 thin services emit almost nothing — salesforce at 2 per cent, it-security-services at 6, median fleet coverage 2.9. Largest technical defect on the site. Fix by extracting getParityService into a module both the component and the generator import.
• 110 fabricated case cards. The parity outcomeCard / outcomeCard2 defaults at UniversalServicePage.jsx:700-715 invent a client descriptor ("Global Enterprise Organization") and assert "100% operational reliability" / "maximum operational yield" on all 55 services lacking their own. The user has no real clients yet.
• 57 duplicate h1s — every service but a handful renders "<Name> Solutions at Enterprise Scale."
• RPA, BPM and DPA still share ~78 per cent of their content with each other. Intelligent Automation was differentiated out of that cluster; the other three were not. Consolidate-or-differentiate is still an open commercial decision.
• No analytics installed. Nothing shipped is measurable.