Copywriting for Ads with AI

Why copywriting + AI

Copywriting is writing that drives an action — an ad, a headline, a landing page, an email. What AI adds is abundance: twenty angles in the time it used to take to write one, so you edit from options instead of staring at a blank page.

The bottleneck moved, and most people did not notice

Writing used to be the constraint. Now producing twenty headlines takes ninety seconds, and the constraint is knowing which one is good — a skill that abundance does not teach and in fact makes easier to avoid.

The failure looks like this: you generate twenty options, they all read fine, you pick the one that sounds most like an advert, and it underperforms the plain sentence you would have written yourself. Nothing in the process told you the difference, because the model optimises for plausible copy and plausible copy is exactly what bad advertising is made of.

So the useful skills now are research and selection. Everything below is organised around those two, because the generation part genuinely is solved and it is the least important part.

The best copy is quoted, not written

This is the highest-leverage idea on the page.

A model writes what marketing usually sounds like — it is trained on a world of marketing. Your customers do not talk like marketing. They describe their problem in specific, unglamorous words, and copy that uses their words outperforms copy in the vendor's register, because it reads as recognition rather than persuasion.

Those words already exist and you do not have to invent them:

• Reviews — yours and competitors'. The three-star ones are the most useful; people explain exactly what nearly stopped them buying.
• Support tickets and sales-call notes. The question asked most often is the objection your landing page is failing to answer.
• The email a customer sent when they signed up, saying why. That paragraph is frequently better copy than anything you will write about yourself.
• Forums and communities where your buyers complain about the problem.

Then use the model properly: paste the raw material and ask it to find the recurring phrases and the objections, not to write copy from a topic. Now it is working on evidence from your market instead of recalling the average advert. That single change in input does more for output quality than any prompt technique.

Where the reader is, before what you say

The same product needs entirely different copy depending on how much the reader already knows, and getting this wrong is why technically competent copy falls flat. The classic framing — from Eugene Schwartz's Breakthrough Advertising, and still the most useful thing in the field — is five stages:

• Unaware — they do not know they have the problem. Lead with the problem, or a story. Mentioning your product here loses them.
• Problem-aware — they feel the pain, do not know solutions exist. Lead with the problem, precisely described, and introduce the category.
• Solution-aware — they know the category, not you. Lead with why this approach, and what makes yours different.
• Product-aware — they know you and have not bought. Lead with the objection: price, risk, switching cost, whatever is actually stopping them.
• Most aware — they are ready. Lead with the offer and get out of the way.

Tell the model which stage you are writing for. "Write a landing page for product X" gets you copy for a most-aware reader, which is a small minority of traffic — and that mismatch, rather than the writing, is why the page does not convert.

Frameworks, and what they are for

Give the model a structure and the output tightens considerably:

• AIDA — attention, interest, desire, action. The classic.
• PAS — problem, agitate, solution. Strong for short ads.
• BAB — before, after, bridge. Good for showing transformation.
• The four U's — useful, urgent, unique, ultra-specific. For headlines.

A worked prompt: "Five PAS ads for product X, aimed at a problem-aware reader who has tried Y and found it too slow. Use only these customer phrases: [paste three real quotes]. No statistics. Plain speech, no marketing vocabulary."

Be clear about what a framework does, though. It supplies order, not insight. PAS tells you to agitate the problem; it cannot tell you which problem your buyer actually has at 11pm. Frameworks make a weak idea well-structured, which is a real improvement and is not the same as a good advert.

Headlines and hooks

Most people read the headline and nothing else, so it carries most of the weight. The often-quoted figure — that the headline is worth eighty per cent of your spend — comes from David Ogilvy's advice rather than from a measurement, and it is a good instinct rather than a statistic. Treat it as: this is the part to spend disproportionate time on.

Generate twenty, then select against criteria rather than taste:

• Is the benefit legible in two seconds? Not clever — legible.
• Is it specific? A number, a timeframe, a named situation. "Faster invoicing" is a category; "invoices out the same afternoon" is a picture.
• Could a competitor put their name on it? If yes, it says nothing about you.
• Does it promise only what you deliver? An overclaiming headline buys a click and sells a refund.
• Would you say it out loud to someone? If it would be embarrassing to say, it reads as advertising and gets skipped.

That third test is the one that eliminates most generated headlines, and it is the fastest quality filter available.

Specificity is the whole craft

Most weak copy fails in one identifiable way: it is abstract. Models default to abstraction because abstraction is safe and general, and the fix is mechanical enough that you can ask for it directly.

• "Save time" → "stop rebuilding the same report every Monday"
• "Trusted by professionals" → "used by the finance team at a firm you have heard of, if you have permission to say so"
• "Improve your workflow" → "one place to see which invoices are unpaid"
• "Powerful and flexible" → says nothing in any context; delete

Add to every prompt: no adjectives where a fact would do, and invent no numbers. Then supply the real facts yourself. An unedited model will cheerfully produce "increases productivity by 40%", and a fabricated figure in an advert is not a style problem — in most markets it is an unlawful claim.

