How to Brief AI to Write Ad Copy That Doesn't Sound Like AI
Published 2026-06-09
You ask ChatGPT for an ad. It gives you something like: "Unlock the ultimate coffee experience. Elevate your mornings with our handcrafted blends." Technically an ad. Reads like every other ad the internet has ever produced. Nobody clicks it, nobody remembers it, and worst of all, it sounds nothing like your business.
Here's the thing most people get wrong: the model isn't broken. The brief is. AI ad copy sounds like AI because the prompt behind it could describe ten thousand other businesses. Fix the inputs and the structure of the brief, and the output changes dramatically. This article walks through the full process — what to gather before you prompt, how to structure the brief itself (with a complete worked example), how to generate variations that are actually different from each other, and the edit pass that removes the last traces of AI flavor.
Why default AI ad copy all sounds the same
Large language models are trained on enormous amounts of text, and a meaningful chunk of that text is marketing copy. Not the best marketing copy — all of it. Landing pages, press releases, LinkedIn posts, brochure filler. When you ask for "an ad for a coffee shop" with no further detail, the model does the statistically sensible thing: it returns the average of everything it has seen labeled "ad for a coffee shop."
The average of all marketing copy is, by definition, generic. That's where "Unlock," "Elevate," and "Discover the difference" come from. They're not bugs; they're the center of the distribution.
This leads to the single most useful mental model in prompt engineering for copywriting:
An underspecified brief returns the average. A specific brief forces the model away from the average.
Every concrete detail you add — a real customer quote, a named objection, an actual price — narrows the space of possible outputs and pushes the model toward something only your business could say. The work, then, isn't in clever prompt phrasing. It's in gathering the right raw material first.
Gather these four inputs before you write a single prompt
Professional copywriters don't start by writing. They start by collecting. Do the same before you open the chat window. Four things, fifteen minutes, and most of them you already have.
1. Real customer quotes
Pull three to five verbatim quotes from reviews, support emails, DMs, or conversations. Verbatim matters — don't paraphrase. The way customers actually describe your product ("I didn't know coffee could taste like this without sugar") is almost always better ad language than anything you or the model would invent. It carries specificity, rhythm, and proof.
If you don't have reviews yet, ask three customers one question: "What almost stopped you from buying, and what changed your mind?" The answers are an ad brief in disguise.
2. The one objection to overcome
Every ad works best when it's aimed at a single hesitation. Not five. One.
- "It's probably too expensive."
- "I won't actually use the membership."
- "Specialty coffee is pretentious."
Pick the objection your best-fit customers actually voice, and make defusing it the job of the ad. An ad that tries to answer every objection answers none of them — and forces the model back into benefit-soup mode.
3. The specific offer
"Check us out" is not an offer. The model can't write a sharp call to action around a vague one. Nail down: what exactly the reader gets, what it costs (or what's free), what the constraint is (this week, first 50 people, in-store only), and what the single next step is. A concrete offer gives the copy a spine.
4. Voice samples
Find two or three pieces of text that already sound like you — an Instagram caption you were proud of, an email a customer replied "ha, love this" to, even a sign in your shop window. Paste them into the brief as voice reference. This works far better than adjectives. "Write in a witty, warm tone" produces a model's idea of witty and warm, which is its own kind of generic. Three real samples produce your witty and warm. The persona adoption lesson goes deeper on why showing beats describing here.
The brief structure, with a full worked example
With the inputs gathered, the brief almost writes itself. Here's a complete one for a fictional-but-realistic business: a neighborhood coffee roaster running Instagram ads to people within a couple of miles of the shop.
You are writing Instagram ad copy for Hartline Coffee Roasters, a
small-batch roastery in the Riverside neighborhood. We roast on-site
twice a week and sell bags over the counter and via local pickup.
AUDIENCE: People who live or work within 2 miles. They already drink
coffee daily, usually grocery-store beans or a chain. They are not
coffee nerds and are slightly put off by specialty-coffee snobbery.
THE ONE OBJECTION: "Fancy local coffee is overpriced and pretentious —
my normal coffee is fine."
OFFER: Free 4 oz sample bag with any drink order, this week only.
Mention "the sample thing" at the counter. No signup, no app.
REAL CUSTOMER QUOTES (use the language, not the quotes verbatim):
- "I genuinely didn't know coffee could taste like this without sugar."
- "It's like two dollars more than the supermarket and it's not close."
- "I came for the dog in the window, I stayed for the Kenya."
VOICE SAMPLES (match this register):
- "We roasted too much Guatemala this week. Our mistake, your Tuesday."
- "No, you don't need to know what 'washed process' means. It just
means it's good."
CONSTRAINTS:
- Primary text under 125 characters, headline under 40.
- No words from this list: unlock, elevate, discover, indulge, artisan,
curated, experience (as a noun), journey.
- No rhetorical questions as openers. No exclamation marks.
- Mention the free sample bag explicitly.
TASK: Write 10 headline + primary text pairs. Each pair must take a
different angle (price comparison, anti-snobbery, neighborhood
identity, the roasting schedule, the free sample itself, etc.). Label
each angle. Do not rank them or pick a favorite.
Worth pausing on what each block actually does:
- The opening line sets the format (Instagram ad) and grounds the model in a real, specific business — not "a coffee shop."
- AUDIENCE does double duty: it defines who's reading and preemptively bans the snobbery register the model would otherwise reach for, because "specialty coffee" is statistically tangled up with pretentious language in its training data.
- THE ONE OBJECTION gives every output a job. Copy that has a job to do is automatically less mushy than copy that's just "promoting."
