AI Copywriting Tool Examples: Before-and-After Rewrites With Prompts

AI copy and a human rewrite differ in specifics. The raw draft is fluent but generic: vague promises, stacked adjectives and claims nobody can check. The rewrite swaps those for the owner's real numbers, wording and limits. Below are three pairs, a web page, an email and an ad, each with its prompt.

The point of showing the raw output is that you can see what to look for. Nothing here is a verdict on one tool. All three raw drafts are illustrative of what a general chatbot returns from a short prompt, and the businesses are invented: a financial adviser, a mortgage broker and a signage company.

Follow me on Instagram@sagnikteaches

How the test works and who signs what

Each piece goes through the same four steps, and the owner (or whoever is accountable for the claims) signs at the end. First, write a prompt with the audience, the offer and the facts you are happy to stand behind. Second, keep the first output untouched so you can compare. Third, edit it line by line against a short checklist. Fourth, sign off with your initials, the date and the version, in a file or a comment, so you can show later who approved what.

Connect on LinkedInSagnik Bhattacharya

The checklist I use has six questions. Can every number or promise be traced to something the business actually does? Does any sentence work for a competitor if you swap the name? Is there a detail that only this business could write? Is the next step for the reader obvious and honest? Does it break any rule of the channel or sector? Would the person quoted or described be comfortable reading it? If a sentence fails two of these, it is rewritten, not polished.

Subscribe on YouTube@sagnikteaches

The scoring sheet

To compare raw and edited drafts, score each from 0 to 2 on five lines and add up (maximum 10). The scale is mine, not an industry standard.

Rewrite one: a financial adviser's "how I work" page

The adviser runs a three-person practice and wants a page that explains the first meeting without making promises about returns. The prompt below is deliberately ordinary, the sort of thing most owners type first.

Write the "How we work" page for a small independent financial advice practice. Audience: people in their 40s and 50s who want help planning for retirement. Tone: professional but friendly. About 150 words.

The raw output, shown as illustrative, looked like this:

At our practice, we are passionate about helping you achieve your financial dreams. Our dedicated team of experienced professionals provides tailored, all-in-one solutions designed to give you peace of mind. We take the time to understand your unique goals, so you can confidently plan for a secure and prosperous retirement. Whether you are just starting out or approaching retirement, we will guide you every step of the way with expert advice you can trust. Contact us today for a free, no-obligation consultation and take the first step towards the future you deserve.

Nothing in it is false, which is exactly the problem. It scores 2 out of 10 on the sheet: no specifics, a "secure and prosperous retirement" that implies an outcome, "expert advice you can trust" with no basis, and a voice any practice could use. "Free, no-obligation consultation" is also a claim the adviser must be certain is true for every enquiry.

The second prompt gives the model real material, in the adviser's own words, and bans the weak phrases:

Rewrite the page below using ONLY the facts I give you. Do not add promises, outcomes or claims.
Facts:
- First meeting: 45 minutes, we ask for no payment and no commitment.
- Before the meeting we send a 12-question form; clients bring last year's pension statements.
- Written plan arrives within 10 working days of the second meeting.
- We review the plan once a year; we do not manage day-to-day trading.
- We cannot guarantee returns and the plan shows ranges, not promises.
Voice: plain, calm, first person plural. Avoid: passionate, dedicated, tailored, all-in-one, peace of mind, journey, dreams.
Length: about 150 words. End with one next step.

The edited version the adviser signed off, after two small changes by hand (one sentence admitting a goal might be out of reach, and a softer close):

We start with a 45-minute meeting. You pay nothing and commit to nothing. Beforehand we send a 12-question form, and we ask you to bring last year's pension statements so we are working from your real figures. After a second meeting we write your plan and send it within 10 working days. It shows ranges, because nobody can promise a return, and we will say so plainly if we think a goal looks out of reach. Once a year we sit down again and update it. We do not trade your money day to day. If that sounds like what you need, book the first meeting below.

