It looked fine. It said nothing.
Nothing on that page was broken. It was fast, responsive and accessible. It was also the statistically most likely version of itself: near-black and electric cyan, three blurred glow blobs, a frosted glass nav, an animated film grain layer, and a scrolling marquee of service names. Swap five nouns and it sells SaaS, a barbershop, or a crypto exchange. It had no idea it was about cars.
The content was worse than the visuals. Every headline used the same formula, filled in five times. A stats band invented round numbers nobody could source. The gallery was exotics only, which anyone in the trade reads as fabricated, because real shops photograph a Civic and an F-150 next to the one nice car. The marquee advertised six services directly above a heading that said we do not offer a menu of fifty things. The page argued with itself inside one scroll.
And the signature interaction, the before and after slider that the spec called the value proposition made tactile, was the same photograph twice with one side dulled in CSS. It was honestly disclosed, which was to its credit. It also meant the one thing that has to be real on a paint correction site was not.
Write the faults down, then let a script enforce them.
A design document did not prevent version one. It had a coherent one, and the build executed it faithfully. The spec had been written before any visual target existed, so it locked in the defaults and the code did as it was told. The rebuild inverted that order.
- References first, then the plate, then the spec. A single photographic reference was locked before a line of the design document was written, and the document was derived from it rather than the other way round.
- Every fault became a machine check. The audit findings turned into a grep gate that fails the build on the banned colour literals, on backdrop-filter, on blur radii large enough to be a glow, on grain, on marquee, on more than two gradients in the file, and on em dashes. A judgement that lives in a script cannot be forgotten at two in the morning.
- The builder does not mark its own homework. Sections were built one at a time and judged by a separate pass that never saw them being built: deterministic checks first, then contrast and Lighthouse, then a cheap model ruling pass or fail against written criteria. Taste stays at criterion level and never becomes a score.
- Measure, do not assume. Two tap targets on the finished page were under the 44 pixel guideline while Lighthouse scored accessibility 100. Both were found by measuring the rendered box in a browser, not by reading a report.
Bright bay, real rows, honest proof.
- Light instead of dark. A warm off-white studio, because that is what a real correction bay looks like under 5000K inspection light, and because the dark luxury default was the loudest tell in version one.
- Prices as rows on hairlines, not three icon cards. Three services with real figures and real units, and correction quoted per car because it honestly cannot be priced from a menu.
- An inspection map instead of a fake slider. Four numbered points on a panel that explain where a detailer looks and what each defect class is. It claims nothing about that car, and the caption says so.
- A process with stop conditions. Every stage states its real elapsed time and the point at which the shop stops and phones you. A marketing step list has no failure conditions. A real workflow does, and that difference is the section.
- Ordinary cars. A compact sedan, a commuter, a family crossover. Each caption says what that kind of car usually needs, and labels the image as licensed stock or as an AI illustration, whichever it is.
- Trust as terms, not a stats band. The section names what the warranty does not cover and what happens when the work is wrong. Trust on a real shop site comes from stating things that could be held against you.
The form tells you nothing was sent.
The booking form validates properly, handles a pasted phone number with brackets in it, keeps your values on error and moves focus where a screen reader will follow. Then it shows a panel that says plainly that this is a concept demo, nothing was transmitted, and no data left your browser. A demo form that says "thanks, we will be in touch" is making a claim that is not true.
The same rule runs through the imagery. Where a genuinely licensed before and after pair of one panel did not exist, we did not fake one. Two gallery images are illustrations rather than photographs, and they carry a footnote marker that links to a note saying so. The business is fictional, and that is disclosed in a fixed banner, in the footer, and in the structured data, because a search engine cannot read a banner.
A service page built to be found.
Behind the home page sits a full Ceramic Coating service page as the template the other services would follow: title and heading carrying service plus city, Service and LocalBusiness and BreadcrumbList and FAQPage structured data, and area copy that names real places in York Region rather than a spam list of forty towns nobody serves.
The FAQ markup was checked word for word against the visible copy, because structured data that does not match what a visitor sees is what earns a manual penalty. Its most useful section is the one headed what a coating does not do, which is the question people actually search and the one most service pages refuse to answer.
Fast because of decisions, not tricks.
- Lighthouse mobile on the service page: 100, 100, 100, 100. The home page runs 99 performance and 100 across accessibility, best practices and SEO.
- Largest contentful paint 1.7 seconds, cumulative layout shift zero. Every image is served as AVIF with a WebP and JPEG fallback at four or five widths, so a phone downloads a phone-sized image. The hero went from one 401KB JPEG for everybody to 12KB on a phone.
- Zero external requests. The fonts are self-hosted. The Google Fonts stylesheet was the last render blocking request on the page and removing it took first paint from 3.1 seconds to 1.1.
- WCAG AA across all 163 text elements, checked by computing contrast on the rendered page rather than trusting a spot check. One muted grey in the original palette failed at 2.91 to 1 and had to be replaced.
- 133KB for the whole service page. No framework, no build step, no dependencies.
The gate is the deliverable.
Anyone can generate a competent looking website now. That is exactly why a competent looking website is no longer evidence of anything. What is worth paying for is a process that can look at its own output, name what is wrong with it in specifics, and hold the line the second time.
For a detailing shop the practical result is a site that reads as a real business rather than a template: prices you can act on, a process with honest failure points, terms in writing, and a booking path that is never more than one scroll away. The same discipline that refused to fake a before and after is the one that keeps the rest of it credible.