Web design

Claude Code and ChatGPT Codex: what AI changes in website development

Gabriel Gervais
Gabriel Gervais

Partner · Strategy, advertising and measurement

October 21, 202613 min readLire en français →

Gabriel Gervais

Insights · Web design

Write to Gabriel Gervais

A question about this article or about your situation? Your message goes straight to Gabriel.

A company address, preferably.
Reply within 1 business day

AI now writes a large share of website code. Claude Code (Anthropic's programming tool) and ChatGPT Codex (OpenAI's coding agent) read an entire site, change it on request and check their work before submitting it. For a business, the real change is not the price of code: it is the time between a decision and the moment it goes live. A well-governed site can keep pace with the business, as long as strategy, rules and verification stay in human hands.

Key takeaways
  • In 2026, 90% of professional developers use a coding agent at work every week, according to JetBrains.
  • At Anthropic, more than 80% of the code merged in May 2026 was written by Claude, up from low single digits in early 2025.
  • WordPress, Elementor, Webflow and Shopify are all opening up to agents: changing a site by writing a sentence is becoming the norm.
  • The advantage for a business is agility: reacting, testing and launching faster, in both languages.
On this page
Definition

Coding agent

A coding agent is an AI that works directly in the code of software or a website. You give it an instruction in plain language: it reads the relevant files, makes the changes, runs the planned checks and presents the result for approval. Claude Code and ChatGPT Codex are the two most widely used examples. A chat assistant suggests code to copy, while an agent does the work inside the project itself.

90%of professional developers use a coding agent every weekJetBrains, 2026
80%of the code merged at Anthropic in May 2026 was written by ClaudeAnthropic, 2026
19.2%of Canadian businesses use AI to produce goods or deliver services, three times more than in 2024Statistics Canada, 2026

What changed in twenty months

In early 2025, AI mostly helped write snippets of code that a developer then copied into the project. Twenty months later, it works directly inside the project. Here are the milestones that matter for a business that owns a website.

DateWhat happened
February 2025Andrej Karpathy, an OpenAI co-founder, names "vibe coding": programming by describing what you want, without reading the code. Anthropic releases Claude Code as a research preview the same month.
May 2025Claude Code becomes available to the general public.
October 2025OpenAI makes ChatGPT Codex generally available.
February 2026Anthropic announces that Claude Code's run-rate revenue has passed US$2.5 billion, more than double since January.
May 2026WordPress 7.0 builds an AI client into its core: the software that runs a large share of the web can now talk to AI models. The visible features (images, titles, alt text) come through an official plugin.
June 2026OpenAI reports more than 5 million weekly ChatGPT Codex users. Anthropic publishes that Claude writes more than 80% of the code it merges.
August 2026JetBrains finds that 39% of professional developers use Claude Code at work (18% in January) and 16% use ChatGPT Codex (3% in January).
September 2026Elementor connects its sites to Claude, ChatGPT Codex and Cursor: agents can build and edit Elementor pages.
October 2026Shopify introduces Canvas: its Sidekick assistant made more than 25 million theme edits in the first half of 2026.

The shift is not limited to large tech companies. Statistics Canada found that 19.2% of Canadian businesses used AI to produce goods or deliver services in the second quarter of 2026, up from 6.1% two years earlier. In Quebec, the Institut de la statistique counted 12.7% in 2025. In 2026, most SMBs had not yet made the move: those that do are getting ahead.

What Claude Code and ChatGPT Codex actually do

A coding agent does not replace a page builder or a dashboard. It works on the files that make up the site, the way a developer would. The difference shows in how a request unfolds.

01

It reads the whole site

The agent goes through the pages, templates, styles and project rules before touching anything. It knows that a price shown on forty pages comes from a single file.

02

It makes the change

It changes the text, creates the page, adds the French or English version or fixes the form, on a working copy rather than on the live site.

03

It checks its work

It runs the checks the project provides (broken links, structured data, writing rules) and fixes what fails.

04

It submits the result

A person approves the release. Every change is dated and can be undone.

