Every substantive claim on this site is anchored to a numbered source whose publisher, title, URL and check-date are printed on the same page as the claim, and the citation marker in the prose is a link that jumps to that row. That is the entire standard.
It is enforced by two components rather than by a style guide, which is why it holds on a bad week. This page is written under the same rules it describes — the contents rail on the left, the citation pills in the prose, and the numbered manifest at the foot are the mechanism, not a decoration of it.
§01The standard
The standard in one paragraph
Every substantive claim on this site is anchored to a numbered source whose publisher, title, URL and check-date are printed on the same page as the claim; the citation marker in the prose is a link that jumps to that row. That is the whole standard, and it is enforced by two components rather than by a style guide.
The two components are SourceManifest.astro and InlineCite.astro. A style guide is a document somebody has to remember; a component is a thing the page will not render without. Once the citation apparatus is a build artefact rather than a habit, the failure mode changes shape: I can still cite the wrong source, but I cannot quietly cite nothing.
What follows is a description of code, not of intentions. Where the model has a weakness — and it has at least two — the weakness is named in its own section rather than left for you to find.
§02Source record
What a source record contains
A source is a plain object with seven fields: id (the anchor key), title, publisher, an optional shortPub used as the mono tag in the list, url, date, and an optional note. Nothing is free text — the record is the data the manifest row renders from, so a missing publisher or a missing date is a build-visible gap rather than a stylistic slip.
Here is a real record, reproduced exactly as it ships in the manifest on the Shopify schema pillar. It is the record behind every statement this site makes about Google retiring FAQ rich results1:
tsA source record, verbatim from the /shopify-schema/ manifest
{id:"google-faq-policy",shortPub:"Google",publisher:"Google Search Central",title:"FAQ (FAQPage) structured data — eligibility restricted",url:"https://developers.google.com/search/docs/appearance/structured-data/faqpage",date:"2026-05-07",note:"FAQ rich result retired across general Search 2026-05-07. Eligibility now restricted to well-known authoritative government and health authorities. Rich Results Test FAQ support ends June 2026; Search Console FAQ report ends August 2026.",}
The note is the field that does the real work. It is where the specific thing the source says gets written down, so that a later reader — including me, a year on — can tell whether the claim in the prose is still what the source supports without re-reading the whole document. A record whose note just restates its own title is a record that has not been read properly.
§03Inline citation
How a claim connects to its source
A citation in the prose is written as <InlineCite n={N} sourceId="..." />. It renders a mono superscript pill whose href is #source-{sourceId}, and SourceManifest gives every row the matching id. Click the pill and you land on the row; the targeted row runs a short cite-pulse background animation so you can see which of the numbered entries actually answered.
That round trip is the point. A footnote you have to hunt for is a footnote nobody checks, and a citation nobody checks is decoration. The pill is placed immediately after the clause it backs — inside the sentence, before the punctuation — rather than parked at the end of the paragraph, so that when a paragraph makes three claims from three sources you can see which mark belongs to which clause.
Now the limitation, because it is a real one. Both n and sourceId are typed by the author. Nothing in the build cross-checks that n matches the source's position in the array, and nothing checks that the source I pointed at is the source that actually supports the sentence. A mismatched number is a visible editorial error and you can catch it in a second by clicking through — which is the argument for making the marks clickable in the first place. A citation pointing at a real but wrong source is harder to catch, and the only defence against it is the note field and me being careful. If you find one, tell me.
§04Check-date
Where the check-date comes from
The "Last checked" date in every manifest header is derived, never typed. SourceManifest computes it as sources.map((s) => s.date).sort().pop() — the newest date across the page's own source records — and renders that value in a time element. There is no field an author can set to make it say something else.
tsSourceManifest.astro line 30 — the entire freshness mechanism
Two consequences, both deliberate. First: a page cannot display a fresher check-date than its freshest source. If I want this page to say it was checked last week, the only way to get there is to actually re-open a source and re-stamp its record — which means doing the work. Second: rewriting a title, a meta description, or a schema block moves nothing. Those edits do not touch the sources array, so the header keeps reading what it read before.
That is the opposite of the usual arrangement, where a “last updated” stamp is a template variable set to the deploy date. The methodology page takes this further and separates the three distinct dates every page carries, because Published, Updated, and Last checked answer three different questions and collapsing them into one number is how freshness becomes theatre.
