No single AI model writes the best marketing copy for every business, and published rankings will not tell you which suits your voice. The reliable way to choose is to run one brief, one voice sample and one edit note through ChatGPT, Claude and Gemini, then score the drafts on a weighted scorecard.
Published leaderboards rank models on exams and puzzles, which says little about whether a headline sounds like your business. The test below takes about 90 minutes, costs nothing on the free tiers, and leaves you with a brief template you can reuse every time a model updates.
Set up the test so it is fair
Suppose a two-branch optician wants copy for a landing page promoting a children's eye test. The owner has a short brief, a paragraph of past writing she likes, and a list of facts (opening hours, the test is free for under-16s, appointments take 30 minutes). Every assistant gets exactly the same package. If one gets a better prompt, you are testing the prompt.
The package has five parts:
- The task: a 150-word landing page section with a headline, three short benefits and a call to action.
- The audience: parents of children aged 4-12, worried about school reading, not experts.
- The facts, and only these facts: the list the owner supplied. Anything else is off limits.
- The voice sample: 120 to 200 words of past writing, plus three words that describe how she sounds ("warm, plain, a little dry").
- The rules: banned words, spelling variety, length.
Write a 150-word landing page section for a children's eye test at a two-branch optician.
Audience: parents of children aged 4-12 who are not experts.
Use ONLY these facts: free for under-16s; 30-minute appointment; open Saturdays; no referral needed.
Voice: match the sample below (warm, plain, a little dry). British spelling.
Do not use: "state-of-the-art", "peace of mind", exclamation marks.
Structure: a headline, three short benefits, one call to action.
Voice sample: [paste 150 words of the owner's own writing]Open each assistant in a new chat, paste the package, and save the first draft without changing anything. That first draft is round one. Resist the urge to nudge.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowA six-line scorecard weighted for voice and facts
Score each draft from 1 to 5 on six lines, writing a one-line reason for every score. The reasons matter more than the numbers, because they show you what to fix.
| Line | What you are looking for | Weight |
|---|---|---|
| Voice match | Could the owner have written it? Would a regular customer notice a change? | x3 |
| Facts kept | Only supplied facts, none invented | x3 |
| Clarity | One idea per sentence; a parent understands it in one read | x2 |
| Angle | A specific hook, not a generic "care about your child's eyes" | x2 |
| Rules followed | Banned words absent, length within 10% | x1 |
| Edit time | Minutes to make it publishable (5 = under 5 minutes) | x2 |
Weights add to 13, so the maximum is 65. Weight voice and facts heavily: they are the two failures customers and regulators see first, and they are the two that cost most to fix. If you work in a regulated area, such as eye health, make "facts kept" a pass or fail line: any invented claim fails the draft outright.
Reading the three first drafts blind
Before you score, hide which assistant wrote which draft. Paste the three into one document as Draft A, B and C (ask someone else to shuffle them if you can) and unmask them only after the sums are done. Models update every few months, so any reputation you have heard is older than the drafts in front of you. Look for these four things in each:
| What to look at | A good sign | A sign it needs heavy editing |
|---|---|---|
| Headline | Uses a fact from your list and sounds like the sample | A rhetorical question or a generic promise about caring |
| Body copy | Short sentences, the owner's dryness kept | Stacked adjectives, gushing warmth, a softened call to action |
| "Only these facts" | Flags what is missing instead of filling it | Adds a benefit or a sector commonplace you never supplied |
| Angles | Several genuinely different hooks when asked | Ten versions of the same idea |
An illustrative first-round headline set might read: Draft A, "Is your child squinting at the board? Book a free 30-minute eye test"; Draft B, "A free eye test for your child, in half an hour, on a Saturday"; Draft C, "Free children's eye tests, no referral needed". All three are usable, but only one sounds like the owner's dry voice, and the rhetorical question in the first is exactly what her sample avoids. You will not know which assistant wrote the best fit until you unmask the drafts.
Round two: the edit instruction test
A good copywriting model is not the one with the best first draft; it is the one that takes direction well. Send each assistant the same single edit note, written the way a client would:
Better. Make it warmer but shorter (120 words). Keep the Saturday opening in the first
benefit. Replace the headline with one that does not ask a question. Keep every fact.Now score three new things: did it do all four instructions, did it keep the parts you did not mention, and did it add anything new? A common failure is that the assistant fixes the instructions but rewrites other lines, so a sentence you liked disappears. Another is that "warmer" turns into gushing. Whichever assistant follows the note most precisely is easier to work with day to day, and that matters more than a brilliant first draft you cannot steer.
Round three: copy in a different voice
Repeat the test with a second business and a completely different sample, to check the model is adapting to the voice and not just applying its own. Say a veterinary practice with a brisk, practical tone wants an email reminding owners that annual boosters are due. Use a sample that is short and direct: "Your dog's booster is due. It takes ten minutes. Book online or ring us." If the output opens with a warm paragraph about how much the practice cares, the assistant has ignored the sample and defaulted to its own voice.
This second test is where the first round's leader may slip. Treat the two businesses as a small data set: a model that wins one and loses one is telling you its strength depends on the voice, which is useful to know for a client with an unusual one.
