A one-page site in ten minutes: what the workflow really looks like

Every AI website generator advertises a number with "minutes" in it, and every one of those numbers describes the machine, not the person typing at it. When we lined several of them up against the same brief, the generation step was never the bottleneck; it was everything a person still had to decide or fix afterward. We wanted to know what that gap looks like on one generator specifically, so we ran the whole thing with a stopwatch: not "how fast does the server respond" but "how long from opening the tab to having a real URL you'd send someone." We used reach, the CV-to-one-page-site generator, because it makes the narrowest possible claim about speed and is therefore the easiest one to hold to account.

The finding surprised us in the opposite direction from what we expected. The machine part is not just fast, it is faster than the number reach itself advertises. The slow part of a ten-minute site is not the AI. It is choosing.

The run, minute by minute

We filled in a real CV — ours, not a fictional one, because a fake CV produces fake decision-making and we wanted the real kind. Here is what each step actually cost, timed from a cold start.

Step	What happens	Time
Landing form	Name, role, CV upload, photo	~1 minute
Profile details	City, LinkedIn, GitHub, company site	~1 minute
Colour choice	6 background presets, 6 accent presets, or a free picker	2–4 minutes
Template choice	40 hero layouts, pre-filtered by role	3–5 minutes
Generation	The system builds the page	~10 seconds
First look and light edits	Reading the draft, swapping a line or two	1–3 minutes
Publish	Choosing a subdomain name, going live	~30 seconds

Add it up and the honest total for a first-time user sits at roughly ten to fifteen minutes — which is a real number, not the "under two minutes" reach quotes on its own marketing, because that figure assumes you already know which colours and which layout you want and click through both screens without lingering. Almost nobody does, on a first run. What the stopwatch makes obvious is where those extra eight or nine minutes actually go, and it is not where the pitch implies.

Ten seconds of that is the machine, and it is not close

Generation itself runs in two phases. The first takes about two seconds and needs no model call at all — the system composes a complete, structurally correct page deterministically, filling a known set of slots with the CV data it has already parsed. The second phase takes about eight seconds and is a single model call that rewrites the copy, adjusts the palette, and reorders sections for rhythm, then the page is reassembled. Two plus eight is ten seconds of actual machine time, which is faster than the "twenty seconds" reach quotes publicly — that headline figure builds in headroom for slower connections and heavier CVs, not padding for effect. Either way, it is not the bottleneck. If the second phase fails or times out, the first phase's page ships anyway, so the run never ends with nothing to show.

Compare that against the two decision screens sitting immediately before it. Forty hero layouts, pre-filtered by the role you entered, is still forty options, and people do what people do with forty options: they scroll the whole list once before committing to anything, then scroll it again to check they picked the right one. The colour step has the same shape at smaller scale — six presets each for background and accent, plus a picker that invites you to second-guess the presets entirely. Neither decision is hard in isolation. Both are the kind of small, low-stakes choice that expands to fill whatever time you give it, and forty layouts gives it a lot of time.

This is the part the "ten-minute" claims never mention, because it is not the vendor's to control. A generator can only make its own step fast. It cannot make a person decide faster, and the honest comparison — the one we set out in our framework for testing AI generators — is that time-to-first-output is nearly always the wrong number to advertise, because the interval that follows is where the real cost sits. Here, unusually, that interval is mostly deciding rather than fixing, which is a better problem to have than the rebuild tax most generators leave you with. It is still the time that matters.

What costs credits, and what doesn't

Once the page exists, the free pool is forty credits, and premium accounts get six hundred a month. Regenerating a section — hitting the rotate button to get a different hero, a different values block, a different anything — costs a flat five credits, charged before the run starts, because that is the one action that calls the model again. Adding a new section, rebuilding one from a blank state, or moving one around costs nothing and returns in a few hundred milliseconds, because none of that touches the model at all — it is composition from the same vetted layout library the deterministic first phase draws from, not generation.

The practical upshot: forty free credits is eight regenerations, which is plenty to try a few different heroes and a couple of alternate takes on the CV timeline before you settle. What it is not plenty for is treating regeneration as a slot machine — pulling the lever repeatedly hoping for something better rather than editing what you have. The editor does inline text, drag positioning, font and colour changes per element, image swaps, and an AI chat that can rebuild a specific part on request, all without spending a credit. Most of the actual shaping of the page happens there, not in the regenerate button.

The blocker nobody puts in the demo video

Here is the limitation worth stating plainly rather than softening: the editor needs at least 820 pixels of width. Below that, on a phone, you get a read-only preview and a share bar, with a PIN you can hand to a laptop or desktop to pick up editing where you left off. Filling in the form, uploading the CV, watching the page generate, and viewing the result all work fine on a phone. Editing does not, and if your plan was to build the whole thing on a train with your phone, that plan does not survive contact with the product. It is a reasonable trade for a tool whose editor supports layer ordering, per-viewport styling and multi-select — that is real interface complexity that a touch screen under 820 pixels genuinely cannot host well — but it is a trade, and the marketing page does not lead with it.

What the parser cannot do for you

A CV parser reads facts: job titles, dates, a list of skills. It cannot tell you which of three similar-sounding roles actually mattered, why you left the second one, or what you want the page to argue about you rather than just report. That gap is not specific to reach — it is the gap in what AI-built portfolios get wrong more generally, and it is structural rather than a bug anyone is going to fix: a document built from your employment history will always need a sentence or two of framing that only you can write, because framing requires a point of view and a CV does not contain one. Budget two or three minutes for that specifically, separate from the colour and template time above, because it is the one part of the process that stays manual no matter how good the generator gets.

The reason this whole exercise is worth doing with a stopwatch rather than trusting the pitch is that the two halves of the process behave completely differently. The machine half is genuinely close to instant and gets faster every time the model improves. The human half — choosing among forty layouts, six colour presets, and the handful of words that describe you honestly — does not get faster just because the tool around it did. Knowing which half you are in, at each point in the ten minutes, is the difference between using the tool well and being frustrated that it didn't feel as fast as the number on the landing page. That distinction between what a system decides deterministically and what a model decides on request is also the one we work through in more technical detail in deterministic assembly versus free-form AI generation, for anyone who wants the mechanism behind why the ten seconds holds up run after run.