§05Primary sources
Primary sources over commentary
Open any manifest on this site and the domains repeat: help.shopify.com, shopify.dev, changelog.shopify.com, apps.shopify.com, developers.google.com, schema.org, llmstxt.org. Vendor documentation, vendor changelogs, and specifications — the thing itself rather than somebody's article about the thing.
The reason is practical rather than principled. Second-hand SEO commentary decays badly: it paraphrases a doc, the doc changes, and the paraphrase stays online looking authoritative. A vendor help page changes under its own URL, which means a re-check finds the change. The llms.txt specification2 is the clean case — a spec published on 3 September 2024, cited at its own publication date, with the record's note describing the file structure the spec actually mandates rather than what a blog post said about it.
The honest exception: some of the vocabulary in this field is industry-coined and no vendor publishes a spec for it. When that happens the prose says so out loud and cites the trade publication instead. The live example is the AEO definition, whose first section states plainly that no AI vendor publishes an “AEO spec” the way Google publishes ranking documentation, and then names the reference it is leaning on. Naming the weaker source class is better than dressing a trade article up as a standard.
§06Undated sources
Undated and evergreen sources
Some cited references genuinely have no publication date — Schema.org type pages and Google's live tool endpoints among them. The repo convention is to write date: "Evergreen" or a bare year rather than invent one. That convention has a real cost, and this page is standing in it.
The Schema.org Product entry used on the digital-products schema guide carries date: "Evergreen"3. Google's Rich Results Test record on the schema pillar carries a bare "2026". Neither is a lie: the type page has no dated revision a reader could check, and the test endpoint is a running tool rather than a document.
But the check-date is computed with a plain string sort, and strings do not know they are meant to be dates. "Evergreen" sorts after every ISO date, because E comes after 2 — so any page carrying one Evergreen record displays “Evergreen” as its check-date regardless of how recently its other sources were re-opened. A bare "2026" sorts before "2026-05-07" and is harmlessly ignored, which is luck rather than design.
This page is the demonstration. Its three sources were opened on the dates their records carry, the newest real date among them is 2026-05-07, and the manifest at the foot nonetheless reads Last checked Evergreen, because of the Schema.org record. The schema layer at least refuses to propagate the error: the lastReviewedISO prop is dropped when the computed value is not an ISO date, so this page emits no lastReviewed at all rather than an invalid one. Showing you a broken derived value and explaining it is the lesser evil against hand-typing a date that looks tidier.
§07Byline
One byline, no ghost authors
Every article on this site sets its author to one Person entity: the profile at /founder/. The layout emits author as a reference to that single @id on every page it renders, so there is no staff-writer pool, no agency byline, and no unattributed content anywhere in the site.
Practically, that means the accountability for a wrong claim has one name on it. Billy Reiner — the page lists the public profiles you can check the claims against, and states in its own words that every guide on the site is written by the same person.
What is deliberately absent from that page is as informative as what is on it: no install counts, no client testimonials, no awards. That constraint is written into the file header of the founder page itself and you can verify it by reading the page — there is nowhere on this site where a number about our own results appears, because no such number could be sourced to anything except us saying so.
§08Coverage
Coverage
The rule, not a frozen number: every editorial page under the six pillars ships a SourceManifest. The pages that do not are the utility routes — contact, sitemap, thank-you, and the two legal pages — which make no external factual claim to source.
At the moment this page was built, 184 routes use the editorial layout and 195 routes ship a source manifest — every editorial page plus the two commercial pages that make factual claims about what the install does. Those two figures are counted from the page files at build time rather than typed here, so they cannot go stale between deploys; a number typed into prose is a claim that quietly stops being true the first time somebody adds a page.
A page with no manifest is a page making no external claim. If you find one that makes a claim and carries no manifest, that is a bug in the same category as a broken citation link, and the corrections route below is the right place for it.
§09Corrections
Corrections
If something on this site is wrong, email [email protected] with the subject line "Correction". Include the URL and, where you can, the source that contradicts us — a correction with a primary source attached gets fixed the same day.
The mechanics of what happens after that — how a source record gets re-stamped, when the modified date is allowed to move, and what the prose says when a cited page is retired — are set out under corrections and re-checks, so that they live in one place rather than two. Everything else about reaching a human is on the contact page.