Skip to content
AIupick

Guide for marketing · 6 jobs

AI for marketing: copy in your voice, ads you may run, numbers you can check

Getting a draft out of AI is the easy part. Whether it can go out under your brand depends on what surrounds it: your voice saved once, claims you can prove, images you may run, and numbers and publishing steps someone has checked.

  1. 01

    One brief, several copy variants, in your brand's voice

    Asking for "10 headlines" in an empty chat gives ten versions of one generic idea, in nobody's voice. Save your voice guide and a few approved examples in one workspace, and ask for variants that each test a different angle, so the set is worth testing.

    Example

    A meal-kit brand launches a spring offer. The workspace holds the voice guide, a list of words the brand never uses and three emails that did well. One brief returns six subject lines, two each on price, time saved and taste, each marked with its angle; you test one from each pair and add the winner to the examples.

    How to do it

    1. 1Create a Project in Claude, or a notebook in Microsoft Copilot, for the brand. Add the voice guide, the words you never use and three to five pieces of approved copy that did well.
    2. 2In a chat inside it, paste the brief with the prompt on this page: audience, offer, the one message, the channel and its character limit.
    3. 3Ask for variants that differ by angle, such as price, time saved or proof, not by synonyms. Drop any two that say the same thing.
    4. 4If the team edits copy in DeepL Write, save your spelling and formatting rules as Style rules (Write Pro add-on, Business and Enterprise), so every suggestion follows them.
    5. 5After the test, add the winning variant to the workspace as a new example.

    Built in

    Prompt to copy

    Brief: [product or offer] for [audience]. One message: [what they should believe or do].
Channel: [email subject line / ad headline / social post], at most [n] characters.
Write [6] variants in the voice of the guide and examples in this workspace: two each for the angles [angle 1], [angle 2] and [angle 3].
For each, name the angle and the claim it makes. Use only facts from the brief; where a variant needs a fact I did not give, write [needs proof] instead of inventing it.

    Where it goes wrong

    • AI miscounts characters. Check the length of each variant in the ad tool's preview before you upload.
    • It invents claims nobody approved, such as "the fastest" or a discount you did not offer. The [needs proof] marker shows where; check the rest against the brief.
  2. 02

    Take a campaign into other languages

    A word-for-word translation keeps the meaning but loses the hook, gets the form of address wrong and runs past the character limit. Give the AI fixed terms and limits for each market, ask for adaptations with a back-translation, and have a native speaker check prices, claims and legal lines.

    Example

    The headline "Dinner, sorted" goes to Germany and France. The glossary keeps product names in English and says whether each market uses formal or informal address. You get two headlines per market, each with a back-translation, so the team at home sees what the German line says; the native speaker then catches a price written with a decimal point instead of a comma.

    How to do it

    1. 1For each market, write down the form of address, the terms that stay in English, approved translations of key terms and the character limit of each field. Keep it in the brand's workspace, such as a Project in Claude.
    2. 2Paste the copy with the prompt on this page: two options per line, each with a back-translation into English.
    3. 3If you polish the result in DeepL Write, add brand and product names to Terms to never edit, with the language they apply to, so its suggestions leave them alone.
    4. 4For the campaign video, Dubbing in ElevenLabs translates it and keeps each speaker's voice.
    5. 5Have a native speaker in each market check the final copy, starting with prices, dates, units, legal lines and every place the AI marked as unsure.

    Built in

    Prompt to copy

    Adapt this campaign copy for [market], in [language]. Keep the message and the effect of the hook; do not translate word for word.
Address the reader [formally / informally]. Keep these terms unchanged: [terms]. Use these translations: [term = translation].
For each line: two options of at most [n] characters, a back-translation into English, and a note on anything that may not work in [market], such as a pun, a reference or a claim.
Write prices, dates and units the way [market] does, and mark every place you were unsure.

