In marketing copywriting, generative AI (ChatGPT, Claude, Gemini or Copilot) produces usable copy when your team gives it the material: the offer, the proof and your brand’s words. Before equipping the whole team, you also need to decide who reviews the copy, who checks the facts and how to switch tools. Marketing copywriting here means the texts that sell: offer pages, emails, sales pitches, product descriptions, social media posts. This article first presents two recent surveys, then covers what AI does well, why its copy sounds the same, how to write a request and what to settle before rolling the tool out.
What surveys say about AI in marketing copywriting
Two surveys conducted in 2025 measure the use of AI in marketing copywriting, one of them in France.
- The Content Marketing Institute and MarketingProfs publish a content marketing survey every year. The edition covering 2026, B2B Content and Marketing Trends, is based on responses from 1,015 B2B marketers, collected between June and August 2025. Almost nine in ten respondents already use AI to produce written content, so most of your competitors also write with AI. Among users, 12% say the quality of their content has fallen with AI, and 21% see no improvement on that front. The tool alone does not improve a team’s writing.
- Infopro Digital Média publishes its Baromètre des investissements et tendances marketing B2B (B2B marketing investment and trends barometer, in French) every year. The 2026 edition covers 235 B2B decision-makers surveyed in France between September and October 2025. According to the results presented by BtoB Leaders (in French), a media outlet in the same group, writing copy remains the most common use of AI in B2B marketing: 65% of respondents write copy with AI. Yet this is also the use that has fallen most, by 21 points in a year. For a marketing department that equips every writer with an AI tool, this decline means that buying the tool does not guarantee lasting use.
Before explaining the fall in quality and the decline in use measured by the two surveys, the next section describes what AI does well.
What AI does well in marketing copywriting
The table below separates four uses and shows, for each, what AI contributes and what your team must give it for the resulting copy to serve your brand.
| Use | What the tool contributes | What you need to give it |
|---|---|---|
| Analysing | Reviewing a body of existing texts and identifying the themes, the recurring vocabulary and the differences in tone from one text to another. | The texts to analyse and the precise question the analysis must answer. |
| Rewriting | Shortening, clarifying or adapting a text for another channel, for example turning an offer page into an email. | The original text, the target channel, the words you use and the words you reject. |
| Finding ideas | Suggesting topics, angles and headlines from a theme. | The intended readers, what they already know and the topics you have already covered. |
| Writing a first draft | Producing a first version of a text in a few seconds. | The description of the offer, the proof to cite and the brand’s writing rules. |
The information in the third column determines the quality of the result. Without it, AI fills the gaps with what it has encountered most often in its training data.
Why AI-written marketing copy all sounds the same
A language model, or LLM, is the program behind ChatGPT, Claude and Gemini. It picks the most probable next word, one word at a time. Without precise instructions, it therefore produces the most frequent phrases of the type of text requested: for an offer page, “innovative solution”, “tailored support”, “trusted partner”. These phrases fit every company, so they do not distinguish yours from any other. Another article on this blog, “Why AI content is bland: the cause is mathematical”, explains this mechanism in detail.
The two surveys do not say why quality is falling or why use is declining. The mechanism described above offers one possible explanation for both. Two assistants configured for a fictitious accounting firm, created for a workshop, show the mechanism at work. The first, which had received only the firm’s editorial charter, wrote “your trusted partner, supporting you with peace of mind”. The second, which had also received the description of the offer, a list of banned words and a glossary, named the possible legal structures and the criterion for choosing between them. To get copy specific to your company without giving up the tool, give it what it cannot guess: what you sell, to whom, with what proof and in what words.
