Websites built to be read by people and by machines

This is a specification rather than a philosophy, so it is worth writing as one. One indexable URL for every thing worth finding — every property, every episode, every program, every service somebody might search for on its own. Exactly one heading that says what the page is. A title and a description written for that page rather than inherited from the last one. A canonical address. A share image. Structured data describing what is actually visible. A sitemap whose dates mean something. A robots file that names the AI crawlers explicitly rather than leaving it to inference. And a machine-readable summary at llms.txt.

None of that is exotic and none of it is expensive at build time. All of it is expensive to retrofit, which is the entire argument for doing it while the site is being made rather than as a project the following year.

Any page type that repeats gets generated from a template and a data file rather than made by hand. That is not a preference about tooling; it is what makes the fiftieth page as correct as the first. A hand-made set of forty pages has forty opportunities to forget the canonical, and it will take several of them.

The corollary is a rule we hold to strictly: generated output is never edited in place. If a generated page is wrong, the template or the data is wrong, and fixing it in the page guarantees the next build undoes the fix. We also run a per-page audit script over the result that checks the things a human eye slides over — structured data present and valid, breadcrumbs, description length, an address where there should be one, heading count, the analytics tag actually on the page — and logs what it finds rather than asserting the build was fine.

The largest image on a page is usually the thing the visitor waits for, and it is routinely the thing nobody measured. So every output in our pipeline has a byte budget, and exceeding it fails the build rather than quietly making the site slower. The same check refuses to upscale: an output wider than the source it was cropped from is a soft image, and it is easier to catch it at build time than to notice it on a phone six months later.

Two smaller disciplines travel with it. Metadata comes off on the way out, because photographs carry more than the picture. And filenames are versioned, because an image CDN caches each width for a long time and will happily keep serving the old version of a file you replaced.

Before anything moves, we inventory every URL that currently ranks or has links pointing at it, and each one gets a redirect to its replacement. This is ordinary discipline and it is skipped constantly — we have inventoried a platform migration where every one of forty-two article addresses returned an error afterwards, and found pages still ranking on the first page of results while serving nothing at all.

The part that is not ordinary is knowing when a redirect is the wrong instrument. A legacy page whose content was selected by the part of the address after the hash cannot be redirected per item, because that fragment never reaches the server to be read. Redirect the base address and every working legacy link breaks at once. The right move is to leave the page reachable and mark it so it is not indexed while its links are still followed, and let the new pages take the rankings instead.

A site is not paid for sessions or time on page, and reporting those is how a website conversation stays theoretical for a year. We instrument the two or three actions the business is actually paid for — the download, the call, the enquiry, the registration, the booking — and once those are recorded, the argument about which pages matter stops being an argument.

Monitoring is part of the work rather than an upsell, for an unglamorous reason: we have had buttons stop working on a site we built and nobody noticed, because a dead button produces no error and no complaint. It produces a slightly lower number that looks like a quiet week.

Where you are on a platform we do not control, we deliver the copy, the files and the exact specifications, and we implement everything the system lets us touch. We do not promise to rebuild somebody else's templates, because that promise depends on access we may not get.

Saying which parts of a job are ours and which are advisory before the work starts is less satisfying than saying yes to all of it. It is also the difference between a scope that finishes and one that stalls halfway through with everyone annoyed.