What marketing teams use it for
The jobs below are the ones where AI saves a marketing team real hours without putting the brand at risk, as long as the inputs are facts you can stand behind. Each one links the tools and templates that fit.
Search and social ad variants
A Google responsive search ad takes up to 15 headlines of 30 characters and 4 descriptions of 90, and a model counting characters by eye will overrun them. The ad variants template asks for ten options that between them cover the main benefit, the problem, a specific feature, a proof point and the audience, with a character count beside each headline and description.
Make the limit a test, not a hope, because a model’s own character count can be wrong. Change the output to one headline per line starting with “H:”, and a Matches a pattern check in Prompt Tester with the pattern ^H: .{31,}$, the m flag, and At least and At most both set to 0 fails any output with a headline over 30 characters, so you find the overruns before the ad editor does.
SEO titles and meta descriptions in batches
A site migration or a content refresh leaves two hundred pages that need title tags and meta descriptions. The SEO template writes each one from the page’s own text, so the description matches what the page says rather than what you wish it said.
Google cuts snippets by pixel width, not a fixed count, so most teams settle on a house limit of roughly 50 to 60 characters for titles and 150 to 160 for descriptions. A Length check in characters enforces it across the whole batch, and Version Compare shows what changed when someone edits the prompt before the next batch.
SEO titles and meta descriptions Prompt Tester Version Compare
Landing pages and launch copy from a fact sheet
Write the fact sheet before the copy: what the product does, for whom, the numbers you can prove, the claims legal has cleared and the ones it has not. The landing page template builds from that sheet and nothing else, and the press release template does the same for announcements, with approved quotes only.
Then put the draft and the fact sheet through the claims checker. It lists every claim in the copy with the line of the fact sheet that supports it, or flags it as unsupported, which is the list your reviewer would have built by hand.
Landing page copy Press release from facts Check claims against sources
One article, several channels
A good article can carry a newsletter section, three LinkedIn posts and a short thread, as long as each piece carries one real point from it rather than a summary of all of them. The repurposing template works that way, and it quotes the source sentence behind each post so an editor can check it in seconds.
Save your voice notes and banned-phrase list as one versioned prompt in My Prompts, so the whole team drafts from the same version and any change to it can be compared word by word.
Repurpose an article Newsletter from notes LinkedIn post from a project My Prompts
Customer evidence: case studies and survey themes
Case studies go wrong when the draft improves on what the customer said. The case study template works from the interview transcript, keeps every quote verbatim and lists the facts to confirm with the customer before publishing.
For NPS or post-purchase surveys with hundreds of free-text answers, the survey themes template returns themes with real counts and exact quotes, so the slide says how many people said something rather than how strongly one person did.
Case study from an interview Survey response themes Competitor comparison
Images and short video
Image and video models read prompts differently from chat models: they want the subject, the setting, the light and the framing, not instructions. The image and video templates are laid out for those models, and the background edit template keeps the product itself unchanged, which matters when the picture goes on a listing or an ad.
The Prompt Grader scores image and video prompts too. Pick the model under Written for, and it shows which details the prompt leaves to chance.
Header image with text Product photo background Product video shot Prompt Grader
Tests and campaign readouts
Write the A/B test plan before launch: the hypothesis, the primary metric, the sample you need and the rule for calling it, agreed in writing so nobody moves the goalposts after one good Tuesday. The readout template reports what the numbers show, how sure you can be and what they do not show.
When a channel metric moves, the metric change template separates what the data supports from the story someone would like to tell in the Monday meeting.
A worked example: launch week for a feature
A hypothetical: a B2B software company is launching a scheduling feature, and the product marketer owns the landing page, the launch email, a set of search ads and the social posts. The work runs through six steps, and each one has a check.
- 1Write the fact sheet, then the pageThe landing page template drafts only from what the sheet says.
- 2Score the briefThe free Grader estimate flags a missing audience, length or format.
- 3Test the ad prompt on past launchesCharacter limits, banned phrases and the product name, checked on every output.
- 4Pick the model by blind voteTwo marketers choose the best drafts before the model names appear.
- 5Check every claimEach claim in the copy is matched to a line of the fact sheet, or flagged.
- 6Keep the voice prompt versionedSave the prompt that passed, so the next launch starts from it.
The fact sheet does most of the work
The marketer writes one page: what the feature does in a sentence, the three jobs it does for the customer, the plans it is on, the numbers they can prove (with the source of each), the claims legal has cleared and two that it has not. Every later prompt pastes this sheet. Copy built on it is specific because the sheet is specific, and the claims check at the end has something to check against.