The edited version scores 9. The human contribution was never the prose, it was the facts: the 12 questions, the 10 working days, the "we do not trade" line. The sentence admitting a goal might be out of reach was written by hand, because the model had been told to be friendly and left it out. For a regulated service, the signed-off page should still go to whoever handles compliance, because the signature on a prompt is not a regulatory approval.

Rewrite two: a mortgage broker's follow-up email

This time the task is a follow-up to a client who has gone quiet after a first call. The raw prompt:

Write a follow-up email to a client who has not replied after our first mortgage call. Keep it short and persuasive.

The raw output (illustrative):

Subject: Don't miss out on your dream home!
Hi [first name],
I hope this email finds you well! I wanted to reach out because mortgage rates can change quickly, and you wouldn't want to miss out on the perfect deal. Our team is ready to help you secure the best possible rate. Time is of the essence, so please get in touch today to avoid missing this unique opportunity!
Best regards,
[your name]

It fails on three counts. "Best possible rate" is a promise the broker cannot make. The urgency ("time is of the essence", "unique opportunity") is manufactured. And "I hope this email finds you well" says nothing. In a regulated sale, a message that pressures a client is also a risk the broker owns.

The second prompt supplies the situation and a rule for tone:

Write a 90-word follow-up email from a mortgage broker to a client who has not replied for 8 days after a first call.
What we agreed on the call: I would compare lenders once I had their payslips and a recent bank statement. They have not sent either.
Rules: no urgency, no claims about rates or "best deal", no exclamation marks. Offer two easy options: send the documents, or tell me they have gone elsewhere and I will close the file. One question at the end. Subject line under 8 words, plain.

The edited version:

Subject: Your payslips and bank statement
Hi [first name],
When we spoke last week I said I'd compare lenders once I had your latest payslips and one recent bank statement. I haven't received them yet, which is fine; things get busy. You can reply with them attached whenever suits. If you've decided to go another way, just tell me and I'll close your file, no hard feelings. Would it help if I sent a list of exactly which pages I need?
[your name]

Only one edit was made by hand: "no hard feelings" replaced a stiffer sentence about the file closing. The subject line is dull on purpose, and that is the point. A follow-up that names the two missing documents is easier to reply to than one that sells.

What the numbers looked like

This is the illustrative main example, with figures an owner might record for the email (word counts are for the body, without subject line, greeting or sign-off). Raw draft: 58 words, scored 3 of 10. Second draft, after the better prompt: 81 words, scored 8 of 10. Edited and signed: 72 words, 9 of 10, because the stiff sentence about closing the file was much longer than the three words that replaced it. Time spent: 1 minute for the first prompt, 4 minutes writing the better prompt with the facts, 3 minutes editing, 8 minutes in total. The same email written from a blank page might take 8 to 10 minutes, so the saving here is small, and it grows only when you reuse the second prompt as a template. That is an honest finding: for a single email, AI saves little, and for the tenth similar one it saves a lot.

Rewrite three: a search ad from a signage company

Ads punish vagueness harder than pages do, because the space is fixed. Say a signage company wants a responsive search ad for shopfront signs. First prompt:

Write 5 headlines and 2 descriptions for a Google Ads search ad for a signage company that makes shop signs.

The raw output (illustrative):

Headlines: Premium Shop Signs | Stand Out From the Crowd | Quality Signage Solutions | Boost Your Brand Today | Get a Free Quote Now
Descriptions: Transform your storefront with eye-catching, high-quality signs. Our expert team delivers outstanding results every time.

Responsive search ads have tight character limits: Google Ads help lists up to 30 characters for each headline and 90 for each description (as of November 2026), so check the help page if you are reading this much later. "Outstanding results every time" is the kind of claim an ad reviewer or a customer can challenge, and "Stand Out From the Crowd" would suit any trade.

The second prompt contains the facts the owner would have given a new sales assistant:

Write 8 headlines (max 30 characters each) and 3 descriptions (max 90 characters each) for a search ad aimed at shop owners who need a new fascia sign.
Facts: we make aluminium and acrylic fascia signs; most jobs take 2 to 3 weeks from approved artwork; we survey the site before quoting; we fit as well as make; quotes are free.
Do not claim "best", "cheapest" or "guaranteed". No exclamation marks. Count the characters of every line and show the count in brackets.

