AI ad copy that does not sound like AI

Why AI copy sounds generic

A language model predicts likely wording. If you ask for an ad for a moisturiser, the likeliest wording is what thousands of moisturiser ads have already said: radiant, glowing, unlock, elevate, transform. None of those words is wrong, but none says anything about your product, and readers have learned to skim past them.

The cause is mostly in the input. A short prompt such as "write a Facebook ad for my candles" leaves every important decision to the model, and the model's defaults are the average of the internet. The fix is to supply the decisions yourself: who the buyer is, what happened in their day that led to the purchase, what the product actually does, and what you would never say.

Start with a copy brief

A usable copy brief fits on half a page. Include these elements.

• Product facts: materials, sizes, quantities, how it is used, what it costs, what is included.
• The customer: one specific person and situation, not "women aged 25 to 54".
• The problem in their words: phrases from reviews, emails and support tickets, anonymised.
• The promise you can support: a statement you could defend if asked for evidence.
• Proof you hold: real review counts, real certifications, real guarantees, stated accurately.
• Voice: two or three sentences of your own writing, plus words you never use.
• Constraints: placement, character limits, mandatory wording, restricted claims.

The research buyer language for ads and turn reviews into ad angles guides explain how to collect the third item. It is the single most effective way to avoid generic output, because real customer phrasing is not what the average of the internet sounds like.

Ask for angles before sentences

Resist asking for finished ads straight away. First request ten angles, each as a single plain sentence describing the reason to care. Read them, cross out the weak ones, and only then ask for copy on the best three. This keeps the creative decision with you and uses the model for what it does best, which is breadth.

Before and after: a hypothetical example

Imagine a store selling a compact travel kettle. The first prompt is "write a Facebook ad for a travel kettle". A typical draft might read: "Unlock the joy of tea anywhere. Our premium travel kettle elevates every journey. Shop now."

Now improve the brief. The kettle folds flat, holds 500 millilitres, has dual voltage, and the founder's customers mention hotel rooms that only have a coffee machine. The revised draft might read: "Hotel room, no kettle, jet lag. This one folds flat, boils 500 ml and works on both voltages, so your own tea is ready in your own mug. Check the voltage rating for your destination."

The second version is not literary, but it is specific, it uses a situation the customer recognises, and it includes a responsible note. Notice also that it contains no claim about speed that the founder has not measured. If the founder had timed the boil, a figure could go in, provided the conditions were stated.

Phrases to ban

Maintain a short list and paste it into every prompt. Common candidates include: unlock, elevate, revolutionise, game-changer, seamless, effortless, in today's fast-paced world, take it to the next level, and any superlative you cannot prove. Ask the model to avoid them, then search the output yourself, because models sometimes ignore instructions.

Also watch structure. Models love tidy triplets, rhetorical questions that answer themselves, and openings such as "Are you tired of". If three drafts in a row start the same way, change the instruction, not just the wording.

Match the placement

Different placements want different shapes. A short caption on a vertical video needs a first line that works with the sound off. A feed ad with a static image can carry a little more explanation. Search ads are constrained: Google's help documentation for responsive search ads states headlines of up to 30 characters and descriptions of up to 90 characters, and notes that you provide multiple of each so that combinations can be assembled. Always confirm current limits in the official help pages, as they can change.

Use the ad character counter to check lengths, and the ad preview mockup to see how text is truncated in context. For Meta placements, check the Meta Ads Guide (opens in a new tab) rather than assuming limits from memory.

The edit pass

Edit in four passes, in this order.

• Truth: Underline every factual statement and confirm it. Remove or soften anything you cannot support.
• Specificity: Replace general nouns with concrete ones. "Great quality" becomes the material and the thickness.
• Voice: Read aloud. If you would never say a sentence to a customer, rewrite it.
• Length: Cut the first sentence if it only warms up. Most drafts improve when the opening is removed.

The ad copy grader can help you spot unclear or overlong wording, but it is a prompt for your own judgement and not an approval.