    Where it goes wrong

    • Translated lines often run longer. Check each one in the real ad preview, not only in the document.
    • A claim allowed in one country can be restricted in another, for example on health or the environment. Ask someone who knows the local rules before launch.
  3. 03

    Ad images with readable text, one style, that you may use commercially

    Many image models garble the words in an ad, and every new prompt drifts a little from the brand's look. Save the style once from your own visuals, use a model built for text, and check what your plan allows before the image goes into a paid campaign.

    Example

    Four social ads for a running shoe, each with a three-word headline. The style comes from ten past campaign images; the headline is in quotation marks in the prompt and placed in the top third. When the headline changes, you edit the text layer instead of generating the image again, and because they were made on a plan that keeps images private, nobody saw them before launch.

    How to do it

    1. 1Save the brand look once: Custom styles in Recraft turns up to 10 of your own visuals into a style that new images follow.
    2. 2For text in the image, use Ideogram: put the exact words in quotation marks, mention them early and say where they go. Keep the text short, and add prices and legal lines in your design tool.
    3. 3For several variants in one layout, write a JSON prompt in Ideogram, which fixes the layout, text placement and hex colours, and change only the words.
    4. 4To change the words later, open the image in Studio in Ideogram and use Layerize text, which turns the text into editable layers.
    5. 5Before the ad runs, check your plan. Recraft's paid plans give you ownership and commercial rights for images made while subscribed; Free plan images belong to Recraft and are not licensed for commercial use. Ideogram does not claim ownership of generated images and allows commercial use under its Terms of Service.

    Built in

    Prompt to copy

    Social ad, [4:5]. Headline "[3 to 5 words]" in large bold letters in the top third, [colour] on a plain background. Below it: [product], [angle], [lighting]. Leave the bottom fifth empty for the price and logo. No other text in the image.

    Where it goes wrong

    • On the free plans of Ideogram and Recraft, what you generate is public. Make unreleased campaign images on a plan that keeps them private.
    • One wrong letter is easy to miss at full size. Read every word at the size the ad runs, on a phone.

    From the official docs: Pricing and plans - Recraft · Pricing — Ideogram · Plans and Credits

  4. 04

    Research customers and competitors, every claim with its source

    Asking a chatbot what customers want or what a competitor charges gets a plausible answer you cannot check. Run deep research that names the sources that count, ask for a link and date beside every claim, and open the sources of the claims you will act on.

    Example

    Before relaunching its pricing page, a project-management app wants to know how three competitors price and what their users complain about. The plan you approve uses the competitors' own pages, public reviews and your survey answers in Drive. The claims table shows that one competitor's price comes from a blog post two years old, so you check their pricing page before it goes into the comparison.

    How to do it

    1. 1Write down the decision the research is for and which sources count: competitors' own pages for prices and features, public reviews for what their customers say, your survey or interview notes for your own.
    2. 2Start Deep Research in Gemini and, if your company allows it, click Sources to add Drive; or open Researcher in Microsoft Copilot and answer its follow-up questions.
    3. 3Review Gemini's research plan before you start it, and check that it looks where your sources are.
    4. 4Ask for the claims table in the prompt on this page: each claim, who says it, the link and the date.
    5. 5Open the source of every claim you will act on or publish, and check competitors' prices on their own page the day you use them.

    Built in

    Prompt to copy

    Research [competitors or customer group] for this decision: [decision], by [date].
Use these sources: [competitors' own websites, public reviews on named sites, our files]. Treat what a company says about itself as a claim, not as evidence.
Return a table: claim, who says it (the company itself, its customers, a third party), link, date of the source.
Keep prices and features apart from opinions. Write "no source found" rather than fill a gap.

    Where it goes wrong

    • A few loud reviews can look like a trend. Ask for the quotes behind each point and count them yourself.
    • A comparison you publish has to hold on the day it runs. Keep a dated copy of each competitor page you rely on.
    • Survey answers and support tickets contain names and emails. Remove them before you upload, unless your company has approved the tool for that data.
  5. 05

    Make sense of a campaign's numbers in a spreadsheet

    Paste exports from several ad platforms, ask how the campaign did, and you get totals that add up conversions each platform counts its own way. Put the numbers in one table with a definition for each column first, then ask what changed and why, with the formula behind every figure.

