SEO you can't forget to do: how we made GEO a build gate

Every team says SEO matters and then ships a page with no <title>, no canonical, and a client-only shell that renders nothing until JavaScript runs. Not out of malice — it just isn't anyone's job on the day the page ships. So in krispyai we did the only thing that reliably changes behavior: we made the omission fail bun run check. This post is the honest version of what that gate does, and the bigger thing it deliberately does not.

What a script can enforce

scripts/check-seo.ts walks every public page.tsx across all apps and refuses to pass if any of three things is missing. The rules come straight from Google's AI optimization guide:

• The page is indexable with a snippet — it must export metadata or generateMetadata. In practice that means it goes through @krispy/seo's pageMetadata(), the one door that fills canonical, Open Graph, and Twitter cards from a single config so no page hand-rolls them and drifts.
• The content is server-rendered — a public page can't be a root "use client" component. If a crawler (or an AI answer engine) gets an empty div, you don't exist. Interactivity is allowed; it just has to live in a child component.
• The site is discoverable — every app serving public pages needs robots.ts and sitemap.ts. The blog's sitemap is generated from the content directory, so a new post can't be left out of it.

Because the check is wired into bun run check, the lefthook pre-push hook, and CI, there is no path to main that skips it. The structured data (articleJsonLd + <JsonLd>) that earns rich results is built the same way — from @krispy/seo, never a hand-inlined <script> tag that goes stale.

What a script cannot enforce

Here is the part the tooling can't do for you, and pretending otherwise is where most "GEO" advice goes wrong: the lever that gets you surfaced and cited is original, first-hand, expert content with a point of view. Google says this plainly — there is no special schema, no magic file, no phrasing trick that substitutes for having something worth reading on a crawlable page.

So the guide is blunt about what not to do, and we put the same list in AGENTS.md:

• Don't chunk your writing into bite-size fragments "for the AI". Write sections and paragraphs because humans read that way — that's readability, not an AI format.
• Don't write in "AI syntax". There isn't one.
• Don't mass-produce scaled, recycled, or commodity content, or spin a page per keyword variation.
• Don't treat llms.txt as a ranking lever. Google Search ignores it — it "will neither harm nor help" your visibility. We still ship one for the non-Google tools that read it, but we're honest that it's not why anyone finds us.

The division of labor

That's the whole philosophy: automate the hygiene so it's impossible to forget, and spend the freed attention on the writing only a human with real experience can produce. The check makes sure this very post has a canonical URL, an Article JSON-LD node, an OG card, and a sitemap entry. Whether it's worth citing is on us — and on the folder structure that made adding this blog a three-line change.