Hooks and calls to action

The first line decides whether anything else is read. Use the ad hooks guide to generate openings by angle, then ask the model to adapt them to your product. For the closing line, write clear ad CTAs explains why a plain next step usually beats a clever one. Ask the model for five call-to-action options and choose the one that matches the landing page, since a mismatch between ad and page is a common source of wasted clicks.

Claims, regulation and tone

Some categories carry extra risk. Health, skincare, supplements, finance and anything involving before-and-after results attract scrutiny from regulators and platforms. AI tools do not know the rules for your market and may happily write a promise that would be a problem. Keep wording factual, avoid implying that a product treats or cures a condition unless you are authorised to say so, and consult the relevant advertising standards for your country. If in doubt, ask a qualified adviser. This guide is general information and not legal advice.

Never ask a model to write customer testimonials. Fabricated reviews mislead buyers and break platform and consumer rules. Use real, permissioned reviews, quoted accurately, or none at all.

Build a reusable prompt

Once you have a brief that works, save it as a template with blanks. A simple structure is: role (you are helping an ecommerce founder write ad copy), product facts, customer situation, customer phrases, voice sample, banned phrases, placement and limits, and the output you want, such as ten angles as one sentence each. Update the facts and phrases for each product, and keep a record of which versions produced drafts you actually published. Over a month you will learn which inputs matter most for your store.

Measure, but carefully

Copy tests are easy to muddy. If you change the hook, the body and the image at once, you cannot tell which change mattered. Change one element at a time where you can, and compare on a metric tied to profit rather than clicks alone. The AI ad creative testing guide covers structure, and the AI ads pillar places copy within the whole workflow. Remember that platform attribution has limits, and small samples mislead, so avoid declaring a winning phrase after a few orders.

A worked editing example

Take a first draft for an imaginary stainless steel lunch box: "Elevate your lunch with our premium, game-changing lunch box. Shop now and transform your meals." Run the four passes. Truth: nothing is checkable, so "premium" and "game-changing" go. Specificity: the founder adds that the box has two compartments, a lid with a silicone seal and fits a standard laptop bag. Voice: the founder would say "keeps the rice away from the pickle". Length: the warm-up sentence goes.

The result might read: "Two compartments, so the rice stays away from the pickle. Silicone-sealed lid, fits flat in a laptop bag. Dishwasher use is fine for the box, but check the lid care label." The last sentence only appears if it is true, and it shows how a short practical caution can build trust. The point of the exercise is that each edit replaced a general word with a fact the founder could defend, which is the main difference between copy that reads as generated and copy that reads as written.

Keep a swipe file of your own voice

Models imitate what they are shown. Keep a document of ten to twenty pieces of your own writing that sound like you: product page lines, replies to customers, packaging inserts, emails. Paste three or four into the prompt as voice samples and name the traits you want, for example plain, a little dry, no exclamation marks. Update the file when you write something you are proud of. After a few weeks the prompt carries a recognisable voice, and edits get quicker because the first draft lands closer to your own register. Also keep a list of corrections you make repeatedly. If you always delete the first sentence or replace the same phrase, add that rule to the prompt so the model stops making the same mistake. This small habit does more for quality than switching between tools.

Input → raw draft → edited draft

Illustrative worked example, not a measured customer result. Replace assumptions with checked facts.

Common questions

How do I make AI copy sound like my brand?

Give the model a sample of your own writing, a list of words you never use, and real customer phrases. Then edit the output aloud until it sounds like something you would say.

Can I publish AI-written copy without editing?

You can, but it is not advisable. Check every claim, limit and tone issue first. Unedited output is the main reason AI copy reads as generic.

What character limits apply to Google search ads?

Google's help documentation for responsive search ads gives 30 characters for headlines and 90 for descriptions. Confirm against the current help page before writing.

Should I tell the AI to write like a famous copywriter?

It rarely helps. Specific facts and real customer language help far more than imitation of a style.

Is it acceptable to generate reviews or testimonials with AI?

No. Invented testimonials mislead buyers and can breach platform and consumer protection rules. Use real, permissioned feedback.

How many versions should I generate?

Generate plenty of angles, but publish only a few of your best edited drafts. Testing needs enough spend per version to show a difference.