SEO: one URL per page, and give the blog its engine axis - #52
Merged
Merged
Conversation
Three findings from a crawl of all 122 pages, all of them structural. **One URL per page.** `trailingSlash: 'ignore'` with `build.format: 'directory'` meant the host served `/features/` and answered `/features` with a 301 — but every link, every canonical and the RSS feed were written without the slash. So all 2,949 internal links spent a redirect, and 121 pages carried a canonical naming the redirect rather than the page it was stamped on. The sitemap, meanwhile, listed the slashed form: three signals, three answers. `pagePath()` in src/lib/site.ts is now the single place that decides, and it leaves assets alone — `/og/default.png/` is not a file. **Posts that relate to nothing.** "Keep reading" took the two newest posts and showed them to all 104, so a reader on a Postgres page was offered the same two links as a reader on a Redis page and never the other eight Postgres posts. The blog has always grouped by engine — nine PostgreSQL, eight MySQL — but the grouping lived in the filename where nothing could reach it. src/lib/posts.ts reads it from there, sourcing the id list from engines.ts rather than keeping a second copy, and matching longest-first because sqlserver, sqlite and libsql all contain `sql`. That gives each post same-engine recommendations, sequential prev/next within its engine, and a clickable chip. The declared tags stay plain text on purpose: "Engineering" is on all 104 posts and "Databases" on all but four, so a link on either leads to a page that duplicates /blog instead of narrowing it. **Archives.** /blog/engine/<id>/ for the seventeen engines with two or more posts. One post is the post, not an archive of it. Also: BlogPosting gains the image it needed for Article rich results (og:image was already computed one line away), BreadcrumbList lands on every post and archive, mainEntityOfPage names the served URL, and the sitemap carries lastmod where a date is actually known — the posts' own front matter, and each archive's newest post. Marketing pages stay undated; a lastmod refreshed on every deploy is the signal Google learns to ignore. `updatedAt` is optional and uses the preprocessor pattern, not a bare `.optional()` — tests/content.test.ts explains why. Tests 166 → 198. The three new suites assert the built output, because the helpers were never the part that broke. Not done, deliberately: aggregateRating. It would put stars in the SERP, but Google requires it to come from real reviews on the site, and there are none.
Two mechanical passes over the posts, no prose rewritten. **Slashes.** 176 body links were written `/features`, which the host answers with a 301 to `/features/`. Same fix as the templates got. **Cross-links.** 226 places named another engine in plain prose with no link — 59 posts mention PostgreSQL, 41 SQLite, 38 DuckDB. 157 of those first mentions now point at that engine's archive, which is where a reader who just read the comparison wants to go. The rules the pass applied, in case it is ever run again: - at most three per post — a body full of chips reads as stuffing and dilutes every link on the page - never a post's own engine - first mention only, never every occurrence - nothing inside code fences, inline code, headings, tables, existing links, or block quotations. A link injected into someone else's quoted sentence is an edit to their words. - `libredb` excluded. In this blog "LibreDB" almost always means LibreDB Studio, not the embedded database of the same name, so the archive would be the wrong target. tests/engine-archives.test.ts asserts each of those against the built HTML, so a rerun that gets greedy fails rather than ships.
Google renders about 60 characters of a title. 96 of 146 pages ran past it — but the writing was rarely the reason. ` — LibreDB Studio` is 17 characters charged to every page, and removing it alone drops the count to 11. The brand was crowding out the words that tell the two pages apart. So the suffix is now fitted rather than fixed: the full name where it fits, `— LibreDB` where that fits, nothing where neither does. 75 pages still show the brand, none of them at the cost of their own words, and the overrun falls 96 → 13. The thirteen are editorial headlines that are long before any suffix is added; `seoTitle` lets a page title itself for a listing without touching its H1. Descriptions had the same shape of problem from a different cause: the string is both the meta description and the lede printed on the page, so it is written to open a post, not to end at 160 characters. 23 ran over, up to 264 — losing the closing clause, which is usually the part that earns the click. The trim now happens in buildSeo, for the tag only, and stops at a sentence rather than mid-word. The lede stays whole. `seoDescription` overrides it where a post wants a different listing. Also: /security, /compare and /docker-compose declare TechArticle and named no author. Same Organization the posts use. tests/serp-budget.test.ts asserts the rules and the built pages both.
104 posts in one flat list, and the only way in was to scroll it. The per-engine archives exist now, so the listing names them: a row of engines with their post counts, above the cards. Plain links, not a client-side filter. The grouping is then crawlable, each subset has an address someone can send, and it works with JavaScript off — which a filter would not, for no gain the filter offers over seventeen links.