The signed-off set, after the owner removed one line that promised a fixed fitting day, a fact the company had not supplied:

• Fascia Signs Made and Fitted (28)
• Free Quote After a Site Survey (30)
• Aluminium and Acrylic Fascias (29)
• Most Jobs in 2 to 3 Weeks (25)
• Description: "We survey your shopfront, quote free, then make and fit your sign. Most jobs take 2 to 3 weeks from approved artwork." (117 characters, well over the 90-character box, so it was cut to "Site survey, free quote, then we make and fit your sign. Most jobs take 2 to 3 weeks." at 85)

The last bullet shows a mistake worth noticing: the first description the model wrote was too long for the box, even though the prompt asked for counts. The bracketed counts the model returned were also off by one or two on several headlines, and the figures above are recounted. Language models are unreliable at counting characters, so treat any count they report as a guess. Paste the lines into a spreadsheet and use =LEN(A2), or into the ad editor, which flags anything over the limit.

The five habits visible in every raw draft

Across the three pairs the same patterns appeared, and it is quicker to learn them than to relearn each page.

1. The warm-up sentence. "I hope this finds you well", "At our practice, we are passionate". Delete the first sentence of any raw draft and see whether anything is lost. Usually it is not.
2. Stacked adjectives. "Tailored, dedicated, experienced". A noun with a number beats three adjectives: "a 45-minute meeting".
3. Outcome promises. "Secure and prosperous retirement", "best possible rate", "outstanding results". Replace with what you do, not what happens.
4. Manufactured urgency. If the deadline is real, state it with its date. If not, drop it.
5. The invented specific. When asked for something concrete without facts, models can make up plausible ones: a delivery time, a certification, a number of years in business. This is the dangerous habit, because it looks like a specific. Check every figure against the facts you supplied.

If you see habit five in a draft, the fix is in the prompt, not the edit. Add "use only the facts below, and write [MISSING] where you need one that is not given", and the model will leave gaps you can see rather than fill them silently.

Prompt changes that moved the score most

In these three illustrative rewrites the biggest gains came from changes to the prompt, not from a better tool. Here is what each change did, in rough order of effect.

The last row deserves a longer note. Pasting one paragraph you wrote, with "match the sentence length and tone of this sample", does more for voice than adjectives like "friendly" and "warm". The adviser example above would improve again if the sample came from the owner's own client letters. For a method that builds on that idea, see how to build a brand voice guide that AI can follow.

A fourth, shorter pair: the headline

Headlines show the gap fastest, so it is a useful ten-minute exercise before you tackle a whole page. Ask for ten headlines for the adviser's page with the one-line prompt and you will usually get lines such as "Your Retirement, Simplified" and "Plan Today, Enjoy Tomorrow". They sound finished and say nothing. Ask again with the facts block and the instruction "each headline must contain a number or a concrete step", and the set changes: "A 45-minute first meeting, no charge", "Your plan in writing within 10 working days", "12 questions before we meet, so the hour is yours". Pick one, then read it aloud next to the competitor's page. If the competitor could have written it, it is not done.

This also gives you a cheap way to test your own pages. Copy the headline of three pages on your site into a document, remove the business name, and ask a colleague which business wrote each. If they cannot tell, the facts are missing from the page, and no tool will invent the right ones for you.

Running your own test in an afternoon

You do not need a tool comparison to learn what works for your business. Pick three pieces you write regularly, say the home page intro, a quote follow-up and a social post. For each, run the basic one-line prompt and save the answer. Then write the facts-led prompt and save that. Score both on the sheet above, and note minutes spent. After the three pieces you will have a table like this one, and the numbers will be yours.

