Marketers · problem and fix
AI first draft marketing copy in your brand voice on Mac: fixes for marketers in 2026
You can get an AI first draft of marketing copy in about a minute. Getting it to sound like your brand takes a voice card, a few real samples, and a human edit before anything goes live. This page covers what off-brand copy costs you, why AI drifts away from your voice, and fixes you can use on a Mac today.
The short answer
- An AI first draft is fast. A draft in your brand voice needs a voice card and real samples.
- Ahrefs' 2026 survey found that 66% of marketers trust AI content only after a human reviews it.
- Models drift toward a typical, generic voice, and AI-assisted writing tends to look more alike across writers.
- Start free: a one-page voice card, three to five real samples, and a banned words list.
- Lazur drafts and types text and keeps snippets. It cannot know your voice without the samples you give it.
What ai copy that sounds on-brand costs you
Editing time you did not plan for
A draft that is mostly right still needs a full read. You rewrite template-like sentences, fix claims, and check tone. The minutes you save on the first draft can return as edit time if the voice is off. Track edit time for one asset type for a week and compare it with your old process.
Readers who sense the gap
Ahrefs' 2026 survey found that 56% of AI users say human-written content is better quality than their AI content, and 23% say it is way better. Readers may not name the difference, but they feel it in the first line.
Trust that depends on a person
Only 2.7% of respondents trust AI content with no review at all. If your team publishes AI copy without a check, your process sits outside the norm the survey describes. That matters when a reviewer asks why a post sounds different from the rest.
Workload that already feels heavy
CMI's February 2026 survey of 644 full-time marketers found that 59% say their workload feels overwhelming, and 76% say they do the work of more than one job. The survey is about general workload. A draft that needs heavy rewriting can add work instead of cutting it.
Claims you never approved
Brand voice is not only tone. Generic copy can drift into promises and numbers that nobody signed off on. A Shorthand VP quoted in Ahrefs' article says AI copy can make a brand sound like every other content platform, so her team keeps personality-heavy copy human-written.
Why it happens
Models default to the average voice
Without strong signals, a model returns the most typical version of a post. Padmakumar and He found that writing with InstructGPT significantly reduced diversity and made different authors' writing more similar. The sameness came from the model's text, not the writers' text.
Assistant framing leaks into your message
In a CHI 2023 study with 1,506 participants, an assistant set to argue one side shifted the opinions people wrote and their later attitudes. Cornell reports that people using it were twice as likely to write a paragraph agreeing with the assistant. A default slant in a tool can become your slant.
Shared prompts produce look-alike copy
Doshi and Hauser's dataset of 293 participants found that AI-enabled stories were more similar to each other than stories by humans alone. If your team uses one generic prompt, the posts will rhyme. The fix is to give each drafter the same specific samples, not the same generic instruction.
Informal voices are hard to copy
Wang and colleagues tested more than 40,000 generations per model across writing from more than 400 real authors. Models did better with structured formats like news and email. They struggled with nuanced, informal blogs and forums, which is where many brand voices live.
Vague prompts ask for vague output
A prompt that says make it on brand gives the model nothing to match. Anthropic's guide says examples are among the most reliable ways to steer tone. Without examples, the model guesses, and the guess tends toward the middle.
How to fix it
These work with or without any app.
Write a one-page voice card
Keep the card to one page in a Google Doc. Name five traits and give a counter-example for each. Add sentence rhythm, your usual opener and closer, words you use, words you avoid, and rules for numbers and claims. Pin the link where your team drafts. This costs nothing and works in any AI tool that accepts pasted text.
Paste three to five real samples
Anthropic's prompting guide suggests three to five examples for best results, and it recommends examples that are relevant and diverse. Pick samples from work you are proud of. Mix an email, a short social post, and a help article. Skip weak samples, because the model copies flaws as easily as strengths.
Draft in steps
Ask for three openers first, then pick one. Ask for the body next. Then ask for an edit against the voice card, and ask the model to name the two rules it broke. Small steps show where drift starts. One long prompt hides the drift in the middle of the draft.
