Editing your website with AI: write the instruction, check, publish
Partner · Strategy, advertising and measurement
On a website managed with AI, a change starts with a sentence written in Claude or ChatGPT, in your own words: change a price, add a question to the FAQ, create a page or fix a legal notice. The AI makes the change on a working copy, then automatic checks verify what can be verified without you. A person reviews the list of affected pages and approves the release. If the result is not right, the previous version comes back in a few minutes. The marketing team's work shifts: fewer requests waiting in line, more time to decide and review.
- A good instruction states the change, its scope, the exception and the reason. Everything else lives in your written rules.
- The AI works on a working copy: nothing reaches the live site before a person approves it.
- Written rules guide the AI, but only automatic checks block an error before it is published.
- Every change is dated and reversible: you undo one specific change without restoring the whole site.
- Time saved can be measured by logging the delay between a request and its release, before and after switching tools.
On this page
Working copy
A working copy is a version of the site, invisible to your visitors, on which the AI makes the requested changes. It lets you see the result, run the checks and review before anything changes online. The public site only receives the change when a person approves it. A rejected change leaves the live site as it was.
Our article on what AI changes in website development explains why businesses are moving to this way of working. This one describes what happens between your sentence and the page your customers see.
What a well-written instruction looks like
An instruction is written like a message to a capable colleague who does not know the file. Anthropic, which builds Claude, suggests this test in its documentation: show your request to a colleague who lacks the context. If they would not know what to do, Claude will not either. A vague instruction gives a vague result.
A useful instruction often fits in two or three sentences. It says four things:
- The change. What should be different once the work is done, described as a result.
- The scope. One page, one type of page or the whole site, in French only or in both languages.
- The exception. What must not move: a campaign page, a text approved by your legal advisor, an archive.
- The reason. Anthropic notes that explaining the motive behind an instruction helps the AI aim right. The reason also helps the person who will review.
The rest does not need repeating in every instruction. Your tone, the exact names of your services or how a price is displayed live in your written rules (your brand guide), which the AI reads before working. Our glossary entry on the AI instruction defines it in detail. Here is the difference, with fictional examples.
| Instruction that is too vague | Instruction that works |
|---|---|
| “Update the prices.” | “The basic plan goes to $99 a month on January 1. Change it everywhere it appears, in French and English, except in the article about the history of our prices.” |
| “Add a FAQ about delivery.” | “Add three questions our customers ask on the phone to the Delivery page: delivery times to the regions, the cost of a return, Saturday delivery. The answers are in the attached document. Keep the tone of the page's other questions.” |
| “Make a page for the new service.” | “Create a page for our commercial cleaning service based on the Residential cleaning page. It is aimed at property managers. Add it to the Services menu right after Residential cleaning, then prepare the French version.” |
| “Fix the Law 25 notice.” | “In the footer and on the Contact page, the finance department becomes responsible for the protection of personal information. Write the job title, not a name. Do not touch the privacy policy: our legal advisor is reviewing it.” |
Four common requests and what they touch
A request that looks simple often touches more places than you think. The agent searches the whole site, including what the visitor does not see: that is where it saves the most time.
| The request | What the AI touches beyond the visible page | What you check |
|---|---|---|
| Change a price | The other pages that show it, the structured data read by Google and AI tools, the meta descriptions, the other language version | The number of pages affected and that the old price is gone |
| Add questions to a FAQ | The FAQ markup in the page's code, the other language version | The accuracy of the answers, word for word |
| Create a page | The menu, the sitemap, links from neighbouring pages, the title tag, the other language version | The message, the page's address and its place in the menu |
| Fix a notice | Every page that carries the notice, including the footer and the forms | The exact wording, especially if it is legal |
A page address that changes needs one more precaution: the old address must redirect to the new one. Otherwise, outside links and Google results lead to a page that cannot be found. Our article on website migration explains why.
What the AI does after the instruction
Claude Code and ChatGPT Codex (OpenAI's coding agent) follow the same steps.
It reads before it writes
The agent goes through the relevant files and the project's rules: Claude Code reads a file named CLAUDE.md at the start of each session, ChatGPT Codex reads its equivalent, AGENTS.md. That is where your way of writing lives.
