# Google Scholar discovery and citation metadata Each publication retains its bibliography DOI and permanent `works/{citation_key}.html` URL. Distinct template manuscripts retain separate identities even when they share a tooling release. Generated work pages expose title, ordered authors, citation DOI, date, and the selected primary PDF in HTML citation tags, with matching work identity in JSON-LD. Person/organization author types come from verified DOI-matching registration records; a comma-free display name alone does not establish an organization. Unverified literal names remain untyped. They are linked from the static publication browse interface and included according to the sitemap policy. `build_work_pages.py` reads reviewed `citation_metadata` from paper metadata for journal title, volume, issue, ISSN, and page range. Crossref enrichment must match the curated DOI and title before adoption; Zenodo deposits are not invented as journal publications. Full citation dates must be valid and agree with the curated citation year and DOI. Otherwise year-only precision is kept. `seo_invariants.check_work_citations` verifies title, all authors in order, DOI, date, journal fields, and the actual eligible primary PDF against sources. ## Bounded same-directory PDF pilot [Google's inclusion guidance](https://scholar.google.com/intl/en/scholar/inclusion.html) requires the PDF selected by `citation_pdf_url` to share the abstract page's subdirectory. The reviewed pilot declares byte-identical, content-addressed `works/.pdf` aliases for selected journal articles in [`data/scholar-pdf-projections.json`](../../data/scholar-pdf-projections.json). The [dated source review](../../reports/scholar_pdf_placement_review_2026-11-05.md) records the five selected documents, rights evidence, searchable first-page text, ordered bylines, dates and complete author-written abstracts. The two MDPI abstracts were adopted in full from the retained publisher PDFs. The shared [projection resolver](../../code/src/scholar_pdf_projection.py) validates exact work identity, source bytes, review bindings and immutable destination reservations. The Pages assembler creates aliases only in its output; no duplicate PDFs are added to the repository. Each file must be at most 5,000,000 bytes, and all reserved aliases together must stay within a 5 MiB aggregate allowance. Retired reservations remain hosted and budgeted. Changing a selected edition never reassigns an existing alias to new bytes. Reviewed remote primaries retain priority over aliases; original `papers/` PDFs and their public URLs remain available. The work-page citation tag, download link and structured data use the same selected resource. The readiness audit binds the selected bytes and checks that the complete reviewed abstract remains visible in static HTML. Hidden, collapsed or substituted summaries fail that contract. Declared alias paths are recognized precisely by link and SEO checks, then copied and hashed by the artifact builder under the unchanged 930 MiB CI budget and 950 MiB hard ceiling. The assembled manifest records alias membership and source custody. Local HTTP route checks must serve the assembled Pages output, for example after `uv run python3 code/orchestrators/build_pages_artifact.py --output --check-size`. The reserved `works/.pdf` files are materialized in that output and do not exist in the source checkout. The [October proposal](../../reports/scholar_pdf_placement_proposal_2026-10-13.md) is historical planning evidence. The current registry and dated review own the implemented pilot. Other works can retain cross-directory archival links or reviewed remote editions; the pilot does not establish their eligibility. Google also requires scholarly content, searchable files and freely visible full text or complete author-written abstracts. Source review, generated-page agreement, deployment verification and actual indexing are separate checks. Do not compress or rewrite source PDFs merely to claim inclusion. The 2026-10-12 source pass populated citation metadata for 22 DOI/title-matched Crossref records, including 19 journal articles. Ten further requests returned HTTP 429. A subsequent bounded sequential pass recovered and DOI/title-matched all ten records, completing enrichment of the 32 Crossref-backed paper folders (including 29 journal articles). These counts describe the dated source passes, not a permanent catalog total. ## Repeatable review and local audit `uv run python3 code/orchestrators/review_citation_metadata.py --output /private/tmp/citation-review.json` fetches bounded, sequential Crossref proposals without applying metadata. Use `--citation-key ` repeatedly to narrow a pass. DOI and normalized title must both match; malformed records, title/DOI mismatches, and HTTP errors remain explicit non-adoptable results. No source identity or Scholar metrics are written. Review each proposed field against its linked registration record before adopting `citation_metadata` in the eligible paper's `metadata.json`, then regenerate. Unchanged semantic metadata preserves its earlier adoption timestamp. A fresh request timestamp in the proposal is distinct from that reviewed adoption date. Report destinations inside the repository must be under `reports/`; symlinks and hardlinks are rejected. External destinations need an existing real parent (`/private/tmp` on macOS). `uv run python3 code/orchestrators/build_scholar_readiness.py` builds the [per-work audit](../../reports/scholar_readiness.md) and [JSON dataset](../../data/scholar-readiness.json). The ordered generation plan and standard validation run its exact `--check` mode. It records real head tags, selected public PDF hashes, size, and directory placement. The audit uses a conservative decimal 5,000,000-byte interpretation of Google's 5 MB ceiling. It preserves source PDFs and permanent URLs. Missing local PDFs are observations about this archive, not claims that no external full text exists. Citation evidence excludes template contents and later body head lookalikes; blank required fields, missing catalogued DOI tags, and duplicate singleton tags remain local blockers. Repeated ordered author tags are valid. This lightweight parser is not a complete HTML5 tree builder; the exact source identity gate and browser QA remain separate checks. The configured pilot records reviewed searchable text, scholarly suitability and complete author-written abstracts as prerequisites. Those observations do not verify every catalog item or actual Scholar indexing. The catalog includes software, courses, books, and other material; a local blocker count is not an estimate of how many scholarly articles Google should index. The dataset is discoverable through the website data catalog and agent schema registry. It does not replace authenticated Search Console or Scholar review. ## Verification and follow-up Run `uv run python3 code/orchestrators/regenerate_all.py --validate` and the repository tests after citation-source changes. If a tracked public-source snapshot was refreshed before payload landing, use the explicit narrow `--allow-dirty-prepayload-evidence` mode described in [settle.md](../operations/settle.md); ordinary clean and release checks stay strict. Run `uv run python3 code/orchestrators/gsc_followup_preflight.py --check` for local SEO and live hub readiness. A preflight success does not verify that unpublished PR content is live, that Google indexed a URL, or that Scholar merged versions. After the changes land and deploy, verify the new work pages and PDFs directly, then use the signed-in [Search Console runbook](gsc-followup.md) to submit the sitemap and inspect selected work URLs. Scholar profile membership and indexing must be checked separately. The [2026-11-05 signed-in Search Console observation](../../reports/gsc_observations_2026-11-05.md) records successful sitemap discovery, an initial indexing-request service failure, and a later live-test attempt with an unconfirmed outcome. No accepted indexing request was observed. Actual indexing and Scholar review remain separate checks. Keep `data/scholar-snapshot.json` and its receipt unchanged until a direct authenticated observation is recorded through the declared tool.