Build a banned words and claims list
List the hype words your readers roll their eyes at, and the phrases your legal or product team never allows. Search each draft against the list before you read it. Add every word you delete during edits, because each deletion marks a default the model reached for.
Use Writing Tools for a second pass
Select text in any app, Control-click the selection, and choose Show Writing Tools. Rewrite offers Friendly, Professional, and Concise, and Proofread checks mechanics. These tools push text toward a general tone, so use them to tighten a paragraph you already wrote in your voice. Availability depends on Apple Intelligence support in your language and region.
Edit against the card with a checklist
Read the draft aloud once. Check each opener, each number, and each claim. Replace any sentence that could appear on a competitor's site. Keep the best human line, and delete the rest. A five-item checklist beats a vague instruction to make it sound like us.
Keep personality lines human
Write the hook and the closing line yourself, and let AI handle structure, variants, and boilerplate. Shorthand's approach, described in Ahrefs' article, keeps personality-heavy copy human-written. You do not need to hand over the whole post to get the time savings.
Run a blind two-reader test
Give two colleagues a published post you like and the new draft, without labels. Ask which one sounds more like the brand, and why. Their answers show drift that you stopped seeing after the tenth read. Update the voice card with the reasons they give.
What off-brand AI copy costs you
Most of the cost is hidden. The draft arrives fast, so it feels like a win. Then you spend the saved time rewriting sentences that could have come from any company. Ahrefs' 2026 survey of 301 respondents found that 280 of them, or 93%, use AI for content creation. Two-thirds said they trust AI-generated content only if a human reviews it first.
66%
of marketers trust AI-generated content only if a human reviews it first
69% of Marketers Are Publishing More AI Content Than Last Year (Ahrefs)56%
of AI users say human-written content is better quality than their AI content
69% of Marketers Are Publishing More AI Content Than Last Year (Ahrefs)59%
of full-time marketers say their workload feels overwhelming (CMI survey, February 2026, 644 respondents)
2026 Marketing Career Outlook: AI's Surprising Impact (Content Marketing Institute)1,506
participants in a CHI 2023 study where a biased AI assistant shifted what they wrote and believed
Co-Writing with Opinionated Language Models Affects Users' Views (arXiv 2302.00560)Those numbers describe a team that expects a review step. Only 2.7% of respondents trust AI content with no review at all. If your draft ships without a check, you are in a small group, and you are the one who answers for the tone.
Readers notice too. In the same survey, 56% of AI users say human-written content is better quality than their AI content, and 23% say it is way better. The gap is not only about grammar. It is about the feeling that a real person wrote the post.
| Where it shows up | What it looks like | Quick check |
|---|---|---|
| Openers | Every post starts with the same generic hook | Read the first line aloud. Would you say it in a meeting? |
| Vocabulary | Words your brand never uses appear | Search the draft against your banned words list |
| Claims | Promises or numbers appear that nobody approved | Mark every number and check it against a source |
| Rhythm | Every sentence has the same length and shape | Count words per sentence in one paragraph |
| Personality | Jokes, slang, and asides disappear | Compare with your three best human-written posts |
Time is the other cost. CMI's February 2026 survey of 644 full-time marketers found that 59% say their workload feels overwhelming, and 76% say they do the work of more than one job. That survey measures general workload, not AI-specific work. Still, a draft that needs heavy rewriting adds work to a week that already has no room.
Why AI copy drifts away from your voice
A language model writes from patterns in a large amount of text. Without strong signals, it returns the most typical version of a post. That version is smooth, polite, and easy to forget. Your brand voice lives in details the model cannot see, such as how you open, which words you avoid, and how bluntly you say things.
Padmakumar and He (arXiv 2309.05196) tested writing with language models. Writing with InstructGPT significantly reduced diversity in the text, and different authors' writing became more similar. The parts the writers contributed stayed unaffected. The sameness came from the model's own sentences, not from the people. Those models date from 2023, so test current tools on your own samples.