If the second prompt gets you to 7 or more and the edit takes under half the time of writing from scratch, keep the routine for that kind of piece. If you are rewriting more than half the words every time, the piece may need more of your voice than a prompt can carry, and it may be better written by hand with AI used only to check it. That is a reasonable result, not a failure.

Spot the invented claim

One more quick example, because this habit costs the most. Imagine a broker's second-draft email contained "our clients typically save thousands". It had come from nowhere: no facts in the prompt mentioned savings. The fix was a one-line instruction in the prompt: "Do not mention savings, rates or typical outcomes." Search each draft for the words typically, average, most clients, guarantee and always. Each one is a claim waiting for a source.

Sign-off record to keep

A one-line record is enough, kept in the same folder as the copy. Here is a filled-in example for the adviser page:

Piece: How we work page | Version: 3 | Prompt file: how-we-work-prompt-v2
Facts checked against: engagement letter template, process sheet (checked 4 Nov)
Edits by hand: sentence about out-of-reach goals added; softer close
Approved by: [owner initials] | Date: [date] | Compliance review needed: yes, sent

It costs a minute and settles later arguments about who approved a claim. Pair it with a short fact-check routine, such as the one in how to fact-check AI marketing copy before it goes live, and with a banned list of the phrases you keep deleting, as in a banned-phrase list for your brand.

When a person should write it from scratch

Use AI for first drafts of routine copy where the facts are stable: process pages, follow-ups, ad variants, product descriptions. Write by hand when the piece turns on judgement or on a relationship: an apology, a price rise, a complaint reply, a story about a client. Models can still help there, as a reader that points out unclear sentences, but the first words should be yours.

Google's guidance on generative AI content says that using AI tools to generate many pages without adding value for users may break its spam policy on scaled content abuse, and it points site owners to its advice on creating helpful content. It is worth reading once at the Search Central page on generative AI content so you know where mass-produced, unchecked copy gets into trouble. For the causes of the flatness you have seen above, read why AI marketing copy sounds generic.

Questions owners ask before trusting AI copy

Which AI tool wrote the raw versions?

The raw versions are illustrative of what a general chatbot such as ChatGPT, Claude or Gemini typically returns from a short prompt. Different tools word things differently, but the habits shown here (generic openings, stacked adjectives, invented claims) appear in all of them. Run your own prompt through whichever tool you use and mark it up the same way.

How much editing does AI copy usually need?

Expect to rewrite or replace a good part of a first draft and to touch almost every sentence at least lightly. The figures in the worked example are illustrative. Time your own edits on three pieces before you decide whether the tool saves you time.

Should I tell customers a page was drafted with AI?

There is no single rule for ordinary marketing copy, but some platforms require labels for realistic AI-generated images or video, and regulated services may need compliance approval whatever wrote the words. Check the rules of each channel and your own sector, and make sure a named person stands behind every claim.

Can I paste client details into the prompt?

Avoid it. Marketing copy rarely needs names, account details or case facts. Describe the audience and the problem in general terms, and add real specifics yourself afterwards. If you must use client material, check your tool's data settings first.

Further reads

• AI Tools for Copywriting: A Freelancer's Draft-to-Delivery Workflow — A draft-to-delivery routine for people who write copy for clients.
• AI Software for Copywriting: A Freelancer's Monthly Tool Budget — What the tools cost month by month.
• How Agencies Keep Brand Voice Consistent in AI-Written Content — How agencies keep one voice across many writers and drafts.
• Best AI Model for Copywriting: How to Test ChatGPT, Claude and Gemini on One Brief — Run the same brief through three assistants and compare.
• Best AI for Copywriting Ads: Variants, Testing and Client Sign-Off — Go further on ad variants and testing.
• How to Turn Customer Reviews Into Marketing Copy With AI — Source real customer wording instead of inventing it.
• Using AI in Copywriting: Protect Each Client's Voice Step by Step — Build client-specific voice guides, draft from evidence and edit every line before it reaches the client.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Google Search Central guidance on using generative AI content; Google Ads Help on responsive search ads (character limits); YouTube Help on disclosing altered or synthetic content. All sample copy and figures are illustrative.