This is not vibe coding. Simon Willison, a well-known developer whose blog is widely followed on the subject, reserves the term for software built with AI without reviewing the code it writes. When the AI's work is reviewed, tested and blocked by checks, it is AI-assisted programming. For a business, the distinction is essential: the first approach suits a prototype, the second a site that serves customers.

The real gain: the time between decision and launch

For years, changing a business website followed the same path: a written request, a queue, an estimate, a technical intervention, a review, then the release. That path is reasonable for a redesign. It becomes a brake when the market moves every week.

What the business wants to doThe usual pathWith a website managed with AI
Change a price or a condition everywhere on the siteFind each page, edit each one, check that none was missedOne instruction, then a check that confirms the price is the same everywhere
Publish a page for a new serviceTemplate, writing, integration and French version in several stepsThe page, its FAQ, its structured data and its French version prepared together
Create a landing page for a campaignOften a single template for every adOne page per campaign, matched to what the customer searched for
Test a new version of a pageOne test a quarter, for lack of production timeMore frequent variants, so faster learning
Fix a legal notice (Law 25, Bill 96)One more request in the queueThe fix the same day, applied to every page concerned

Four situations show what this agility changes for your sales and leads.

React the day the market moves

A new rule, a competitor's promotion or a season that starts early: the page that responds to the situation can be live the same day. The site keeps pace with the business instead of slowing it down.

Test more often

When a page variant takes an hour to prepare rather than a week, you can test what convinces your customers more often. Production time stops being the limit: the limit becomes the traffic a test needs to be conclusive. Our article on A/B testing explains how to read one.

Keep both languages up to date

In Quebec, one language version of a site often falls months behind the other. An agent prepares both versions in the same request. The translation is reviewed like everything else, but it no longer waits in line.

Launch without waiting for a redesign

A new service, a new market or a new offer no longer has to wait for the next big website project. The page is built from existing components and follows the brand from day one.

Our own site, by the numbers

We built falia.co with Claude Code, on the Astro framework. The site is pre-built (pages reach the visitor already assembled) and its code lives in a versioned repository. Here is what the site's history shows as of October 20, 2026.

01

One month from first file to launch

Work began on August 20, 2026. The site has been live since September 22, 2026.

02

782 URLs, in French and English

That is the number of pages listed in the sitemap: service pages, articles, glossary entries and client cases. The 587 URLs of the former WordPress site lead to their new page.

03

28 automatic checks before every release

Broken links, structured data, tags, writing rules, prices, consistency between French and English: if a check fails, the release is blocked.

04

386 changes in four weeks

That is the number of changes recorded in the site's history between September 22 and October 20, 2026, including 76 on launch day.

What stays human on our side also shows in these numbers. The 28 checks translate decisions we made: how we write, which prices are shown, what the site must never say. Claude Code executes, a senior expert decides and signs off. It is the same method we apply to our clients' websites managed with AI.

Every platform is opening up to agents

Changing a site by writing a sentence is no longer unusual. In 2026, the main platforms all opened a door to agents.

PlatformWhat changed in 2026
WordPressVersion 7.0 (May) includes an AI client that lets the software talk to AI models. An official MCP adapter (MCP is the protocol that connects an AI to a tool) lets Claude and other agents act on the site.
ElementorSince September, Claude, ChatGPT Codex and Cursor can build and edit Elementor pages. Everything the agent produces is saved as a draft rather than published.
WebflowSince February, an official connector lets you ask Claude for bulk content updates and SEO audits, directly on the site.
ShopifySince October, Canvas shows Sidekick's changes in real time on the store's actual code.
v0 (Vercel)Since February, "anyone on a team, not just engineers" can ship code through a Git workflow, with a pull request for each change.

Consumer tools are heading the same way. Lovable, which builds applications from a description, reports that 80% of the people building on it are in non-technical roles.

The consequence for a business is simple: the question is no longer whether AI will change your site. The question is who governs it. A draft no one reviews, overly broad permissions or a lack of checks can undo in a minute what the site took years to build.

What AI does not do for you

The speed is real, but reliability has not kept up at the same pace. Two recent sources show it.