Logging the edits each draft needed
Keep a log of the edits you make, because that is your real cost. A simple tally works.
- Cut: filler openings, stacked adjectives, throat-clearing sentences. Count the words removed per 150 words.
- Fix: any claim not in your facts list. Count these separately; one is a warning, three is a reason to drop the draft.
- Reshape: rhythm and structure, such as every sentence the same length, or every paragraph ending on a neat summary line.
- Restore: your own turns of phrase the assistant flattened.
An illustrative tally for the optician page: Draft A, 38 words cut, 1 added claim, 2 rhythm fixes; Draft B, 12 words cut, 0 added claims, 1 phrase restored; Draft C, 20 words cut, 1 added claim, 3 phrases restored. The numbers are invented for the example, but this is the shape of record to keep. If an assistant's drafts consistently take you 25 minutes to fix, it has not saved you time, however good the first line looks.
A filled-in scorecard, with the sums
Here is how one scored round might look for the optician page, with invented but realistic marks and the drafts still masked. Each score is multiplied by its weight.
| Line (weight) | Draft A | Draft B | Draft C |
|---|---|---|---|
| Voice match (x3) | 3 = 9 | 5 = 15 | 3 = 9 |
| Facts kept (x3) | 4 = 12 | 5 = 15 | 4 = 12 |
| Clarity (x2) | 4 = 8 | 4 = 8 | 5 = 10 |
| Angle (x2) | 5 = 10 | 3 = 6 | 3 = 6 |
| Rules followed (x1) | 3 = 3 | 5 = 5 | 4 = 4 |
| Edit time (x2) | 3 = 6 | 4 = 8 | 4 = 8 |
| Total (of 65) | 48 | 57 | 49 |
The pattern in this invented round is worth expecting: one draft wins on voice and discipline, another on the angle. Notice what the sum hides. Draft A scored 5 on angle, the line that matters most when you are stuck for ideas. Once you unmask, a sensible owner might draft with Draft B's assistant and ask Draft A's for ten alternative headlines. The scorecard is there to show which job each model is best at, not just to declare one winner.
Editing a stock paragraph back into the owner's voice
Seeing an edit makes the scorecard concrete. Here is an illustrative first-round paragraph from an assistant that did not match the sample, followed by the edited version.
Before: "At our practice, we are passionate about your child's vision. Our experienced team uses the latest technology to ensure every child sees the world clearly. Book today and give your child the gift of sight!"
After: "A child who cannot see the board rarely says so. They fidget, lose their place and decide they are bad at reading. The test takes 30 minutes, costs nothing for under-16s, and you do not need a referral. We are open on Saturdays."
What changed: three stock claims ("passionate", "latest technology", "ensure") became one concrete observation; the exclamation mark and the gift line went; the facts moved up. Only the facts in the owner's list remain. The edit took about six minutes, and most of it was deleting. Time yourself like this on each assistant's draft; deleting is quick, while rebuilding a voice from scratch is not.
Short copy needs a different test
Subject lines, ad headlines and button text are a separate skill from paragraphs. A model that writes a fine section can still produce ten near-identical subject lines. Add a short-copy round: ask each assistant for twelve subject lines for the booster email, grouped by approach (urgency, plain statement, question, number, benefit, curiosity), then score variety, not quality. How many of the twelve are really different? How many are over your 45-character limit? Which ones sound like your voice sample?
A practical rule: keep any three that differ in approach and test them in your email tool against each other. A model's guess about the best line is a guess. The test result is data. If you run ads, the ad-variants tutorial in the further reads carries on from here.
Mistakes that make the test misleading
- Different prompts. Copy-paste the same package to all three. Retyping it by hand introduces differences.
- Reading the name. Strip the model name before scoring if someone else can paste and shuffle the drafts for you.
- Testing one brief. A model that wins an eye-test page may lose on a tradesperson's quote follow-up. Use at least two.
- Weak voice sample. If your sample is two lines, every model will guess. Use 150 words of writing you are proud of and that was not written by an assistant.
- Ignoring plan limits. A free tier that cuts you off after a few messages makes round two impossible. Check the daily limits on the vendor pages before you start.
- Treating a tie as a failure. A tie means any of the two will do, so pick on cost and convenience.
Where the owner stays in charge
Copy carries your name, so a few decisions are not for the assistant. What you promise, what you charge, who you compare yourself with and what you leave unsaid are business choices. A model will happily write that you are the "leading" optician in the area, and that is a claim you must be able to support. Keep a short list of claims you may make and claims you may never make, and paste it into every brief under AVOID. When the list is part of the template, you stop relying on memory at 5pm on a Friday.
It also helps to save every winning draft and its brief together. After a few months you have a folder of copy that worked, which is a better voice sample than anything you could write fresh, and a record of which assistant produced which. That folder is worth more than any subscription.
A last word on budgets of time. Plan 90 minutes for the first full test and about 30 for every rerun, since the brief, sample and scorecard already exist. Put a reminder in your calendar for the week after any major model release, and treat each rerun as a short, boring, useful habit: same brief, same sample, same sums, new scores. Over a year, three or four reruns will tell you more about the assistants than any review, because they are measured against your own copy and your own customers.
