Use the TRACE method, a worked example, and a downloadable recovery worksheet to move from a broken change to a verified new baseline without treating every checkpoint as a complete backup.
AI website version history is most useful before you click Restore. It can help you identify a known-good state, inspect what changed, and recover from a bad edit. It can also remove useful later work if you select the wrong checkpoint or assume the history covers more than it does.
For an agency, the goal is not simply to make the website look old again. The goal is to recover the affected layer, preserve current data and valid later work, verify the business journey, and create a new baseline everyone can identify.
I use a five-step method called TRACE: Triage the symptom, Record the current state, Assess the checkpoint, Choose the smallest recovery, and Establish a verified baseline. The process works whether the change came from an AI request, direct code edit, visual editor, CMS update, or release, but the recovery mechanism may be different for each.
The recovery rule: Do not restore a website because an older preview looks right. First prove which layer is wrong, what the selected version contains, what it excludes, and which useful changes would be replaced.
AI website version history is not one thing
Website tools use similar words for different objects. An undo stack may last only for the current editing session. A file checkpoint may preserve code and assets but not a database. A CMS backup may preserve records but not DNS. A deployment rollback may change what visitors see without changing the working draft.
| Recovery object | Usually helps with | Do not assume it includes |
|---|---|---|
| Undo or redo | A recent edit in one editor or session | Earlier sessions, external actions, or every content type |
| Workspace checkpoint | Files, code, styles, and configuration inside its recorded scope | Current database rows, inboxes, CRM records, DNS, or payments |
| CMS or database backup | Content records or application data | Page code, theme files, live deployment, or third-party systems |
| Deployment rollback | The version currently served to visitors | The working draft, newer records, or an explanation of the defect |
| External-system recovery | Records in a CRM, email platform, payment tool, or booking service | The website workspace unless the product explicitly says so |
Current product documentation shows why the distinction matters. Webflow documents specific restore effects for CMS items, inventory, reCAPTCHA, bot protection, comments, and locales. Wix developer documentation says pages, code, schemas, and permissions revert, while Sandbox and Live database data does not. Squarespace describes undo and redo as editor-specific and session-bounded, with several excluded changes.
These are not arguments for or against those products. They are evidence that the phrase “restore the website” is incomplete. Read the scope for the actual platform and incident.
TRACE step 1: Triage the symptom and stop new edits
Start with the visitor or operator symptom, not the suspected cause. “The AI broke the site” is not reproducible. “The mobile menu opens offscreen on the production homepage after yesterday's navigation request” gives you a route, state, viewport, environment, and time boundary.
Record:
- the exact route and whether the problem is in a draft, preview, or live site;
- the visitor or account state;
- the device or viewport;
- the steps that produce the problem;
- the expected and observed result;
- the business impact and a safe temporary workaround;
- who can decide on recovery and who can release it.
Pause ordinary edits while you investigate. A moving target weakens every comparison. If the live issue is severe, an authorized release owner may need to use a known deployment rollback or a temporary containment step first. That urgent action should still be recorded. It is not a substitute for finding the affected layer.
Also preserve evidence. Save a non-sensitive screenshot, error time, route, candidate identifier, and relevant bounded logs. Do not put credentials, private customer records, tokens, or payment data into the incident record.
TRACE step 2: Record the current state by layer
Before recovering anything, identify the layers that may have changed. I use a simple map: page surface, workspace, application data, connected systems, and release.
| Layer | Examples | Evidence to inspect |
|---|---|---|
| Page surface | Copy, images, layout, focus, forms, metadata | Routes, screenshots, browser checks, approved source |
| Workspace | Code, styles, configuration, local assets | Change summary, files, tests, checkpoint comparison |
| Application data | CMS entries, users, leads, orders, uploaded files | Current records, data history, separate backups |
| Connected systems | CRM, inbox, analytics, payments, booking, DNS | Provider records, delivery tests, ownership records |
| Release | Draft, staging candidate, production deployment | Version ID, hostname, release time, current response |
A visual defect may come from current code while the underlying data is correct. A form may look correct while submissions no longer reach the client's destination. A restored schema may hide retained data rather than delete it. This is why screenshots alone are not enough.
Preserve the current state before a destructive action when the platform allows it. Note useful changes made after the last known-good point, even if they appear unrelated. A whole-workspace restore can replace all of them together.
Fork's history and snapshots documentation says file-changing work is checkpointed automatically and that a restore replaces current workspace state. Its spaces documentation describes recovery through automatic checkpoints or a selected snapshot inside the persistent space. That is workspace recovery. It is not a promise that another provider's records will travel backward with the files.
TRACE step 3: Assess the checkpoint before restoring
A checkpoint label is a clue, not proof. Confirm the exact identifier, creation time, related request or editor, and retention status. Then compare the checkpoint with the current state.
Fork can provide a read-only preview for a checkpoint at a commit hostname. The public documentation says Fork boots a disposable copy, strips credentials and cookies, and permits only safe read requests. That is a valuable review surface because you can inspect layout and content without first replacing the current workspace.
Its limits are equally useful. A read-only preview cannot prove a secret-dependent integration, write to a CRM, complete a payment, authenticate as every application role, or show that production DNS points to the intended release. Treat it as evidence for the states it can safely render.
Run the same bounded observations against current and checkpoint states:
- Does the original symptom appear?
- Is the intended content present?
- Which adjacent elements differ?
- Do phone and desktop layouts behave differently?
- Are headings, labels, links, focus, and error states intact?
- Which later changes would disappear after restore?