  • Trust. In Stack Overflow's 2026 survey, nearly half of developers trust AI when they can verify its work. Only 6.6% trust it for an important decision.
  • Time saved. In 2025, a METR study found that experienced developers were 19% slower with AI on their own projects. Its February 2026 update leans toward a time saving, evidence METR itself describes as weak.

One incident sums up the risk. In July 2025, Replit's coding agent deleted a user's production database during a period when all changes were frozen. Replit then separated development and production environments. The lesson applies to a website: the agent should never work directly on what your customers see.

Warning

A site built with AI can look perfect on screen while hiding a form that sends nothing or pages Google cannot read. What cannot be seen is verified with automatic checks and a person who signs off.

Several decisions stay human, whatever the tool:

  • Strategy. Which pages should exist, for which customers and which searches.
  • The brand voice. AI writes fast, but it writes like everyone else unless someone gives it rules.
  • Legal matters. Bill 96, Law 25 and accessibility call for judgment the agent does not have.
  • The sign-off. A person agrees to publish and answers for what goes live.

What is changing in the website market

The financial reports of the large freelance platforms already show the shift. Fiverr revised its 2026 outlook because of persistent weakness in the categories "most exposed to AI automation." Its management summed it up: AI absorbs high-volume, low-value, transactional tasks. At Upwork, management notes that lower-complexity work continues to shift toward automation while demand for more complex projects grows. Spending on AI-related work there rose more than 22% in a year.

For a business buying a website, the message is clear. Producing code takes less effort, so the value moves to what code does not do on its own: understanding your customers, organizing the site around their searches, writing in your voice, making the site readable by Google and by AI, then evolving it every week. Our article on designing a site for visitors, SEO and AI covers that side.

Implementing a site you can change with AI: six decisions

Moving to a website managed with AI is an organizational project as much as a technical one. These six decisions come before the first instruction.

01

Who writes the instructions

One person, the whole marketing team or your provider. Both approaches can coexist: you write an instruction on Tuesday and hand over a page redesign on Thursday.

02

Where the site lives and who owns it

The code should live in a versioned repository in your name. The site should depend neither on a single AI provider nor on a single host. Agents like Claude Code and ChatGPT Codex can pick up the same code.

03

The rules the AI must follow

Your brand guide becomes a set of written rules: the words to use, those to avoid, how to show a price, what the site never promises. Without these rules, AI writes correct text with no personality.

04

Automatic checks

Every important rule becomes a check that blocks the release when it is broken. That is what lets you move fast without losing quality.

05

Who approves the release

A named person approves each release. Versioning lets you go back to the previous version in one click.

06

What you measure

The time between a request and its release, the number of tests run and the leads generated by new pages. These numbers tell you whether agility pays off.

A point to watch in Quebec: if your instructions contain personal information (customer names, form data), Law 25 also applies to the AI tool that receives them. The simplest approach is to keep personal information out of the instructions: AI does not need it to change a page.

Three ways to get there

ApproachFor whomStrengthLimit
An AI website builder (Wix, Lovable, Framer)A fast launch, a prototype, a small siteA visible result in a few hours, without codeThe site stays on the platform. SEO, security and brand rules are up to you
A CMS open to agents (WordPress, Elementor, Shopify)A business that wants to keep its current toolThe team keeps the interface it knows, with AI on topPlugins and security updates still need managing
A code-based site managed with an agent (Astro, versioned repository)A business that changes its site often and wants to move fastSpeed, security, control over every change and ownership of the codeRequires written rules and checks up front

None of these approaches is right for everyone. An online store will often stay on Shopify. A WordPress site that works well can stay on WordPress and benefit from AI connectors. Our article on the WordPress business website covers its costs and maintenance.

If you like your current site, you do not have to start from scratch. It can be recreated as is (same design, same text) in an architecture you can change with AI. Your visitors see only one difference: loading speed. The migration is prepared like any other, with its redirects.

What it means for you

Five questions to ask before you start

  • How long does it take today for a decision to become a live page?
  • Which requests are waiting in your website's queue right now?
  • Who on your side would write the instructions, and who would approve releases?
  • Are your brand rules written clearly enough for an AI to follow them?
  • Do you own your site's code? Can you change providers?

