AI and your website: how to get quality, and where you still need a human
Two years ago, AI was a single assistant whose every line you had to double-check. Today it writes, proofreads, searches the web for sources, reads your documents (price list, old website, client emails) and draws. Quality no longer depends on whether “AI can do it”, but on how you organise it. Here is how to build an AI team for your website’s text and images, why a review in a fresh chat saves time, and where you still need a human.
The rules below come from Google’s official documentation for site owners, the French Consumer Code (Code de la consommation) and the EU AI Act, as of 16 October 2026. AI tools change every few months; the methods in this article do not depend on any particular tool.
AI today is a team, not a single assistant
The most common mistake is asking one chat to do everything at once: write the text, check it and praise it. What works is giving AI different roles one after another, as in a small newsroom. Each role is a separate request with a clear task.
- Writer. Drafts the page from your facts: what you do, where, at what price, what sets you apart. Without facts, you get the same text as a thousand other websites. Ready-made prompts are below.
- Native-speaker editor. Reads the text as a native speaker would: cuts stiff, bureaucratic wording, word-for-word translations and awkward phrasing, and evens out the tone.
- Demanding client. Reads the page through the eyes of someone choosing between you and a competitor: what they do not believe, what they did not understand, what answer they need before they call.
- Fact-check with web search. Today’s chatbots search the web and give links. Ask for a source for every figure, date and legal reference, then open those links yourself: AI can still cite a page that does not say what it claims. Search is switched on with the “Search” button under the input box; some chatbots use it automatically.
- Illustrator. Draws illustrations, diagrams, backgrounds and article covers — more on this below.
The order matters: content and facts first, then language, then images. Polishing a text whose figures are still going to change means doing the work twice.
Review in a fresh chat: the step that saves time
The chat that wrote a text is poor at spotting its mistakes: it remembers what it meant to say and reads that, not what is actually written. A fresh pair of eyes with no conversation history finds more. So run the reviewing roles separately, each with its own request:
- A saved assistant — a chat whose role is set once in its settings: a Gem in Gemini or a project in Claude (both available for free), or a custom GPT in ChatGPT (creating one requires a paid plan). From then on, you simply hand it each new text. Paid plans also offer full agents that check work on their own against rules you set.
- A new chat — click “New chat” and paste in the finished text and the request. A temporary chat is even better (that is what it is called in ChatGPT and Gemini; in Claude it is incognito mode): chatbots now remember past conversations, while a temporary chat starts from a clean slate. It works on any plan, just by hand.
Native-speaker editor:
You are a native French speaker and an editor. Read the text below. Mark the phrases that sound unnatural, like a translation or like advertising, and suggest how a French person would say them. Reply in English. Text: […]
Demanding client:
You run a small business in Nantes and are choosing who to hire for this service. Read the page below. What is unclear? What do you not believe? What answer is missing before you would call? Reply in English, point by point. Text: […]
Fact-check (with search switched on):
Find every figure, date, price, deadline and legal reference in the text below. For each one, give the official source with a link and flag anything that is not confirmed. Reply in English. Text: […]
Swap in your own language, town and line of work. A prompt for the writer, built on your facts, is in “Five tasks AI can take off your hands today”.
How this works in practice. We translated “Five tasks AI can take off your hands today” into French, Ukrainian and English, and each translation was then read by a separate agent playing a native speaker. The agents found 16 fixes in the French version, 15 in the Ukrainian and 17 in the English: literal translations, phrasing that sounded too much like advertising, terms a local reader would not understand. All of this was in texts that looked finished at first glance.
Yes, it is an extra step and a few more minutes. But a mistake caught before publishing is one that you, a native speaker or — worse — a client will not have to find. Saving time is the whole point of AI, so spend a few minutes checking now rather than fixing things later.
ImportantAn AI review does not replace a real native speaker or the original text of the law. It removes most of the mistakes, so the human gets a clean text and their review becomes shorter and cheaper. How to set this up for French: see “Writing French with AI as a non-native speaker”.
Images: what AI draws well, and where the line is
Image generation is available in ChatGPT, Gemini, Mistral’s Le Chat and Canva, including on free plans. For a small business website this is a real saving: you no longer need a stock photo library for every illustration.
- What works well: article illustrations and covers, section backgrounds, abstract “mood” scenes, banner variations, removing the background from your product photos.
- Describe your style once — colours, look, what not to draw — and add that description to every request. Then the images on your site will look like one series rather than the work of ten different artists.
Text and screensAI writes text on images better than a year ago, but it still mixes up letters in long words and fills phone and website screens with plausible-looking gibberish.
What to doAdd text in an editor on top of the finished image, and draw diagrams with text in an ordinary graphics editor. That is how the diagrams on this blog are made: drawn separately, together with their text, so there are no misspelt words.
“Photos” of your workA generated salon interior or a “finished renovation” presents something that never happened as real. A client who turns up and sees something different will come back — with a review.
