An open packing crate spilling straw, a dinner plate standing half-unwrapped from tissue paper against its front
Peace of mind5 min read · a short film

How do I approve or undo changes?

Letting anything edit your website - a person or an AI - only feels safe if nothing reaches your visitors by surprise. On a bernard-hosted site, changes to your pages land on a draft first, and you see a preview before they count.

The prompt

Show me what’s waiting for my approval on my site, and explain each change in plain words before I decide.

Signed in, Get my prompt fills in your site for you. Otherwise press Copy and fill in the [bracketed] parts. First time? Connect your AI to bernard once: Use your own AI (opens in a new tab).

One thing works differently, and it is worth knowing before it surprises you: the words inside a course lesson are not held on a draft. Adding a lesson goes through approval like any other page - but editing what an existing lesson says reaches the people already enrolled the next time they open it. There is no preview step and no Approve button for it. Course content has its own route back instead: every lesson keeps a history of what it said before, so any edit can be put back (Change the same thing in every lesson of my course covers doing and undoing this across a whole course at once). Everything below is about your pages.

How approval works

Every change the AI makes goes onto a draft - a complete copy of your site with the edits applied. The AI sends you a preview link; you open it, see exactly what your visitors would see, and read the AI’s note for each change. Press Approve and it goes live. Press Discard changes and the draft is thrown away, live site untouched.

Where the button is. Whenever changes are waiting, a green bar appears in your workshop just above the message box: it counts what’s pending, names the first few changes in plain words, offers a preview, and carries the Approve button. Approve doesn’t publish on the spot - it opens the full list first, as a checklist. Everything starts ticked; untick anything you don’t want, and the button becomes Make 4 of 5 live (whatever your numbers are). Unticked changes are removed for good when the rest go live, and it says so before you press anything. If two changes are tangled together - one builds on the other - bernard will say so and suggest keeping both or asking the AI to undo the one you don’t want.

Beside Approve, in quieter type, is Discard changes - the other answer, for when you’ve read the list and don’t want any of it. It asks you to confirm before anything is thrown away. The same list, with the same two choices, sits on your site’s own page under Site history. If you can’t see the bar, that page always knows.

Who presses that button depends on which AI you’re using. Bernard in your bernard dashboard can’t publish at all: the Approve button is yours alone, signed in to your account. If you’ve connected your own AI over MCP, it can publish a draft when you ask it to, because it acts with your authority - but it’s instructed to do that only on your say-so, never on its own initiative, and every publish it makes is recorded in your history saying what changed and that it came from your assistant. See Is it safe to let an AI edit my site? for the full picture.

Checking his work before you approve

When Bernard changes how something looks, he opens the page in a browser and looks at it. Those pictures now come back to you: a small strip of thumbnails under his reply, labelled What I saw. Click one to see it full size.

It exists for one reason. Bernard sometimes finishes a visual change confident it worked when it didn’t, and a confident sentence is hard to argue with when you can’t see what he saw. With the picture beside the sentence you can, in about five seconds - so when he asks “does that match what you meant?”, you can answer honestly rather than take his word for it. The pictures are kept for a week, so an old conversation keeps its words but not its thumbnails.

Approving isn’t the end of the story, either. After a fix goes live, bernard re-checks it - and shows the result honestly on your site’s page. What gets checked and when is covered in How Bernard checks your site - and what he does about it.

When a change would do something, not just say something

Most changes are words and pictures. Some change what your site actually does - where a form sends what people type, whether a page forwards visitors somewhere else, what the page loads from another company. Those are worth knowing about before they go live, and they are invisible in a preview: the page looks identical either way.

So bernard says them out loud, above the list of changes, in the plainest terms it can:

A form on this page will send what visitors type to example.com, not to you.

Anyone opening this page will be sent to example.com instead.

Two things to know about these notes. They name the destination, because that is the part you can actually judge - you know whether you use that company. And they stay quiet on ordinary work: rewording a page, swapping a photo, changing a colour produces nothing, so a note appearing at all is a reason to slow down and read.

They are not a security check, and bernard does not claim they are. If a change came with something bernard could not read, it says so rather than pretend it looked. The question they help you answer is the one you are actually qualified to answer: did I ask for this, and is that the right destination?

How undo works

Before approval, undo is just not approving. After approval, bernard keeps a full year of version history: ask the AI to restore an earlier version (which itself arrives as a draft for your approval), or do it yourself from your dashboard under your site’s Site history, where every published version has its own Restore button. Either way the old version comes back as a draft waiting for your approval, so nothing changes for your visitors until you say so - and because there’s only ever one draft, it replaces any unapproved changes you already had waiting. “Put it back how it was last week” always works.

One last thing worth knowing if you are not the only person working on your site: there is a single drafting table per site, shared by everyone and every assistant working on it, so approving picks up whatever is on it (Two of us are working on the site at the same time).

Questions people ask

Does the AI publish changes to my website by itself?
No - not on its own initiative. Every change lands on a draft first, and it’s instructed to publish only when you ask it to in plain words. Bernard in your bernard dashboard can’t publish at all: the Approve button is yours alone. You’ll find it in a green bar just above the message box whenever changes are waiting, and on your site’s own page under Site history. Your own connected AI (ChatGPT, Claude, Gemini) can put a draft live when you say so, and every publish is recorded in your version history and can be put back.
How do I undo a change before it goes live?
Just don’t approve it. To drop one change but keep the rest, press Approve to open the list, untick the one you don’t want, and press Make the rest live - the unticked change is removed for good. To drop the whole draft, use the quieter **Discard changes** link beside Approve: press it, confirm with **Yes, discard**, and everything waiting is thrown away with your live site untouched.
How do I undo a change after I’ve approved it?
bernard keeps a full year of your site’s history. Ask the AI to put a page back how it was, or restore any published version yourself from your bernard dashboard (Sites → your site → Site history).
What if the AI made several changes - do I approve each one?
They collect on one draft with a plain-English note beside each change. The green bar shows the first few notes, and pressing Approve opens the full list as a checklist - every change starts ticked. Untick any you don’t want and the button becomes Make 4 of 5 live (whatever your numbers are); the unticked ones are removed for good when the rest go live. If a change needs correcting rather than removing, ask the AI to fix it before you approve.
How do I know a change is safe to approve?
You are never asked to judge code, because nobody could. Instead, when a change would alter what your site DOES rather than what it says, bernard puts it in plain words above the list - naming the place, not the technique. “A form on this page will send what visitors type to example.com, not to you.” “Anyone opening this page will be sent to example.com instead.” Ordinary edits - new wording, a swapped photo, a colour - say nothing at all, so a note appearing means something worth a second look. The question to answer is not “is this safe?” but “did I ask for this, and is that the right destination?”
What is the “What I saw” strip under Bernard’s reply?
When Bernard changes how something looks, he opens your page in a browser and looks at it - and those pictures now appear as small thumbnails under his reply, labelled “What I saw”. Click one to see it full size. It’s there so you can check his work against his words: if he says the grey bar now runs edge to edge and the picture shows it stopping halfway, you can say so straight away instead of finding out later.

Next