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
… renders `building-universal-database-provider-typescript` arrived from somewhere self-contained and kept the furniture it needed there. On this site that furniture is duplicated: - `> **Author:** Cevheri & The LibreDB Studio Engineering Team` sat two lines under the byline the template already prints from front matter, so the page named the author twice in a row. `Topic:` restated the tags; `Target Audience:` is an artefact of the original format. - A hand-written `## Table of Contents` sat beside the generated one in the sidebar — and, being an `<h2>`, appeared *inside* that sidebar as its own first entry. The page offered a table of contents of its table of contents. Both removed, along with the `---` rule that closed the header block: with the block gone it would have been left stranded against the front matter, where the parser stops treating what follows as body. That is not hypothetical — it swallowed the Introduction section on the first attempt at this edit. Only this post is affected; the other 103 carry neither pattern. Verified after the change that all thirteen headings still render and the sidebar leads with Introduction again.
…lace Two follow-ups on this branch. serp-budget's rendered() chained six .replace() calls, so each ran over what the previous produced: '&lt;' decoded to '<' and then to '<'. The count it exists for came out right either way, but CodeQL reports it as js/double-escaping at high severity and the required check stays red. Matching once and consuming the match is the only ordering that cannot double-decode. The sitemap serializer re-derived a post's engine with a bare startsWith(), which is the second copy of a mapping this file's own comment warns against, and it drops the longest-match rule that stops 'sqlite' claiming a 'sqlserver-' post. It now calls postEngine() like everything else does. Sitemap output unchanged where it was already right: engine/sqlite 2026-07-20, engine/sqlserver 2026-06-17, engine/libsql 2026-05-06, engine/libredb 2026-05-28. Gate green, 212 tests.
--radius-md is not in the vocabulary: the scale is xs/s/m/l/xl/2xl. An undefined custom property with no fallback makes the declaration invalid at computed-value time, so border-radius resolved to 0 and the card rendered with square corners while every card around it is rounded. .post__seqlink is a bordered surface card in a grid next to PostCard, so it takes --radius-l (12px), the token the design system documents as the card radius, and now matches them.
…and-engine-archives
/helper landed while this branch was open, so it was written against trailingSlash: 'ignore' and never saw these tests. Its link to /get-started spent a 301, and its TechArticle node had no author, which is the one Google treats as required on an article. Both are what the tests on this branch exist to catch; nothing else on the page changes. Author is the organisation, matching what every blog post already emits.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
A crawl of all 122 pages turned up three structural findings and two budget overruns. Four commits, split so the content diff can be read separately from the code.
What was wrong
Every internal link spent a redirect.
trailingSlash: 'ignore'withbuild.format: 'directory'means the host serves/features/and answers/featureswith a 301 — but links, canonicals and the RSS feed were all written without the slash, while the sitemap used it. Three signals, three answers. 2,949 links and 121 canonicals affected; a canonical naming a redirect is not a canonical.Posts related to nothing. "Keep reading" took the two newest posts and showed them to all 104. A reader on a Postgres page got the same two links as a reader on a Redis page, and never the other eight Postgres posts. The blog has always grouped by engine — the grouping just lived in the filename where nothing could reach it.
Titles and descriptions overran what a listing renders. 96 of 146 titles past 60 characters — but
— LibreDB Studiois 17 of them, charged to every page, and removing it alone dropped the count to 11. 23 descriptions past 160, because the same string is the lede printed on the page and is written to open a post rather than to end at a limit.What changed
BlogPosting.image/BreadcrumbListlastmodpagePath()insrc/lib/site.tsis the single place that decides a URL's shape, and it leaves assets alone.src/lib/posts.tsreads the engine from the slug, sourcing the id list fromengines.tsrather than keeping a second copy, matching longest-first becausesqlserver,sqliteandlibsqlall containsql.Judgement calls worth reviewing
/bloginstead of narrowing it.— LibreDBwhere that fits, nothing where neither does. 75 pages still carry it, none at the cost of their own words. The 13 still over 60 are editorial headlines that were long before any suffix;seoTitleis the per-post escape hatch and leaves the H1 alone.seoDescriptionoverrides.lastmodonly where a date is known — the posts' front matter, and each archive's newest post. Marketing pages stay undated: a lastmod refreshed on every deploy is the signal Google learns to ignore.libredb. In this blog "LibreDB" almost always means LibreDB Studio, not the embedded database, so the archive would be the wrong target. It also caps at three per post and never touches code, headings, tables, existing links or block quotations — a link injected into a quoted sentence is an edit to someone's words.aggregateRating. It would put stars in the SERP, but Google requires real reviews on the site and there are none.Content commit
content:touches 101 posts with two mechanical passes — 176 body links given the slash the host serves, and 157 first mentions of another engine turned into links. No prose rewritten. The rules are in the commit message and asserted intests/engine-archives.test.ts, so a rerun that gets greedy fails rather than ships.Verification
bun run gatepasses. Tests 166 → 212; the four new suites assert the built output rather than the helpers, because the helpers were never the part that broke.