Jakesch and colleagues ran a CHI 2023 study with 1,506 participants (arXiv 2302.00560). Participants wrote with an assistant set up to argue that social media was good or bad. The assistant's opinions shifted the opinions in participants' writing and their later attitudes. Cornell's write-up says people using the biased assistant were twice as likely to write a paragraph agreeing with it. For marketers, a model's built-in framing can leak into your message.
Doshi and Hauser's dataset covers 293 participants who wrote short stories with and without generative AI ideas. The authors describe AI-enabled stories as more similar to each other than stories by humans alone. If five people on your team use the same prompt, expect five posts that sound alike.
Why a voice guide alone often falls short
Wang and colleagues (arXiv 2509.14543) tested whether language models can imitate an individual's writing style from a few samples. They ran more than 40,000 generations per model and used writing from more than 400 real authors. Models approximated styles in structured formats like news and email. They struggled with nuanced, informal writing in blogs and forums. Many brand voices are informal, so expect that gap.
A voice card made of adjectives still helps, but it is weak on its own. Friendly, bold, and smart describes almost any company. Samples carry more signal. A banned words list carries even more, because it removes the defaults the model reaches for first.
Fixes you can use today, without buying anything
Each fix below uses tools you already have, such as Google Docs, a notes app, and Writing Tools on a Mac where Apple Intelligence is available. None needs a new subscription. Start with the first two, because they solve most of the drift. The rest tighten the result.
Write a one-page voice card
Keep the card to one page in a Google Doc. Name five traits and give a counter-example for each. For example, write: Plain, not cute. Say 'ships Tuesday,' not 'get ready to be excited about Tuesday.' Add sentence rhythm, your usual opener and closer, words you use, words you avoid, and how you handle numbers and claims. Pin the link where your team drafts.
Paste three to five real samples
Anthropic's prompting guide says examples are one of the most reliable ways to steer tone. It suggests three to five examples for best results, and it recommends examples that are relevant and diverse. Pick samples from work you are proud of. Mix formats, such as an email, a short social post, and a help article. Skip weak samples, because the model copies flaws as easily as strengths.
Draft in steps, not in one prompt
Ask for three openers first. Pick one, then ask for the body. Then ask for an edit against your voice card, and ask the model to name the two rules it broke. Small steps show you where drift starts. One giant prompt hides it in the middle of a long draft.
Use Writing Tools for a second pass, not the first
Select the text in any app, Control-click the selection, and choose Show Writing Tools. Apple's guide lists Rewrite with Friendly, Professional, or Concise, plus Proofread and Summarize. Those rewrites move text toward a general tone. Use them to tighten a paragraph you already wrote in your voice, not to write it from scratch. Compose uses the ChatGPT extension and is only available where ChatGPT is. Apple Intelligence is not available in every language or region, and usage limits may apply.
Keep a banned words list and a claims list
Write down the hype words your readers roll their eyes at, and the phrases your legal or product team never allows. Search each draft against that list before you read it. Then add the words you deleted during edits. Every deleted word marks a default the model reaches for.
A worked example: one launch email, start to finish
Say you launch a reporting tab in your app. You have a voice card, four past emails, and a launch note from product. Open the Google Doc with the card and samples. Ask for three subject lines in your voice, then pick one. Ask for a body in three short paragraphs, each with one concrete detail from the product note.
A weak first draft often reads like this: 'We are thrilled to announce a powerful new way to see your data.' The fix is specific: 'The new reporting tab shows last week's numbers on one screen. No more copying rows into a sheet on Monday morning.' You removed the empty adjective. You added a fact a reader can check.
Then run the checklist. Search for banned words. Mark every number and confirm it against the product note. Read the email aloud once. Ask one colleague which lines sound like you. Log how many minutes the edit took, so you can compare it with your usual process next week.
Where Lazur helps, and where it does not
Lazur is a voice assistant for the Mac. It lives in the notch, listens and looks at your screen only while you hold the shortcut, and answers out loud. For this problem, it can draft text and type it into the app you are working in. Dictation can speak your first rough version at the cursor. It cannot know your voice without the card and samples you give it. It does not post, send, or click for you, so the review step stays with you.
