You shipped a homepage redesign. The team agrees it looks better. But "looks better" is not a measurement, and the real question is whether a cold visitor now understands who you are for, what you claim, and how to start. A website teardown you ran before the redesign gives you a baseline. Run the same study again and the repeat review leads with what got fixed, so you can hold the new version against the old one line by line.
Here is the full workflow, from the first teardown through the comparison after launch.
Step 1: Run the first teardown before you touch anything
Open the AI assistant you already use and add Cobalt as a connector. The entry point for running studies walks through the exact setup; in Claude chat it is Settings, Connectors, Add custom connector, then paste the Cobalt MCP URL. Once it is connected, type /workflows to see every study and the plain phrase that starts each one.
Ask for the teardown in plain words: "Run a website teardown of yoursite.com". The AI renders your homepage and key pages itself, on desktop and phone, and comes back with a review: real screenshots of each page it judged, findings ranked by what they cost you, and the verdict at the top. No browser driving, no code, nothing for you to do while it runs. This is the baseline. Save it to an account so it does not go read-only on you mid-project.
The outcome: a dated review that says what the current homepage puts first, what it claims, whether it proves those claims, and where a cold visitor gets stuck. If you want a sharper read on the positioning side, the positioning audit covers the same study under its other name.
Step 2: Fix against the ranked findings, not against opinions
A teardown ranks findings by cost, so the top of the list is where a visitor is most likely to bounce or misread you. Hand those to whoever is doing the redesign. If that is a human designer, the review link reads cleanly on its own. If it is an AI coding agent, the same review is available as clean markdown the agent reads directly, screenshots included.
Work the findings in order. The point of starting from the ranked list is that your redesign targets the things the baseline said actually hurt, not just the parts someone felt like changing. If your redesign work happens inside an agent, the feedback loop that makes vibecoding ship shows how to keep the capture-fix-recheck cycle tight.
Step 3: Rerun the exact same study after launch
Once the redesign is live, go back to the same assistant and ask for the website teardown again on the same URL. Use the same phrasing. The study renders the new pages and judges them against the same questions: who is this for, what does it claim, does it prove it, how do you start.
A re-run does not just hand you a fresh report. It leads with what got fixed. The review opens with the findings from last time that no longer appear, so you see the before-and-after at the top instead of reading two reports side by side and diffing them yourself. The outcome is a review that reads like a changelog for your positioning: these three problems are gone, this one is still here, and here is a new one the redesign introduced.
Step 4: Read the delta, including what the redesign broke
Redesigns fix things and break things. A homepage that finally states who it is for might bury the pricing link, or a cleaner hero might drop the proof a cold visitor needed. The re-run surfaces both. Treat the still-open findings and the new ones as the next round of work.
This is where honesty matters more than a clean report. If the re-run shows the redesign did not move a top finding, that is the finding to escalate, not to explain away. The screenshots make it checkable: you can see exactly what the study saw on each page, on desktop and phone.
Step 5: Share the before-and-after with people who were not in the room
Every review publishes to a short link anyone can open, plus a read-only page for people who only need to read it. Send the re-run to the stakeholder who approved the redesign budget. They get the verdict at the top and the list of what improved, without you writing a status update. If they open it on a phone, it opens on a phone with no app to install.
You can also export the review as a PDF or Word doc if someone wants it in a deck. Same content, pick the format the receiver can act on.
What the rerun costs and when it goes read-only
Running and rerunning studies is free to try, with no credit card. Cobalt asks for an email when a review is ready so it can send you the link. The catch worth knowing: after 7 days a review goes read-only for you, so editing, exports, and agent access stop, though the shared link keeps working for everyone you sent it to. Because a redesign spans weeks, keeping both the baseline and the re-run editable matters. Pro is $5 a month or $50 a year, keeps every review editable with no 7-day cutoff, and reopens a read-only review instantly if your baseline already locked. The full breakdown is on the pricing page.
The move is simple: capture the baseline before you start, fix against the ranked findings, then rerun the identical study and read the top of the review. If it leads with three fixes and one regression, you know exactly what the redesign bought you and what it still owes you.