Site review

What it asks of each page

• The five-second test. After five seconds, could a stranger say what this is and who it is for?
• The promise. Does the top of the page promise a concrete outcome, or polished vagueness?
• The main button. Does it say what happens when you press it?
• Proof. Does the page show the product working, or only claim that it works?
• Trust. Would someone hand over their details here? Does anything look generated or thrown together?
• The phone. Does the page feel finished on a phone, judged from the phone screenshots?
• Difference. Is there a reason to choose this over the alternatives you named in Settings, Market?

In the same pass, Site review reads the words line by line, the way a strict editor would: vague claims, filler, jargon, headlines that do not say what the product does, and buttons that do not say what comes next. Then it scores how ready the product is to launch.

What comes back

• Findings, ranked, each quoting the words or naming the element it is about, with a fix prompt for your coding tool.
• Rewrites for the lines that matter most, with your current words beside the proposed ones.
• A launch score out of 100, weighted across nine areas such as value clarity, trust and privacy, and pricing clarity, plus any blockers that should hold a launch on their own. The weighting is done in code, and the AI scores only the areas it saw evidence for.

Each finding lands in the department that owns its kind of problem. A screen a stranger cannot follow or trust goes to IT. A claim without proof, or a line of copy, goes to Marketing. A missing way to pay or sign up goes to Finance. The launch score feeds the verdict on Home and the score history there, and the run itself, under Runs, keeps the page-by-page read with every rewrite beside its line.

How a claim gets checked

When Site review says something about behaviour, such as “this button does nothing” or “this panel cannot be closed”, it has to give steps a browser can follow. nth Labs runs those steps in a real browser before the finding is filed:

• A claim the browser confirms stays, marked as reproduced in a browser. Only a confirmed claim can be ship-blocking.
• A claim the browser disproves is dropped.
• A claim nothing could check stays on your list as unproven, and cannot be ship-blocking.

Opinions are labelled opinion in the console. They are worth acting on, but they are never ship-blocking and never become a regression check, because no browser can settle them.

Running it

On a paid plan, every press of Check now includes Site review, and so do Full check and IT’s and Marketing’s checks. To run it alone, open New run, then Run one check instead, and pick Site review. Two options on the run screen are worth knowing:

• Focus on one section judges one part of each page, such as Pricing or How it works.
• Signed in reviews your pages as a signed-in visitor, with the test account from Settings, Sign-in.

What it costs

Site review runs on nth Labs’ AI and counts against your plan’s monthly fair use. Page text and screenshots go to the AI provider that reads them. See AI and your plan and pricing.

From the command line

If you run nth Labs yourself from the repository, Site review uses the AI key set on your machine:

node cli.cjs audit                  # the full review
node cli.cjs audit --focus Pricing  # one section of each page
node cli.cjs audit --passes 3       # three reads of one capture, each finding labelled with how many reached it
node cli.cjs copy                   # the copy only, with a rewrite beside each line
node cli.cjs launch                 # the launch score and the blockers only

Each extra pass costs about as much as another review.