What a paid month costs once you have a winner
For individual use, ChatGPT Plus is $20 a month, Claude Pro $20 a month ($17 monthly-equivalent billed annually) and Google AI Pro $19.99 a month. Team plans on ChatGPT Business and Claude Team are $25 a seat monthly or $20 billed annually, with a two-seat minimum. Free tiers exist for all three, with tighter limits. All of these were on the vendors' pages in November 2026; check them on the day you buy.
For a small business writing its own copy, do the test on free tiers, pick the winner, and pay for one plan only if you hit its limits. A freelance copywriter should read the monthly tool budget tutorial for what a full stack costs. Shared instructions help too: Claude Projects (Free accounts can create up to five) and ChatGPT Projects can hold your voice sample and banned-word list so you do not paste them each time. Gemini's Gems do the same job; on personal Google accounts they are becoming skills from November 2026 (work accounts follow in 2027), and Google recreates existing Gems as skills automatically.
A reusable brief template
The brief you wrote for the test is the real asset. Keep it as a template and fill it in per job.
JOB: [what is being written, where it will appear, word count]
AUDIENCE: [who reads it, what they already know, what worries them]
GOAL: [the single action you want]
FACTS (use only these): [list]
VOICE: [3 words + paste sample]
AVOID: [banned words, claims you cannot make, tone traps]
FORMAT: [headline, sections, call to action]
PROOF AVAILABLE: [reviews, numbers, guarantees you can state]The "proof available" line is the one most people forget. Without it, an assistant invents proof. With it, you can say "use the 4.8 average from 212 reviews" and it will. Illustratively, filled in for a veterinary practice's booster email, it might say: "Proof: 14 years open; same-day appointments on weekdays; reminders go out two weeks before due date". Nothing more.
Copy checks no assistant does for you
Whichever model wins, five checks stay with you before anything is published.
- Claims: every claim about results, prices, qualifications or availability matches reality today.
- Customer words: compare against real review language; if customers say "check-up" and the copy says "assessment", change it.
- Sameness: read the first sentence aloud. If it could open any competitor's page, rewrite it.
- Rules for the channel: advertising and platform rules for your sector can limit what you may say; read the current rules for any platform you pay to advertise on.
- Ownership of the voice: paste a few of your own paragraphs into the voice sample, not another writer's work.
For a method that digs real customer language out of reviews, the voice-of-customer tutorial pairs well with this test, and for protecting a client's voice over time see the client-voice workflow.
Reading your results and deciding
Add up the weighted scores. If one assistant leads by more than eight points out of 65 on both businesses, use it as your main model. If the lead is under five, choose on practical grounds: which has the interface you prefer, which your team already pays for, which handles your file types. If each wins a different business, keep a primary model and use the other as a second pair of eyes: paste the draft into it with "critique this against my voice sample and list three weaknesses". A second opinion from a different model is one of the cheapest edits available.
Rerun the test whenever a model is replaced, or twice a year at least. It takes under two hours with the template, and it protects you from paying for last year's best. If you would like help running it on your own brief, that is what my AI implementation consultation covers.
Questions about testing AI models on your own copy
How many briefs should I test before choosing?
Three is enough for a small business: one short piece (an ad or subject line set), one medium (a service page section) and one that needs your voice (a founder email). A single brief can mislead, because each assistant has a type of copy it handles best.
Should I let the assistants see each other's drafts?
Not in the first round. Keep each draft independent so your scoring compares like with like. In a later round you can paste the best draft into another assistant and ask it to critique, which is a good use of a second model.
Do paid plans write better copy than free ones?
Paid plans give you stronger models, larger limits and more attachments, but the biggest gains come from a better brief and voice sample. Run the test on free tiers first, then repeat on a paid month if the free results are close.
Can I use the winning model's copy without disclosure?
That depends on your clients, platform rules and any contract wording. Read the terms of any ad platform or marketplace you publish on, and the terms of any client you write for. When unsure, ask. Make sure every claim in the copy is true and that you own the facts.
Further reads
- AI Copywriting Tool Examples: Before-and-After Rewrites With Prompts — Before-and-after rewrites with the prompts used.
- AI Tools for Copywriting: A Freelancer's Draft-to-Delivery Workflow — A full workflow from draft to client delivery.
- ChatGPT Prompts for Copywriters: Research, Angles and Edits — Prompts for research, angles and edits.
- Can AI Help Copywriters? Research, Drafts and Edits for Client Work — Where AI helps copywriters across research, drafts and edits.
- Best AI for Copywriting Ads: Variants, Testing and Client Sign-Off — Take the winning model into ad variants and testing.
- Best AI for Writing Copy: A Freelance Copywriter's Edit Workflow — How an experienced copywriter edits AI drafts.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: claude.com/pricing, gemini.google/subscriptions, chatgpt.com/pricing, OpenAI's ChatGPT Business pricing page and Google's Gems-to-skills help page, checked November 2026; the scorecard is a method, not a benchmark.