Sequencing

The ordering rule

The principle

Sort by impact divided by effort, then apply one hard constraint: the foundation group ships before anything else, regardless of what the arithmetic says.

The constraint exists because the foundation checks are prerequisites rather than improvements. Structured data on a site serving the same content on two hostnames is decoration on a broken frame: the markup is correct and the site is still telling a crawler it is two sites.

The implementation

const EFFORT_WEIGHT: Record<CheckEffort, number> = { S: 1, M: 3, L: 6 };

export function returnOnEffort(check: Check): number {
  return check.impact / EFFORT_WEIGHT[check.effort];
}

The effort weights are not linear, and that is deliberate. A large task is more than three times the cost of a small one once you include the context switching, the review and the risk that it turns out to be two large tasks.

Two weeks

The principle

With ten working days, do four things and stop.

1. Foundation. One origin, one host, one trailing-slash policy, real 404s. Half a day.
2. Metadata. Title template, unique titles and descriptions, absolute self-referencing canonicals, explicit robots directives. Two days.
3. One schema graph. Organization, WebSite, and a page node with breadcrumbs. Two days including the integrity test.
4. robots.txt as a deliberate decision. Half a day, and it is the highest-leverage half day on the list.

That is five days of work. The remaining five are for verification, which is the part everyone skips and the reason half-finished implementations are so common.

Why not the rest

Programmatic tiers and AEO are worth more in absolute terms and worth nothing on a broken base. A programmatic tier on a site with duplicate hosts multiplies the duplication by 1,689. The AEO layer describes a site that a crawler cannot correctly resolve as one entity.

A quarter

The principle

With three months, the order changes because you can now afford work that pays off later.

Weeks 1 to 2. The two week plan above.

Weeks 3 to 4. Sitemaps and feeds with differentiated lastModified. Measurement: tier segmentation, event taxonomy, the analytics proxy. Instrumenting before optimising is what makes the remaining ten weeks attributable.

Weeks 5 to 8. Content and E-E-A-T. Author registry, freshness dates, internal linking. This is the slowest-paying work on the list and the most durable.

Weeks 9 to 10. The AI layer. llms.txt, canonical answers, citation guidance, and the seed query baseline. Establish the baseline before you optimise, or you will never know what moved.

Weeks 11 to 12. Programmatic, if and only if you have data worth templating. Otherwise performance and accessibility.

Do not build a programmatic tier before you have the data

The principle

The value of a programmatic tier comes entirely from the per-entity data. Building the template first produces a doorway page generator waiting for content, and the temptation to launch it anyway is enormous once it exists.

Write one page by hand first. If the hand-written version is thin, the generated version will be thin 1,689 times, and the strip-the-entity-name test will confirm it after you have already published.

Instrument before optimising

The principle

Shipping ten changes in a week without measurement means you learn nothing about which one worked. You will then repeat the useless ones, because they are cheap and they feel productive.

The measurement work is small: tier segmentation is a URL convention, the event taxonomy is one file, the proxy is two rewrite rules. It belongs in the first month.

The quarterly review

The principle

Two chapters age fastest: bot economics and AI answer engines. Crawler user agents change, engines are added, and citation behaviour shifts. A quarterly pass on those two is the difference between a reference and a fossil.

The rest of the checklist gets an annual pass, plus a re-run of the strip-the-entity-name test on every programmatic tier, because content decays as source data goes stale.