Testing the ad prompt
In Prompt Tester, the eval uses the fact sheets of the last eight launches as test cases, which the team already knows well. Four checks run on every output: the headline pattern above, a Does not contain check for the banned list (“best-in-class”, “world-class”, “revolutionary”), a Contains text check for the product’s name, and an AI judge with one narrow criterion, “every claim in the ads is supported by the fact sheet”. When a version passes all eight, it is ready for this launch.
Choosing the drafts
Model Compare runs the landing page prompt on three models side by side. With Blind vote on, two marketers pick the best draft before the names appear, which keeps the choice about the copy rather than the brand of the model. If the cheaper model’s draft wins, or loses narrowly, that settles which one the team uses for the rest of the quarter.
What it costs
Assume a five-person marketing team making about 60 model calls a working day across ads, SEO batches, repurposing and drafts. Each call sends around 2,500 tokens (a fact sheet or an article section, plus the voice guide and instructions) and gets about 700 back. Half of each prompt is the same voice guide and banned list, read from the provider’s cache. Twenty-two working days a month.
Monthly cost of 1,320 calls with about 2,500 tokens in and 700 out each, 50% of the input read from cache. List prices checked October 20, 2026; your bill may differ.
At this volume the model bill is small on every tier, and the cost that matters is an editor’s time on drafts that miss. Spend the difference on testing: the model that wins a blind vote on your own launches is worth more than the one that is cheapest in the chart. Image and video generation is priced on different terms, usually per image or per second of video, and is not included here.
A prompt to start from
A launch email brief that keeps the model to the fact sheet and in your voice. Paste it into any chat model, or into the Grader first to see what it is missing for your case.
You write launch copy for {{company}}. The readers are {{audience}}.
Use only the facts in the fact sheet. If a claim you want to make is not
in it, leave it out and list it at the end under "Claims I left out".
Fact sheet:
"""
{{fact_sheet}}
"""
Three samples of our writing, for rhythm and word choice only. Do not
reuse their content:
"""
{{voice_samples}}
"""
Never use these words or phrases: {{banned_phrases}}
Write:
1. Five subject lines, each under 50 characters.
2. One launch email of at most 150 words. Open with what the reader can
do now that they could not do last week. One call to action:
{{call_to_action}}.
After the email, list each factual claim it makes and the line of the
fact sheet that supports it.
Guardrails for marketing
- Have the evidence before the claim runs. In the US, the FTC expects advertisers to be able to back up an objective claim before the ad appears, and its advertising guide for small businesses explains the standard. A model will happily write “the fastest on the market”. The fact sheet and the claims check are how you keep it out.
- No invented reviews, testimonials or quotes. The FTC’s 2024 rule on fake reviews and testimonials covers AI-generated ones explicitly. Quotes come from real people, word for word, and are approved by them.
- Keep customer data out of drafting prompts. A CRM export with names and emails does not belong in a tool your company has not approved for personal data. Aggregate or remove identifiers before you paste survey answers or lists.
- Respect likeness and marks in images. Do not generate real people, a competitor’s logo or another brand’s product. Some platforms require a label on realistic AI-generated imagery, particularly in political and social issue ads, so check the current policy for each one you buy on.
- Publish pages a person would stand behind. Google’s spam policies target pages generated at scale mainly to game rankings, whoever or whatever wrote them. Each published page gets an editor, a fact check and something only your team knows.
Key takeaways
- Write the fact sheet first. Every prompt pastes it, and the claims check needs it.
- Turn platform limits into checks: a pattern check catches a 31-character headline before the ad editor does.
- Give the model samples of your writing and a banned-phrase list, and keep both in one versioned prompt.
- Choose drafts and models by blind vote on your own past launches, not by the model’s name.
- No invented testimonials, no unsupported claims, no real people in generated images.
Related templates, guides and posts
Questions
Does Google penalize AI-written content?
Google says it ranks content on how helpful it is, however it was produced, and its spam policies target pages generated at scale mainly to manipulate rankings. In practice the risk is publishing many thin pages, which is a problem whether a person or a model wrote them. Edit every page, add what only your team knows and check the facts.
How do we stop AI copy from sounding like AI?
Give the model something specific to say and a voice to say it in: a fact sheet, two or three samples of your own writing, a list of phrases you never use and a request for several variants to choose from. Generic copy is usually the result of a generic brief, not a weak model.
Can we use AI to write testimonials or reviews?
No. A testimonial has to come from a real customer and reflect what they actually said, and in the US the FTC’s 2024 rule on fake reviews covers ones generated by AI. AI can help you draft the request to a customer, or tidy a quote that the customer then approves, but it cannot supply one.