What to doYour work, your team, your premises: real photos only. AI can improve them by fixing the lighting or cleaning up the background. If a generated image resembles real people, places or events and could pass for authentic, the EU AI Act requires you to disclose that it was made by AI — details in our article on the AI Act.
For Google, an “AI-generated” label on an image is mandatory only for products in Google Merchant Center (the service for online shops). On an ordinary website it is not required, but Google recommends stating how the image was created, for example in its metadata. And check the alt text (the image description used by search engines and screen readers) yourself: Google asks you to verify everything that was generated, alt text included.
Where you still need a human
- Facts and the law — from the original source. Prices, timescales, guarantees, contribution rates: whatever AI writes becomes your promise to the client. Check the law on service-public.gouv.fr or with your accountant, even if AI gave you a link.
- A native speaker for key pages. Home, services, prices, contact: this is where the client decides whether to call. An AI review makes a translation good; a native speaker’s review makes it sound natural.
- Your experience. Real cases, your area, your clients, the things you argue about with colleagues — AI has none of this, and it is exactly what sets your site apart.
- The decision and the byline. You publish, under your own name: Google advises signing content with its real author and treats made-up names, AI-generated headshots and false credentials as deception.
Google’s rules in brief
Google judges quality, not who wrote the text: using AI is not a violation in itself. The violations are something else:
- Volume without value. Google’s spam policies call this “scaled content abuse”: lots of pages made to rank, not to help the reader. This also covers machine translations of other people’s content made to multiply pages; translating your own site is not a violation. A blog of forty generated “10 tips” posts will not bring you clients, and your site may drop in the rankings.
- A page for every town. Near-identical pages such as “Plumber in Nantes” and “Plumber in Rennes” that all funnel to the same page, or blocks of text listing towns, are explicitly called spam by Google. A page for each service, yes; one for each town with the same text, no. A line such as “Area covered: Nantes, Rennes, Angers” is normal for clients; spam is a long block of towns written for the search engine.
- Fake reviews. A review written by AI in a client’s name is fake. Google will remove it and may suspend the account, and in France it is a misleading commercial practice (Consumer Code, article L121-4, item 28): up to two years in prison and a €300,000 fine, which can rise to 10% of turnover (article L132-2). Using AI to reply to genuine reviews is fine — see how in the Google Business Profile guide.
Hand over to the AI team
- Page drafts from your facts
- Proofreading, tone, simplifying
- Alt text, meta descriptions, headings
- Illustrations, diagrams, backgrounds
A human has the final say
- Prices, timescales, legal rules
- Translating your main pages
- Questions and answers (FAQ)
Never
- Client reviews
- A page for every town
- Dozens of articles “for SEO”
- “Photos” of work that never happened
If your site is already full of AI text
Do not delete everything. Check prices, timescales and promises; merge near-identical pages into one; delete thin articles that nobody reads or remove them from indexing (tell search engines not to show them — in WordPress it is a tick box in Yoast). Run your main pages past two reviewers, the native speaker and the demanding client, and rewrite the introductions with your own facts and examples.
A chatbot on your site. The EU AI Act requires people to know they are talking to a machine. The obligation falls on the service provider, not on you, but check that your bot introduces itself as one — and if it does not, label the chat window yourself (details in our article on the AI Act).
Who owns text written by AI
In France, copyright protects the work of a human being. In July 2026, a report by the Higher Council for Literary and Artistic Property (CSPLA) concluded that anything created without meaningful human involvement is not protected by copyright, while “hybrid” work (human plus AI) is protected if your creative contribution makes it original. It is a report, not a law, but the practical takeaway is clear: text you have reworked and filled with your own facts is yours; a chatbot’s raw answer can be copied by anyone.
Check your website
- You have checked prices, timescales and guarantees yourself, against the original source.
- Your main pages have been read by a “fresh” AI playing a native speaker and a client — in a new chat or as a separate agent.
- Translations of key pages have been proofread by someone who knows the language.
- There are no near-identical pages for different towns or districts.
- All reviews are genuine; photos of your work, team and premises are real.
- Images carry no misspelt text and no “screens” full of gibberish.
- The text includes things only you know: your experience, your examples, your town.
Is your site failing to bring in enquiries for other reasons? Check them against seven points. And if you would like an outside view, my audit looks at both the text and how the site appears in search.
Information correct as of 16 October 2026. Sources: Google Search Central — on AI-generated content, using generative AI content, spam policies, helpful content; Google Maps content policy; Consumer Code, article L121-4; article L132-2; EU AI Act, article 50; CSPLA report (in French).
Not sure about your website copy?
Send me your website address. I’ll look at what in your copy could hurt you in search, and tell you where AI can help and where it’s better to write it yourself.
Want your whole situation looked at?
Mentoring covers the whole path: from registration to reporting and choosing your tools.