It works on a working copy
The change happens away from the live site. In October 2026, the ChatGPT Codex documentation described a separate workspace for each task, with the change saved when you choose.
It lists what changes
You know how many pages move before anything is published. A list longer than expected often signals a misunderstood instruction.
It runs the checks
The agent runs the verifications set up for the project and fixes what fails before showing you the result.
It waits for approval
The change stays on the working copy until a person agrees to publish it.
An example from our own site. On September 22, 2026, we changed the starting price of one of our offers. The 20 pages that show it, in French and English, almost all read that price from a single source. The agent found the exception, a page where the price had been typed by hand: it connected that page to the source at the same time. The old amount was then added to a check that blocks publication if it ever reappears.
Checks: what gets verified without you
Well-written rules are not enough. The Claude Code documentation says so plainly: the agent treats its rules file as context, not as enforced configuration. The more specific and concise the rules, the more consistently they are followed. Nothing guarantees they will be followed every time.
We confirmed this on falia.co. On September 8, 2026, we added a punctuation rule to our written rules. It did not hold for a day: we turned it into a check that blocks publication the same day. As of October 20, 2026, our site had 28 checks of this kind.
A check verifies what has a right and a wrong answer, for example:
- No link leads to a page that no longer exists.
- Every page title keeps its keyword.
- The request forms are connected.
- The structured data is valid for Google and AI tools.
- An old price or a withdrawn promise does not come back.
A check does not judge the tone or the relevance of a message. An automatic check verifies what can be verified, then a person decides the rest.
An error that comes back twice deserves a check rather than one more sentence in your rules. A rule the AI forgets now and then will end up online.
Approval: what to review
The checks have already ruled out the mechanical errors. The person who approves focuses on four questions:
Did the right number of pages change?
If you expected five pages and the list shows forty, ask why before approving.
Are the facts accurate?
A price, a date, a delivery time or a condition is checked against the source, never from memory.
Does the text sound like you?
The AI writes correctly. What remains is to judge whether the text sounds like your business or like any other.
Was the exception respected?
The page that was not supposed to move should be absent from the list.
On our websites managed with AI, a senior expert reviews what you hand over to us, or you approve it yourself when you keep control. The choice can change from one request to the next.
Rolling back
A versioned site keeps every change, dated and described. Rolling back means undoing one specific change: the price changed yesterday can return without erasing the page created this morning. Our glossary entry on versioning explains the difference.
On October 3, 2026, we removed two trial layouts of our blog this way, each in a single operation, without touching the rest of the site.
Agents add a safety net during the work itself. In October 2026, Claude Code kept a rewind point for each instruction, up to the 100 most recent in a session. Its documentation adds that these points do not replace versioning, which keeps the site's permanent history.
Who writes the instructions
One person or the whole marketing team can write instructions: it is an organizational decision, not a technical limit. On our websites managed with AI, each person writes in their own Claude or ChatGPT subscription. The same checks apply no matter who wrote the request.
What matters is separating two roles:
- Those who write. Anyone who knows the subject: the sales manager for a FAQ, management for a legal notice.
- Those who approve. One or two named people, who answer for what is published.
One rule protects everyone: no personal information in an instruction. The AI needs neither a customer's name nor the content of a form to change a page. Our article on AI usage policies for small businesses helps put that boundary in writing.
All our websites managed with AI include training, up to two hours depending on your needs, with written documentation you keep. It shows how to word a request, what can be changed in one sentence and how to review what Claude or ChatGPT proposes before publishing.
What is still hard to ask for
These requests deserve more time, or a conversation before the instruction.
| The request | Why it is hard | What helps |
|---|---|---|
| “Make the page more modern.” | Visual taste cannot be measured. The AI guesses, then everyone judges differently. | Show an example of a page you like and say what you like about it |
| Changing a page's address | A forgotten redirect can cost the page its traffic. | Always ask for the redirect in the same instruction |
| Connecting an outside tool (CRM, booking, payment) | It takes access to another system and a real test. | Hand this work to someone who knows both systems |
| Deciding on a promise, a price or a legal text | It is a business decision, not an execution task. | Decide first, then write the instruction |
| Instructions that contradict each other | Two people asking for opposite things on the same day produce an inconsistent site. | A single person who approves |
On WordPress, the logic is the same: our guide to AI in WordPress 7 shows how to limit an agent to drafts.