    Example

    A two-week launch with search ads, social ads and a newsletter. Added together, the platforms report 410 sign-ups; your CRM shows 260. The table keeps platform conversions and CRM sign-ups in separate columns, cost per sign-up is a formula on the CRM count, and the breakdown shows cost per sign-up rose because fewer people clicked, not because the ads got more expensive.

    How to do it

    1. 1Export every channel for the same dates into one sheet with the same columns: date, channel, spend, impressions, clicks and conversions as each platform reports them.
    2. 2Add your own count of sign-ups or sales, from your CRM or shop, as a separate column, and write one line per column on what it counts.
    3. 3In Excel, ask Copilot in the sheet itself with the prompt on this page. Or turn on Work in ChatGPT, attach the exports and ask for a spreadsheet to review, with the formula behind each number.
    4. 4Check one number against the platform itself, such as the spend of the largest channel.
    5. 5To see why cost per result changed, split it into cost per 1,000 impressions, click-through rate and conversion rate, and look at which one moved.

    Built in

    Prompt to copy

    This sheet has one row per channel per day for [campaign], [dates]. The tab [name] says what each column counts.
First describe the data: rows, channels, dates covered, empty cells, duplicates and anything that looks wrong.
Then, per channel and in total: spend, results and cost per result. Use [our CRM column] for totals, not the sum of platform conversions.
Explain the change against [previous period] by splitting cost per result into cost per 1,000 impressions, click-through rate and conversion rate.
Show the formula for every number, and say how many results each conclusion rests on.

    Where it goes wrong

    • Adding up the platforms' conversions counts some customers twice. Use your own sales or sign-up count for totals.
    • Small numbers move by chance. Before you shift budget, check how many results a difference rests on; 12 against 18 is not yet a trend.
  6. 06

    Hand the repeated publishing steps to an automation

    Most teams either copy every post by hand into the newsletter, the social tools and the calendar, or let an agent do it all unchecked. Write down the steps you repeat, let an automation do the ones that are the same every time, and keep a person's approval before anything goes public.

    Example

    Every blog post goes through the same five steps: a newsletter draft, three social posts, tracking links, a message to the team and a line in the content calendar. The automation starts when the post is marked Ready in the calendar, writes the social posts with an AI step and saves everything as drafts; the editor reads them and publishes.

    How to do it

    1. 1Do the steps once by hand and write each one down: what starts it, which app, which fields, and what is the same every time.
    2. 2Describe the list, with the prompt on this page, to Zapier Copilot in a new Zap, to Maia in Make's Scenario Builder, or to Create with Copilot in Power Automate, and review what it builds.
    3. 3Where a step needs writing, such as social posts from the article, add an AI by Zapier step, tell it to use only facts from the article, and click Preview to test it.
    4. 4Make the last step save drafts, not posts, and test the whole automation on an old article first.
    5. 5Turn it on (in Zapier, Publish turns the Zap on) and check the first runs yourself.

    Built in

    Prompt to copy

    Build an automation. Trigger: when an item in [content calendar app] changes to [Ready].
Then, in this order:
1. Create a draft in [newsletter tool] with the title, summary and link.
2. Write [3] social posts from the article with an AI step, using only facts from the article.
3. Save the posts as drafts in [scheduling tool].
4. Send [person] a message in [chat app] with links to the drafts.
Do not publish anything. Ask before you add a step I did not list, and tell me which account each step connects to.

    Where it goes wrong

    • An automation repeats a mistake on every run. Keep it on drafts until its runs have been right for a few weeks.
    • It runs with the access of whoever connected the apps. Connect it with an account the team controls, and write down who owns each automation.

Rules for marketing

  • Do not publish a claim you cannot prove: prices, comparisons, "fastest", health and environmental claims.
  • Paste customer names, emails and order histories only into tools your company has approved for that data.
  • Use a real person's face, voice or words only with their consent, and label AI-generated content where the platform or local rules ask for it.
  • Check what your plan allows before an AI image, video or track goes into a paid campaign.

Go deeper

Esc
↑ ↓ to moveEnter to open