AI has made code fast. The advantage goes to businesses that have decided how to govern it.

A site you change in your own words, with automatic checks 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 AI and website development

Can I build my website myself with ChatGPT or Claude?

Yes, for a simple first site or a prototype. AI builders and agents like Claude Code make it possible. The hard part starts after launch: search visibility, security, legal obligations and weekly changes. That is where governance makes the difference.

If AI writes the code, why hire an agency?

Because code was only part of the work. Strategy, page structure, messaging, SEO, visibility in AI answers and verification still need doing. AI makes execution faster: the agency brings the decisions and the sign-off.

Does a site built with AI rank well on Google and in ChatGPT?

No better and no worse than any other: Google and AI judge the page, not the tool that built it. A pre-built, fast, well-marked-up site starts with an advantage. A site generated without a page plan or structured data can, on the other hand, stay invisible.

Is it secure? Who checks the code?

Security comes from governance: a pre-built site that reduces the attack surface, automatic checks, a working copy separate from the live site and a person who approves each release.

Do I still own my website?

Insist on it. The code should live in a repository in your name and the site should run without AI, which only serves to change it. You can then switch agents, hosts or providers without rebuilding everything.

What happens if the AI makes a mistake on my site?

With a versioned site, every change can be undone and the previous version goes back online. Automatic checks also block most errors before release: a broken link, an inconsistent price, a disconnected form.

Do I have to leave WordPress?

Not necessarily. WordPress 7.0 and Elementor are opening up to AI agents. If your site works well and you rarely change it, it can stay on WordPress. If you change it often and speed or security concern you, a code-based site managed with AI is worth comparing.

What is vibe coding? Is it serious for a business?

Vibe coding means building software with AI without reviewing the code it writes. It is useful for testing an idea in a few hours. For a site that serves customers, the AI's code must be reviewed, tested and approved: that is AI-assisted programming.

Sources and references
  1. JetBrains Research, AI Coding Agents: Adoption Trends, Developer Ecosystem Survey 2026 (more than 15,000 professional developers, May to July 2026), August 2026.
  2. Anthropic, When AI builds itself, June 2026. Source of the share of code written by Claude.
  3. Anthropic, Series G funding announcement, February 12, 2026. Source of Claude Code's revenue and its public release in May 2025.
  4. OpenAI, Codex is now generally available, October 6, 2025; Codex for knowledge work, June 2, 2026.
  5. Statistics Canada, Analysis on artificial intelligence use by businesses in Canada, second quarter of 2026, June 11, 2026.
  6. Institut de la statistique du Québec, L'intelligence artificielle dans les entreprises au Québec (in French), 2025.
  7. Andrej Karpathy, post on vibe coding, X, February 2, 2025.
  8. Simon Willison, Not all AI-assisted programming is vibe coding, March 19, 2025.
  9. WordPress.org, WordPress 7.0 "Armstrong", May 2026; official MCP adapter, accessed in 2026.
  10. Elementor, Introducing the Elementor MCP, September 29, 2026.
  11. Webflow, Webflow meets Claude, February 2026.
  12. Shopify, Introducing Canvas, October 1, 2026.
  13. Vercel, Introducing the new v0, February 3, 2026.
  14. Lovable, A first look at the build economy, June 9, 2026.
  15. Stack Overflow, 2026 Developer Survey, October 2026.
  16. METR, July 2025 study and February 24, 2026 update.
  17. eWeek, Replit agent incident, July 2025.
  18. Fiverr, second quarter 2026 results, July 29, 2026.
  19. Upwork, second quarter 2026 results, August 2026.
Gabriel Gervais
Gabriel GervaisPartner · Strategy, advertising and measurement

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 →

Keep reading

Tout Web design →
01
Web design·4 min read

PHP vulnerability of October 21, 2026: all our clients' WordPress sites are up to date

→
02
Web design·17 min read

How to spot an AI-built website: the tells of Claude Code and ChatGPT Codex

→
03
Web design·21 min read

How much does a website cost in Quebec in 2026? Prices and budget

→
Other topicsStrategySEOPaid advertisingConversionAI visibilityWeb design
← All insights