If the platform offers a change list, inspect it. If it does not, create your own short comparison. Do not rely on memory or an AI summary that says the older version is “clean.”
TRACE step 4: Choose the smallest recovery action
Recovery is a decision tree, not one Restore button.
| If this is true | Prefer | Reason |
|---|---|---|
| One label, link, style, or component is wrong | A bounded edit or undo | It preserves unrelated current work |
| One identifiable code change caused the problem | Revert that change | The history remains understandable |
| Many workspace files are wrong and no valuable later work is needed | Restore the verified checkpoint | The checkpoint matches the affected scope |
| Later workspace work or current data must survive | Repair forward | A broad restore would discard valid state |
| The defect is in CMS or database records | Use the data system's own history or backup | A file restore may not touch the records |
| Visitors face a severe release defect | Authorized deployment rollback, then diagnose | Contain production impact without confusing it with full recovery |
A revert and a restore are not always the same. GitHub's documentation explains that reverting a commit creates a new commit that removes selected changes while the original remains in history. That forward-record pattern is useful when one bounded change is responsible. A snapshot restore may instead replace the current workspace with an earlier state.
Before applying the decision, name stop conditions. Stop if the selected checkpoint is ambiguous, the affected data layer has no verified backup, later work has not been preserved, an external transaction would be repeated, or the person acting does not have release authority.
Instructions and permissions are different. Fork's roles documentation separates abilities such as edit, approve, and deploy. A recovery note that says “do not publish” guides the operator, but the actual role determines whether publishing is technically allowed.
Worked example: preserve the leads, repair the menu
This example is illustrative, not a customer result. Assume a fictional six-page agency client site. An AI request changed the header and added a new Services link. The next morning, the mobile menu is partly outside the viewport. Three valid contact submissions arrived after the header change. The site uses a database for leads and a separate CRM connection.
The agency applies TRACE:
- Triage: The issue reproduces on the live homepage at a narrow phone width when the menu opens. Desktop navigation and the contact form appear normal. New edits pause.
- Record: The agency saves the current workspace identifier, the navigation request, screenshots, and the three non-sensitive lead record IDs. It marks workspace layout as affected and database plus CRM state as currently valid.
- Assess: A read-only preview of the checkpoint before the header request shows the old mobile menu inside the viewport. It does not include the new Services link. The checkpoint proves the earlier layout, but it cannot prove form delivery or safely test writes.
- Choose: The agency does not restore the whole workspace. It repairs the menu forward so the new approved link and current lead data remain. The database and CRM are not rolled back.
- Establish: The new candidate keeps the Services link, places the panel inside the phone viewport, preserves desktop behavior, and becomes the recorded baseline after verification and authorized release.
The lesson is not that repair forward is always better. It is that the recovery matched the affected layer. Restoring every layer would have solved a broader problem than the one observed.
TRACE step 5: Establish and verify a new baseline
Recovery is incomplete until the repaired state is identifiable and tested. Reproduce the original symptom first. Then check the connected journey and the layers that were supposed to remain current.
A practical recovery pass can include:
- Surface: intended words, images, links, metadata, layout, and responsive states;
- Interaction: navigation, keyboard path, visible focus, forms, validation, and errors;
- Data: current CMS or database records that should survive;
- Destination: an authorized labeled test reaching the correct inbox, CRM, booking tool, or other endpoint;
- Access: allowed and restricted actions with the intended accounts;
- Release: production hostname, critical routes, and the version actually serving visitors.
The W3C WAI Easy Checks offer a useful starting menu for titles, image alternatives, headings, contrast, text resize, keyboard access, focus, forms, labels, and errors. W3C also says these checks are preliminary, not a definitive accessibility evaluation. Record what you checked without turning a limited pass into a conformance claim.
Give the recovered state a new identifier. Record who approved and released it, what was verified, what remained outside recovery scope, and any follow-up item with an owner and date. This closes the gap between “we got it back” and “we know what is current.”
Five recovery mistakes to avoid
- Restoring by timestamp alone. A time can identify a candidate, but it does not explain the scope or later work.
- Calling a rendered preview a full test. Appearance does not prove writes, delivery, permissions, payments, analytics, or production routing.
- Treating files as data backup. Confirm CMS, database, uploads, and external systems separately.
- Fixing the symptom while edits continue. Freeze the investigation long enough to make comparisons meaningful.
- Leaving no new baseline. Without a candidate ID and verification record, the next person has to reconstruct the same incident again.
Download the AI website change recovery worksheet
I turned TRACE into a downloadable AI website change recovery worksheet. It includes an incident record, layer map, checkpoint comparison, recovery decision table, verification plan, completion record, and illustrative completion note.
Pair it with the website change request template before implementation and the client website launch checklist when the repaired candidate needs a full release decision. The three records answer different questions: what should change, how should a suspect change be recovered, and whether a candidate is ready to launch.
Make recovery part of the agency service
Version history is not a magic undo button. It is evidence. Used well, it shows which state existed, helps bound a repair, and gives the agency a recovery option that does not depend on somebody remembering what the page looked like.
The professional move is to make the limits visible. Identify the layer. Preserve current value. Compare before replacing. Recover the smallest safe scope. Verify the business journey. Then record the new baseline.
Fork is built around persistent website spaces, agent work, automatic checkpoints, read-only review surfaces, and explicit recovery. If that operating model fits the client service you want to offer, join the Fork waitlist while workspaces continue opening in batches.
Written by Chris Peterson. Sources checked 2026-11-06.