Two features matter most here. Snippets, a personal dictionary, and clipboard slots keep approved wording the same from draft to draft. Hold fn to dictate clean, formatted text at the cursor in any app. Some writers find that speaking gives a looser first draft. Try it on one post and compare the result with your typed draft.
If you are unsure where a control lives, ask Lazur to walk you through it. Lazur guides you one click at a time and waits for each click, so you learn the path yourself. For a quick location, ask it to point. It draws a circle and a label around the control on your screen.
Edge cases: regulated work, shared drafts, and AI checks
If you work in law, nursing, or accounting, follow your employer's rules on client or patient data before you paste anything into a prompt. Remove names, account numbers, and case details from samples. Lazur sends audio and the screenshot to AI providers to answer them, and they are not stored. Your employer's policy still decides what you may share.
Some teams ask AI to check AI. Ahrefs found that 29% of AI users now ask AI to check AI content, up from 14% the year before. Manual self-checking was still the most common method. A second model can catch some problems, but it cannot approve your claims. Keep a person on facts.
If your team drafts in Microsoft Word, Microsoft's Copilot in Word page advises including reference documents and the required writing style to help produce a first draft in minutes. The page does not mention Mac or brand voice. Test that approach on one sample before you rely on it.
Mistakes that bring the drift back
- Writing a voice card of adjectives only, with no samples behind it.
- Using the same generic prompt for every channel, from email to social.
- Letting a Writing Tools rewrite remove the personality from a draft.
- Accepting numbers and claims without checking the source.
- Skipping the review because the draft reads smoothly.
- Never updating the card after the edits teach you something new.
What to do next
Start small. Pick one asset type, such as a launch email or a short social post, and run this routine for one week. Keep the card and samples in one Google Doc so every drafter uses the same version. If your team also reviews copy in Google Docs, the Google Docs guide for marketers covers review and comment habits.
Judge the result by edit time and reader feedback, not by how fast the first draft appeared. A draft that takes longer to fix than to write does not save time. If you want to try the voice assistant, the download page has the Mac app, and pricing lists the plans.
Faster with Lazur
Where Lazur helps with ai copy that sounds on-brand
Hold ⌥ ⌘ and say what you want. Lazur sees your Google Docs window. Lazur reads what's on screen and writes the draft at your cursor for you to check.
- 1
Hold ⌥⌘ and talk
- 2
Lazur reads what's on screen and writes the draft at your cursor for you to check.
- 3
You stay in Google Docs. Nothing to copy into a chat window.
More things to say in Google Docs
- “Draft a launch email from my voice card and these three samples.”
- “Hold fn and dictate the first rough version of this post.”
- “Insert my standard disclaimer snippet.”
- “Show me where Writing Tools lives in this window.”
- “Walk me through using Writing Tools on this selection.”
- “Explain why this sentence sounds generic.”
Free for 3 days, then $18/month. Download for Mac
You say“Draft a three-sentence intro for this post using my voice card.”
Lazur · drafts itLazur types a draft into the document. You compare it with the voice card and cut any line that sounds generic.
Draft at your cursor
Our new reporting tab shows last week's numbers on one screen. Open it on Monday and skip the spreadsheet hunt. Spend the time on the work your customers actually asked for.
What marketers say to Lazur for this
Each one is a single hold of ⌥ ⌘.
“Draft a launch email from my voice card and these three samples.”
Lazur drafts the text and types it into the app you are working in. You still check it against your voice card before it goes anywhere.
Drafts it
“Hold fn and dictate the first rough version of this post.”
Dictation types clean, formatted text at the cursor in any app. Some writers find speaking gives a looser first draft, so test it on one post.
Types what you say
“Insert my standard disclaimer snippet.”
Snippets, a personal dictionary, and clipboard slots keep approved wording identical across drafts, so you do not retype it.
Types what you say
“Show me where Writing Tools lives in this window.”
Lazur draws a circle and a label around the spot on your screen, so you do not hunt through menus.
Points at it
“Walk me through using Writing Tools on this selection.”
Lazur guides you one click at a time and waits for each click. It does not click for you.
Walks you through it
“Explain why this sentence sounds generic.”
