You gave your AI a URL, walked away, and came back to a shareable review. Now you are staring at a verdict, a list of findings, and a stack of screenshots, trying to work out what to act on first. The format is the same whether the AI ran a pricing teardown, a website study, or a comparison, and it is the same format you get when you capture screens by hand for a product review a teammate can act on. Here is what each part is and how to use it.
What screenshots are in the review, and why those ones?
Every page the study judged gets a real screenshot, not a description of a page. For a pricing teardown that means the pricing page as a cold buyer sees it, usually desktop and the phone view, because a pricing page that reads fine on a laptop often breaks on a narrow screen. A website study captures the homepage and the key pages it checked. A comparison study captures the pages for each product it compared.
The screenshots are not decoration. A finding points at a specific place, and the screenshot is the proof. If the review says a buyer cannot tell which plan is theirs, the shot shows the exact pricing table where that happens. You can check the claim yourself in seconds instead of taking the AI's word for it.
What does "ranked by what they cost" mean on the findings?
The findings are not in the order the AI found them. They are ranked by impact: the things most likely to lose you a signup or a sale sit at the top, the smaller stuff sits lower. A pricing page where the solo price is buried outranks a slightly awkward heading. That ordering is the point. You read top to bottom and stop when you run out of time, knowing you dealt with the expensive problems first.
Each finding ties back to a screenshot, so you are never guessing which page or which element it means. This is the same discipline that makes a review a PM can read fast work: the most costly issue first, each one anchored to the spot it describes.
Why is the verdict at the top instead of the end?
The verdict leads so the one person who only reads the first screen still gets the answer. A pricing teardown's verdict tells you whether a buyer can work out what they would pay and which plan fits. A website study tells you whether a cold visitor learns who it is for and what it claims. A comparison study gives an overall lean between the two or three products, backed by a dated, sourced matrix.
You scroll down when you want the detail behind the verdict. You do not have to. That order respects the reader who has two minutes, which is most readers.
Does the cost of a finding come from a number the AI made up?
No. A pricing teardown works out the actual cost for a solo user, a team of 5, and a company of 25, then checks the fine print and the phone view against those cases. The ranking comes from what a real buyer in those shoes would hit, not a guessed severity score. The findings are concrete: this plan hides its price here, this field in the fine print changes the total there.
Can I share the review with someone who did not run it?
Yes. The review is a shareable link. A founder can read it on a phone, a copywriter can rewrite from it, a developer can open it and act on the specific findings. The same review is also available as clean markdown, which is what an AI coding agent reads when you want the fixes done rather than just read. One artifact serves a person who needs the verdict and an agent that needs the structure. That is the idea behind feedback both a person and an agent can use from a single link.
How do the agent-run studies differ from capturing screens myself?
The agent-run studies produce the review for you: you give a URL, the AI renders the pages, judges them, and hands back the finished review with nothing for you to do. The pricing, website, and comparison studies run entirely from a Claude chat with no browser, because Cobalt renders the pages itself. The usability pass and flow walkthrough need an AI that can drive a browser and stop once for your go before they start.
When you capture by hand instead, you drive the same review format: you click Capture screen, crop the still, drop numbered pins, add a comment by typing or dictating, and publish. The output lands in the same shape, a verdict and findings and screenshots on a shareable link. You can see real agent-run examples at Cobalt's published teardowns of products like Cursor and Shopify, and you can start a review of your own either way.
What can I do once I have the review?
Read the verdict, work down the findings until you run out of time, and act on the top ones. People with the link can comment on any individual finding, and if you own the review you can mark each comment resolved as you clear it, so the review doubles as a short worklist without needing a separate bug tracker. If you ship with an AI coding agent, hand it the markdown and it reads the findings and screenshots directly.
Re-run the same study after you make the changes. The new review leads with what got fixed, so you can see whether the expensive problems actually went away. On the free tier every review is editable and exportable for 7 days, then it goes read-only while the shared link keeps working; Pro at $5 a month keeps everything editable with no cutoff. Check the pricing details before you decide which you need.