Writing a one-off request that produces usable copy
A one-off request, also called a prompt, is the message you write to the AI to get a specific text. This message is where the tool receives your offer, your proof and your words. A one-off request differs from the standing instructions of an assistant configured for the whole team, known as a custom GPT. A complete request contains eight elements. Each is illustrated below with the fictitious example of a vendor that sells scheduling software to medical practices.
| Element | What you write | Example |
|---|---|---|
| Role | The function the tool must take on. | “You are the marketing copywriter for [vendor name].” |
| Context | The company, the offer concerned and the moment of the request. | “We are launching an automatic appointment reminder feature.” |
| Task | The expected text and its length. | “Write a 150-word email to practices that are already clients.” |
| Objective | What the reader must do after reading. | “The reader should activate the feature from their account.” |
| Steps | The order of the parts, when it matters. | “Start with the problem of missed appointments, then explain activation in three steps.” |
| Constraints | Banned words, the tone, what the text must not promise. | “Do not use ‘revolutionary’. Do not put a figure on the drop in no-shows.” |
| Format | The form of the answer. | “Give the email subject line, then the text, in paragraphs of three lines at most.” |
| Material | The facts the text must use, and only those. | The feature’s fact sheet and a testimonial from a practice, pasted after the request. |
Of the eight elements, the material carries the most weight: without it, the copy stays generic, even when the other seven are present. When a request has produced publishable copy, save it as a template for that type of text, so that your team reuses it instead of writing it again.
Template to copy
Copy this request template. Replace each passage in square brackets with your company’s information, then paste the material below it.
- Role: you are [the function given to the tool].
- Context: [the company, the offer concerned, the moment of the request].
- Task: write [the expected text] in [number] words.
- Objective: after reading, the reader should [the expected action].
- Steps: [the order of the parts, if it matters].
- Constraints: do not use [the banned words]; do not promise [what the text must not promise].
- Format: [the form of the answer].
- Material: use only the information below. [The offer fact sheet, the proof, a testimonial.]
What to decide before equipping a team
The previous sections cover how one writer works with AI. Equipping a whole team raises five further questions, to settle before choosing a tool.
- Can you switch tools without losing everything? Write your instructions, word lists and template requests in documents that belong to the company, outside the tool. You can then move from ChatGPT to Claude or Copilot without starting again.
- Who reviews, and how long does it take? Copy produced in a few seconds can take a long time to check. Factor review time into the gain you expect from AI. Our article on AI content supervision fatigue describes what happens when review becomes mechanical.
- Can your team still write without the tool? Keep at least one person able to write an offer page or a sales pitch without AI, for the day the tool changes its behaviour or becomes unavailable.
- What data can go into the tool? Depending on the version and settings chosen, the tool’s vendor may reuse conversations to train its models. Do not enter personal data, financial figures or internal documents into the tool, and check the terms of the version your company uses.
- Who checks the facts? A language model can invent a figure, a date or a quotation. Nothing in the generated text distinguishes an invention from an accurate fact. Appoint someone to check every verifiable claim before publication.
To turn the answers to these five questions into a rule shared across the company, see our article on the editorial charter and AI. Web writing as a whole, beyond marketing copy, is covered in “Web writing in 2026: what AI leaves to the writer”.
Frequently asked questions about AI in marketing copywriting
Can AI replace a marketing copywriter?
AI does not replace the marketing copywriter, because it does not know your company’s offer, clients or proof. The tool quickly produces first drafts, variants and rewrites from what a copywriter gives it. That copywriter then checks and corrects the result. The copywriter’s work therefore lies mainly in preparing the material and in reviewing.
How do you stop AI-written copy from all sounding the same?
Give the AI the description of your offer, your proof and the list of words you reject. Ban the generic phrases of your sector, such as “innovative solution”. Then replace your company’s name with a competitor’s: if the text is still true, the request lacked material.
Which AI should you choose for marketing copywriting?
Choose the tool your company authorises and your teams already use. The quality of the copy depends mainly on the request and on the material you provide.
WeAreTheWords’ AI editorial assistant writes only from a closed corpus: your offer, your brand voice and your writing rules. Your team therefore writes from the same offer, in the same voice. Let’s talk about your assistant.