Structured data on every route, and a dateModified derived from git

[robbyrob42]

A GEO audit on 2026-09-04 scored the property at 51/100 on the root page and 43.25/100 across all eight routes, both in the tool's foundation band.

The measured categories made the diagnosis easy. Meta was 14/14 on every page. Robots allowed all 27 AI crawlers. Raw HTML carried 777 words and 19 headings without JavaScript, so the prerender is doing its job. What was missing:

Category	Before
Schema	2.25 / 16
llms.txt	0 / 18
AI Discovery	0 / 6

Six of eight routes emitted no structured data at all.

Structured data on every page

The previous convention emitted Organization on the home page only, reasoning that it describes the property rather than the page, so repeating it on five routes states one fact five times. That is right about authorship and wrong about retrieval — an answer engine parses each URL independently and has no mechanism to carry the root page's entity across to /method.

The fix is a per-page @graph with stable @ids rather than copy-paste. The Organization is one entity referenced from every route, which answers the original objection properly: one fact, stated once, addressed by identity. Pages that support more say more.

Route	Primary entity	Also carries
/	WebPage	Organization with makesOffer
/engagements/	WebPage	makesOffer, FAQPage from the deck's existing questions
/experts/	CollectionPage	ItemList roster linking each profile
/experts/[slug]/	ProfilePage	Person, worksFor as an @id reference
/scope/	ContactPage
/method/, /matters/, /privacy/	WebPage

makesOffer rides only on the two pages that actually display the rate card.

dateModified derived from git

A date written by hand is a claim nobody verifies, and a page that says it was updated last week when it was not is the kind of small untruth this property avoids everywhere else.

scripts/content-modified.mjs records the last commit touching each route's sources before the build. A route with no recorded date emits no dateModified rather than a fabricated one. The deploy workflow now checks out full history — git log on a shallow clone reports the shallow boundary, so every page would otherwise claim the date of the most recent push.

The assumption worth reviewing: a route's inputs are its page file plus the copy deck. The deck is one file, so a copy edit anywhere moves the date on every route that reads it. Coarse, but not false — a change to a page's source inputs is a modification of that page.

AI discovery

llms.txt, /ai/summary.json and /ai/faq.json are generated from the copy deck for the same reason knowsAbout is: a hand-kept second copy of the site's structure goes stale the first time a route changes. /.well-known/ai.txt states the crawling posture explicitly and disallows /api/, which carries live matter detail written before a conflict check has run.

Deliberately not done

• No SearchAction in the WebSite node. The property has no search endpoint, and declaring one that 404s is a claim the site cannot honour. The audit tool suggests it generically.
• No areaServed, still. The copy nowhere states which jurisdictions this practice takes. That remains a copy decision the markup follows rather than leads.
• geo fix --apply was not used. Its generated FAQPage is in Italian ("Cos'è forensics.rootsystem.com?"), its descriptions carry mojibake, and its SearchAction points at a route that does not exist. The plan was taken; the files were written by hand.

Needs your input

sameAs is wired on both Organization and Person, defaults to empty, and is emitted only when populated. The audit scored Social Trust 0/5 and Academic Trust 0/5 on its absence and it is the cheapest E-E-A-T signal available — but every URL has to be a profile that exists and that the practice controls, since a sameAs pointing at something that is not the same entity merges two entities in the index. LinkedIn, ORCID, Google Scholar. The fields are in src/copy/landing.ts.

Verification

• yarn workspace @rootsystem/forensics build — clean, 11 routes prerendered including the three new endpoints.
• astro check — 4 errors before, 2 after. Both remaining are pre-existing parser complaints in scope.astro.
• All eight pages verified to emit exactly one valid JSON-LD block.

Re-audit after deploy to measure the real score against the tool's 79/100 projection. Note that llms.txt alone is 18 of those points and the tool's own output calls it "an organizational signal, not a proven ranking factor" — the schema and sameAs work is the part expected to move citations.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EDibtmcGWcytbgVpb2su7F