Rebuild, redesign, or patch? Five checks decide.

"We need a new website" is a symptom. The scope behind it is one of three things — patch, redesign, or rebuild — and those three carry very different prices and very different amounts of risk. The difference is not how the site looks. It is five measurable signals, and they disagree often enough that measuring all five before anyone quotes you is the cheapest hour in the project.

Almost every rebuild conversation we get pulled into opens the same way: a performance screenshot with a red number on it, and a conclusion already written. The number is real. The conclusion is usually wrong, because the score is a symptom reading, not a diagnosis.

So we ran one measurement pass across four unrelated public homepages with the same tool on the same day, and scored them on five signals. The interesting part is not the ranking: the highest-scoring site had the worst structural problem of the four, and the lowest-scoring site needed the smallest fix.

Why the score screenshot is the wrong input

The Lighthouse performance score is a weighted composite of lab metrics for a single URL, on a single device profile, at a single moment. That makes it useful for tracking a page over time and useless for choosing a course of treatment, for one structural reason: the score cannot tell you whether the slow thing belongs to you.

Two sites can both score 42. On one, the largest contentful paint is a hero image the template renders and the blocking time is first-party JavaScript that ships with the layout. On the other, the page is thin and clean and the entire deficit is fifteen marketing tags and three ad slots. The first is a template problem; the second is a tag governance problem that costs a week, not a quarter. Same number, opposite decisions — which is why the five signals below are weighted by structure. Each one answers a question the score can't: how much of this page is assembled by systems you do not control, and how much of what breaks is baked into the template.

The five signals and their thresholds

Every signal below can be measured from outside the organisation, with public tools, in about an hour for a full site. None of them requires analytics access, and none of them requires a vendor's audit.

Two of these are honest approximations rather than measurements. Template coupling stands in for "how many systems have to agree before one page changes," which is the real question; origins and requests are its visible shadow. Content-model debt only reads from view-source at the shallow end — the deeper version lives in a CMS you have to log into.

The thresholds are judgement. The direction is not: each signal has a band where the work is a repair and a band where the work is a replacement, and mixing them up turns a three-week project into a nine-month one.

Four sites, one measurement pass

Method, so you can reproduce it: Lighthouse 13.0.3, headless Chromium, mobile emulation with simulated slow-4G throttling, one run per homepage, measured 11 November 2026. Accessibility defects are failing rules multiplied by affected elements, not the score. This is a single lab run of a homepage — it is a reading, not a verdict on any of these properties, and it is exactly the kind of reading a buyer runs on their own site before calling an agency.

Read that table by row, not by column, and the decisions separate cleanly.

gov.uk — patch. Structurally the cleanest page we measured: one origin, zero third-party scripts, 84 KB of HTML, 201 KiB total transfer, 17 requests, no accessibility rule failing. And a cumulative layout shift of 0.565, more than five times the 0.1 threshold. On a page that small, a shift that large is almost never an architecture problem; it is a container without reserved space. Verdict: fix the layout, do not touch the platform.

bbc.com — rebuild at the template layer. 508 requests across 119 third-party origins, 5.98 MB of transfer, a 10.2-second mobile LCP and 10.4 seconds of blocking time, plus zero H1 elements on the homepage. A page that assembles itself from 119 external origins has no single owner for its own largest contentful paint. That is the definition of a rebuild — and note the precision: the shells that assemble these pages, not necessarily the whole property.

shopify.com — patch, despite the worst-looking score. Performance 42, mobile LCP 4.8 s, 4.8 seconds of blocking time. Also: accessibility 100, CLS of zero, one H1, heading order intact. The structure is healthy and the deficit is fifteen third-party origins doing work that is largely optional. That is a tag governance project. Anyone who quotes a rebuild off the 42 is selling the wrong thing.

smashingmagazine.com — redesign. Best score of the four at 81, and the only page where the defect pattern points at the template rather than at tags or at a single container: ten affected accessibility elements spread across three different rules, including heading order failing in three places, on a document that weighs 272 KB before images. That combination — score fine, template inconsistent — is the signature of a design system that has drifted across several generations of page. Verdict: consolidate the templates.

Where the scores and the structure disagreed

Three findings from that table are worth pulling out, because each one is a decision error that happens constantly in quoting.

The best-scoring site had the worst single defect. gov.uk's layout shift is the largest structural failure in the set, and its performance score is second-best. If you rank sites by score you rank that problem last.

The worst-scoring site needed the cheapest fix. Shopify's 42 is a tag problem wearing a rebuild costume. The signal that decided it was not the score — it was the accessibility count and the CLS, both of which said the markup is fine.

The rebuild tell is not slowness, it is ownership dispersal. BBC's numbers are bad, but the number that actually decides the verdict is 119. Once a page's critical path is assembled by a hundred external origins, no team can fix the largest contentful paint by editing a template, because the template is not where the problem lives.

There is a fourth reading, less comfortable: governance failures and architecture failures look identical from the outside. Some of BBC's 119 origins are advertising and consent infrastructure, which is removable. Strip the tags and re-measure the same URL, and the test either rescues the rebuild or proves it. Skipping that test is how organisations buy a new site and keep the old numbers.

What this test cannot see

Two of the five signals decide most of the price and are the two weakest to measure remotely: content-model debt and deploy ownership.

The public proxies are shallow. A page can have one H1, ordered headings and no inline styles and still be unmaintainable, because the debt sits in how the model forces editors to duplicate a section to change its order.

The in-account version is three questions for whoever runs the site:

1. Can a marketer publish a landing page without opening a ticket?
2. Does the same content exist twice because two templates render it differently?
3. Who owns the stylesheet?

Three "no idea" answers mean the constraint is the content model, and no amount of redesign fixes a content model. That is the case where a rebuild is genuinely correct — and it is also the case where a rebuild gets sold on a performance score that had nothing to do with it.

The decision rule

Score all five. Then apply one rule: the verdict comes from the worst structural signal, not the average.

• Two or more signals in the rebuild band → rebuild, and the score is irrelevant to the scope.
• None in the rebuild band → you are buying a patch, however bad the number looks.
• Only the vitals signal in the rebuild band → test first. Remove the removable third parties, re-measure the same URL, and only then decide. A vitals signal that improves by half when tags come off was never a template problem.

The reason to write the thresholds down before the measurement rather than after is that everyone, buyers and agencies alike, arrives with a preferred answer. A table you filled in before you knew the result is the only thing in the conversation that cannot be negotiated.

If you are at the point of getting quotes, the useful next step is not a design pitch — it is a scored reading of your own five signals, and a scoped estimate for the work the signals actually justify. For what the three options cost relative to each other, see what a website actually costs and whether it is worth it.

Two related reads if the decision has already landed on rebuild: the migration lessons from a WordPress to Astro move cover what breaks in the move itself, and the EN 301 549 shift to WCAG 2.2 is the accessibility timeline that should be folded into the scope rather than discovered after launch. If the rebuild is really about who edits the content, the headless CMS platform decision is the same five-signal discipline applied to one vendor choice.

Score it before you scope it. The conversation starts here.