Anyone can point a model at a website. These are the rules it applies to every page, why each one exists, and which of them stop a publish rather than warn about it.
Every rule below was paid for by an incident on a site we run — a ranking cliff, a broken translation, a page that rendered its own source code. We are not guessing at best practice; we are writing down what already went wrong so it does not go wrong on your site too.
Rules that block a publish outright: 8 of 12. A blocked page is handed back to the employee with the reason, and it tries again. It does not publish and warn you afterwards.
Exactly one <h1>. A descriptive <title> of 15-60 characters.
Why: Title is the strongest per-page signal Google has. Google rewrites titles for the SERP most of the time, but the signal it forms from yours still affects ranking, and a generic or duplicated title reads as a quality problem.
A meta description of 140-158 characters, unique to the page.
Why: Description does not rank the page, it decides whether anyone clicks it. Too short wastes the snippet; too long truncates mid-sentence.
Every page gets body copy and FAQ answers written for that page. Never a shared template with a noun swapped in.
Why: This is the one that costs months. A site we run shipped ~1700 pages on one generic FAQ body and a site-level quality classifier suppressed the whole domain — pages still indexed, ranking gone, no manual action to appeal. Recovery ran on core-update timescale.
Every page self-canonicalises with an absolute URL. Any hreflang cluster ships alongside that canonical, also absolute.
Why: Relative hrefs are silently dropped. hreflang without a self-canonical fills the index report with "Duplicate without user-selected canonical" — 56,801 pages on one site we run.
Structured data mirrors the visible text exactly. Schema goes on the page types that earn a rich result, not on everything.
Why: Schema is not a general ranking input and not required for AI search -- Google says so directly. Markup that claims a price or a question the page never shows is a violation that can earn a manual action.
Never {# #}. Not even on one line.
Why: Django's template lexer regex has no DOTALL flag, so {# #} only closes on its own line. A multi-line one renders as visible page text, and if it contains a {% %} substring the lexer treats it as a real tag and every render crashes. Both have happened in production.
Any string that will be machine-translated uses {0}, {1} — never {name}.
Why: MT engines translate or transliterate the word inside the braces, or drop a brace entirely. A numeric slot has no word to translate and survives intact. This silently broke about 10% of non-English rows on one site before anyone noticed.
Pages ship noindex. Indexing is a deliberate opt-in by the owner, per site.
Why: These sites share a wildcard domain. One spam tenant indexed on that wildcard damages every other tenant on it, and a half-finished page indexed on day one takes months to undo.
A page either does something useful or says something specific. A page that only describes a tool is not a page.
Why: Thin wrappers around someone else's API are the clearest signal a quality classifier has. Real utility is what survives a core update.
Every indexable page has at least one inbound internal link.
Why: Internal links pass ranking signal and drive crawl. A page reachable only from the sitemap gets crawled a few times and then largely forgotten.
Every <img> has alt text and explicit width/height.
Why: Alt text is an accessibility requirement first and an indexing signal second. Missing dimensions cause layout shift, which is a Core Web Vitals failure on mobile — and mobile is what gets indexed.
A removed page returns 410 Gone, confirmed by an actual request after deploy.
Why: A 404 lingers as a soft-404 for months; a 410 deindexes cleanly. And per-view dispatch often intercepts the request before the 410-returning view is ever reached, so the check has to be a real HTTP request, not a code read.
A rule that only lives in an instruction is a suggestion — the model follows it most of the time and quietly does not the rest, and you find out three months later in a traffic report. So these run as code, after the page is written and before it goes live. A page that fails a blocking rule is handed back with the reason and rewritten. The result is on the receipt either way, so you can see which checks ran.
Every site starts noindex. Not as a default you have to discover — as a rule the linter enforces, because these sites share a domain and one bad neighbour hurts everyone on it. When your site is worth finding, you turn indexing on in settings and it starts being crawled.