Measuring the time saved
Development teams have long measured how long a change takes: the DORA research program, run by Google Cloud, defines change lead time as the time between a change being committed and its release. For a marketing team, the useful measure starts earlier, with the request.
A shared spreadsheet is enough. Log every request for a month before switching tools, then for a month after:
The delay between request and release
The date and time of the request, then those of the release. This is the number that tells you whether the site keeps pace with the business.
The number of requests waiting
A queue that gets shorter often says more than an average delay.
The share of changes rolled back
DORA tracks the same idea for developers: the share of releases that need immediate intervention. If it rises, the instructions or the checks need another look.
What the freed-up time produced
Pages published, A/B tests run and the leads they brought in. Time saved is only worth what it lets you do.
Five questions to settle before the first instruction
- Which requests are waiting in your site's queue right now?
- Who on your side will write the instructions and who will approve releases?
- Are your brand rules written clearly enough for an AI to follow them?
- Which errors do you want blocked by a check rather than left to someone's memory?
- How will you log the delay between a request and its release?
The instruction saves time. The checks and the review make it something you can rely on.
A site you change in your own words, with automatic checks, training and an approval before every release: that is our website managed with AI offer, part of our web design expertise.
Would you rather not write the instructions yourself? Our website maintenance service takes care of it, always with the same contact.
Frequently asked questions about editing a website with AI
Do I need to know how to code to edit my website with AI?
No. You describe the result you want in your own words, as you would to a colleague. The AI does the technical work on a working copy, then automatic checks and a person verify the result before it goes live.
What happens if my request is poorly worded?
The list of affected pages will not match what you expected. Nothing is published before approval: you reject the change, then you clarify the instruction.
Can the AI publish without me knowing?
Not on a properly governed site. The AI works on a working copy and the release waits for a named person's approval. Every published change stays dated in the site's history.
How do I undo a change I don't like?
With a versioned site, the change is undone and the previous version goes back online in a few minutes. The other changes stay in place.
Can several people on my team make requests?
Yes. One person or the whole team can write instructions. Name one or two people who approve, to avoid requests that contradict each other.
How long does a change take?
A change that can be described in one sentence is prepared in a few minutes. The real delay depends mostly on review and approval.
Can I use my ChatGPT subscription?
Yes, with a paid plan: free versions are generally not enough to change a site reliably. The same applies to Claude.
Does the AI also update the French version of the site?
If the instruction asks for it, the agent prepares both versions in the same change. The translation is reviewed like everything else. A check can also verify that both versions keep the same sections.
- Anthropic, guide to writing prompts for Claude (sections “Be clear and direct” and “Add context to improve performance”), accessed in 2026.
- Anthropic, How Claude remembers your project, Claude Code documentation, accessed in 2026.
- Anthropic, Checkpointing, Claude Code documentation, accessed in 2026.
- OpenAI, AGENTS.md, ChatGPT Codex documentation, accessed in 2026.
- OpenAI, Codex Cloud, ChatGPT Codex documentation, accessed in 2026.
- DORA (Google Cloud), DORA's software delivery performance metrics, updated January 5, 2026, accessed in 2026.
- Falia, falia.co repository history (September 8, September 22 and October 3, 2026) and automatic checks as of October 20, 2026.
Gabriel almost always takes your first call and carries out your audit. He builds the strategy starting from your growth goal: where to put your budget, which market to test and how to connect each lead to a real sale in your CRM. He mainly leads engagements for three goals: Optimize the profitability of your digital campaigns, Develop a new market, and Generate demand and growth. With Geneviève, he also works on organic search (SEO), AI visibility (GEO) and conversion rate optimization (CRO). The sales a Google Ads or Meta Ads campaign brings in depend on the page that receives the click. He writes mainly about marketing strategy, paid advertising and measurement.
About Falia →
Strategy
SEO
Paid advertising
Conversion
AI visibility
Web design