Lazur looks at your screen while you hold the shortcut and answers out loud, so you can test a fix right away.
Explains what's on screen
“Research how three competitors describe their tone on their homepages.”
Research runs in Lazur's own Chrome tabs and comes back with a summary and a report. You check the sources yourself.
Researches in the background
A plan you can start this week
- Day 1
Write the voice card
Draft one page with five traits, a counter-example for each, your usual opener and closer, and your banned words. Keep it in one Google Doc and share the link with every drafter.
- Day 2
Collect three to five samples
Pull past posts, emails, and help articles that you think sound like the brand. Pick a mix of formats. Remove client names, account numbers, and any detail your employer's policy restricts.
- Day 3
Draft one asset type in steps
Ask for three openers, pick one, then ask for the body. Log the minutes you spend editing. Do not skip the log, because it is your only proof that the routine saves time.
- Day 4
Run the checklist and update the list
Search for banned words, check every number against a source, and read the draft aloud. Add each word you deleted to the banned list, and note any claim that needed a source.
- Day 5
Run the blind test and review
Give two colleagues the draft and a published post without labels. Ask which sounds more like the brand. Update the voice card with their reasons, and compare this week's edit minutes with your usual process.
Questions people ask
Can AI write marketing copy in my brand voice on its own?
Not reliably. Without samples, the model falls back to a generic middle voice. Give it a voice card and three to five real samples, then review every draft against them.
How many examples should I give an AI for tone?
Anthropic's prompting guide suggests three to five examples for best results, chosen to match the use case and cover different cases. Start with three. Add another only when the output misses a specific pattern.
Does Apple Writing Tools match my brand voice?
Not by itself. Writing Tools offers Friendly, Professional, and Concise rewrites, plus Proofread. Those move text toward a general tone. Use them to tighten copy you already drafted in your voice. Availability depends on Apple Intelligence support in your language and region.
Do I need to buy a tool to fix AI drift?
No. A voice card, a sample set, a banned words list, and a checklist cost nothing. A paid tool might save some pasting, but it does not replace the review step.
How often should I review AI drafts before publishing?
Review every draft before it goes out. Ahrefs' 2026 survey found that 66% of marketers trust AI content only after a human reviews it first, and 2.7% trust it with no review at all.
Can I put client or patient details into an AI prompt?
Follow your employer's rules first. If you work in law, nursing, or accounting, client or patient data may be restricted. Strip identifying details from samples unless your policy allows them. Lazur listens and looks at the screen only while you hold the shortcut.
Why does my AI copy still sound generic with a brand guide?
Guides written as adjectives stay vague, and models struggle to imitate informal, nuanced voices. Samples and a banned list give more signal. Expect to do the final pass by hand.
Does Lazur write in my voice?
Lazur drafts and types text using the voice card and samples you paste into the request. It does not post, send, or click for you, so you check the draft before it goes anywhere.
Sources
- 69% of Marketers Are Publishing More AI Content Than Last Year (Ahrefs) · Ahrefs · checked 2026-11-08
- 2026 Marketing Career Outlook: AI's Surprising Impact · Content Marketing Institute · checked 2026-11-08
- Does Writing with Language Models Reduce Content Diversity? · arXiv · checked 2026-11-08
- Co-Writing with Opinionated Language Models Affects Users' Views · arXiv / CHI 2023 · checked 2026-11-08
- Writing with AI help can shift your opinions · Cornell Chronicle · checked 2026-11-08
- Dryad dataset for Doshi & Hauser, generative AI and creative writing · Dryad · checked 2026-11-08
- Catch Me If You Can? Not Yet: LLMs Still Struggle to Imitate the Implicit Writing Styles of Everyday Authors · arXiv · checked 2026-11-08
- Prompting best practices · Anthropic · checked 2026-11-08
- Copilot in Word · Microsoft · checked 2026-11-08
- Use Writing Tools with Apple Intelligence on Mac · Apple Support · checked 2026-11-08
Fix ai copy that sounds on-brand with Lazur, free for 3 days
Hold a key and ask. Lazur sees your screen, points at the button and talks you through it, in any Mac app. No credit card.