- OFFER makes the call to action concrete. "Mention 'the sample thing'" is a detail no model invents — and it's the most ownable line in the brief.
- CUSTOMER QUOTES hand the model proof and phrasing it could never hallucinate. Note the instruction: use the language, not the quotes verbatim. You want the texture, not fake testimonials.
- VOICE SAMPLES anchor tone by example. Two short lines outperform a paragraph of tone adjectives every time.
- CONSTRAINTS with a banned-word list is the highest-leverage line in the whole brief. Telling the model what it can't say blocks the average outputs at the door. The anti-cliche lesson covers how to build a banned list for your own niche.
- TASK asks for ten labeled angles, not "your best ad." More on why next.
If this feels like a lot, it compresses well once you've done it twice — the 5-line brief template is the short-form version of this exact structure for everyday tasks.
Ask for angles, not "the best one"
A common failure mode: asking for "the best headline" or "3 great options." When you ask for the best, the model gives you its highest-probability answer — which is the most average one. When you ask for three, you usually get the same idea in three outfits.
Instead, ask for ten outputs that must differ by angle, and name the angles or make the model label them. For the roaster, ten angles might include:
- Price reframe — "Two dollars more than the supermarket. Not close in any other way."
- Anti-snobbery — "No tasting notes will be read aloud at you."
- Neighborhood — "Roasted four blocks from your kitchen."
- The schedule — "We roast Tuesdays and Fridays. You can smell which days."
- The offer itself — "Free sample bag with any drink. Just say 'the sample thing.'"
Most of the ten will be mediocre. That's fine — that's how human copywriting works too. You're not asking the model to be the copywriter; you're asking it to be the world's fastest first-draft machine, so you can be the editor who spots the two lines worth keeping. Often the winner is a Frankenstein: the angle from #3 with the closing phrase from #8.
If the first batch clusters around the same two ideas, push: "Angles 1–10 all lead with taste. Give me 10 more that never mention flavor." That single follow-up usually produces the most interesting material of the session.
The edit pass: AI tells to cut
Even with a strong brief, AI-generated copy carries fingerprints. Practitioners who read a lot of model output learn to spot them instantly — and so, increasingly, do regular readers scrolling past your ad. Here's the checklist to run on anything before it ships.
Vocabulary tells
Certain words are so over-represented in AI output that they function as a watermark. Cut or replace on sight:
- Unlock, elevate, empower, supercharge — verb-shaped hype
- Seamless, effortless, hassle-free — frictionless-everything filler
- Game-changer, next-level, revolutionize — claims with no evidence attached
- Discover, indulge, savor, embark — brochure verbs
- Curated, artisan, bespoke (unless you literally curate things)
- "Whether you're X or Y..." — the model's favorite way to address everyone and therefore no one
You don't need to memorize the list. The test: would a real person say this sentence out loud to a friend about your product? "You'll unlock a seamless coffee experience" fails instantly.
Structural tells
- Em-dash overload. Models adore em-dashes — really adore them — often several per paragraph. One per ad, maximum. (Yes, this article is pushing its own budget.)
- Triple parallel constructions. "Bold flavor. Local roots. Zero pretension." The rule of three is a real rhetorical device, but models deploy it constantly, and three of these in one ad set reads as machine-made. Keep one if it's genuinely good; vary the rest.
- Benefit soup. Sentences that chain three or four benefits with commas: "fresh, flavorful, locally roasted, and affordable." Each added adjective makes every other adjective less believable. Pick the one benefit that answers your chosen objection and let it stand alone.
- The summarizing closer. AI loves to end by restating everything: "So if you want great coffee, community, and value, visit Hartline today." Real ads end on the offer or a single sharp image, not a recap.
- Hedging and throat-clearing. "Looking for a better cup?" openers, "In today's fast-paced world" setups. Delete the first sentence of most AI drafts and they improve.
The read-aloud test
Final filter: read the copy out loud, ideally in the voice of the actual person who'd say it — the owner at the counter, not a narrator. Anywhere you stumble or cringe is a line the model wrote and you haven't yet made yours. Rewrite those spots by hand. This last 10% of human editing is what makes the other 90% of AI drafting invisible.
Quick checklist
Before you prompt:
- 3–5 verbatim customer quotes collected
- One objection chosen (only one)
- Offer made concrete: what, price, constraint, next step
- 2–3 voice samples that already sound like you
In the brief:
- Specific audience, including what they're skeptical of
- Banned-word list (start with: unlock, elevate, seamless, discover, game-changer)
- Format constraints (character limits, no rhetorical-question openers)
- Ask for 10 labeled angles, not "the best one"
After generation:
- Cut vocabulary tells; run the say-it-out-loud test
- Max one em-dash, max one rule-of-three
- One benefit per ad, not benefit soup
- Delete summarizing closers and throat-clearing openers
- Rewrite the weakest 10% by hand
If you want ready-made starting points to adapt, the AI prompts for marketers collection covers ads, emails, and landing pages using this same brief-first approach.
Practice the skill, not just the prompt
Briefing is a skill, and skills improve with reps, not with reading. The structure above transfers to almost any writing task once it's in your fingers — try it on something low-stakes first, like the rewrite a social bio challenge, or the trickier voice-matching workout of a dating app bio.
And if you want a daily rep: today's challenge is one real-life briefing task, your submission scored by an AI coach in seconds, with feedback on exactly what your brief was missing. One brief a day, free, takes about five minutes. The first time you jump from three stars to five, you'll feel the difference in every prompt you write afterward.