The offer beats the copy, every time

Worth saying plainly on a page about writing: no amount of copy rescues a weak offer. If the thing is not wanted, or the price is wrong, or the risk sits entirely with the buyer, better sentences will not fix it — they will just get the bad news to more people faster.

An offer is the whole trade: what they get, what they pay, what happens if it does not work, and what they must do to start. Improving any of those four usually moves results more than a month of rewriting.

• Reduce the risk. A guarantee, a trial, a first step that costs nothing. Most hesitation is fear of wasting money, not doubt about value.
• Shrink the first step. "Book a call" is a large commitment; "see the pricing" is not. Ask for the smallest next action that still moves things forward.
• Make it concrete. "Consulting services" is a category. "A two-hour review of your setup, with a written list of what to fix" is a thing someone can decide about.
• Remove the friction you have stopped noticing — the form with eleven fields, the mandatory phone number, the demo gate in front of pricing.

Before generating another twenty headlines, spend ten minutes asking whether the offer itself can be improved. That is the highest-return edit available, and it is the one a model cannot make for you because it does not know what you can afford to promise.

What you are allowed to say

Copywriting is regulated, and a model has no idea what you can substantiate.

• Every factual claim needs evidence you hold before you publish, not after a complaint. Performance figures, comparisons, "number one" — if asked, could you show the basis?
• Testimonials must be real and represent typical results, or say plainly that they do not.
• Urgency must be true. A countdown that resets, or a "last few places" that is never last, is a misleading practice in most consumer-protection regimes — and it is the specific thing regulators have been active about.
• Prices and terms: the total cost, including what is added at checkout, has to be clear at the point it matters.
• Regulated categories — health, finance, supplements — have their own rules about what may be implied at all.

None of this is legal advice, and the specifics differ by market. The habit that covers most of it: keep a file of the evidence for every claim you run. If you cannot fill it, change the copy rather than hoping.

How to spot copy the model wrote

Readers may not name these, but they register them, and they read as "advert" rather than "person".

• The vocabulary: unlock, elevate, seamless, effortless, game-changing, journey, robust, leverage. A find-and-replace list worth keeping.
• Everything in threes. Three adjectives, three benefits, three clauses. Real speech is lumpier.
• Symmetry. Every bullet the same length, every section the same shape.
• The hedge. "Can help you to potentially improve" — three softeners removing all the force.
• The summary that repeats the section above it.
• Beginning with the topic: "In today's fast-paced world" and its relatives. Delete the first paragraph of most drafts and the copy improves.

The editing pass, in order

1. Cut the first sentence. Usually throat-clearing. Check whether the piece is better starting at sentence two — it normally is.
2. Replace every abstraction with a fact, or delete the line.
3. Remove the vocabulary list above.
4. Check every number and claim against something you hold.
5. Read it aloud. Anything you stumble over, rewrite. Anything you would not say to a customer's face, delete.
6. Check there is exactly one call to action, and that it names what happens next — "see pricing" rather than "learn more".

Writing to an audience

The same product, two audiences, completely different copy. Give the model a real persona: who they are, what specifically hurts, what they have already tried, what they are afraid of, and the words they use. Combine with AI content writing for a consistent brand voice across everything.

A persona built from evidence beats an invented one every time. Two real customers you can describe in detail are worth more than a demographic sketch, because you can check your copy against them: would this person read this sentence and feel understood, or sold to?

Variations & testing, done properly

Generating ten versions is where AI earns its place — but ten versions and a favourite is not a test. Pairs well with UGC ads and social campaigns, where variation volume is the whole model.

• One variable at a time. Ten headlines on the same page tells you about headlines. Ten different pages tells you nothing reusable.
• Test the big things first — the offer, the audience, the main promise. Button colours are where testing programmes go to waste time.
• Enough traffic to mean something. Early leads swing wildly, and calling a winner on the second day is how teams confidently adopt the worse version.
• Measure the outcome, not the click. A headline that wins on clicks and loses on sales has found the wrong people faster.
• Write down what you learned, not just which won. "Specific numbers beat curiosity for this audience" transfers to the next campaign; "version B won" does not.

A landing page, section by section

The most common request is a landing page, and "write me a landing page" produces a generic one. Generating it section by section, each with its own brief, produces something usable — and it lets you keep the structure while changing one part.

1. Headline. What this is and who it is for, legible in two seconds. Not clever.
2. Subhead. The one sentence of detail the headline could not carry — usually how it works or what makes it different.
3. The proof, immediately. A real customer quote, a named logo, a number you can evidence. Before you ask for anything, give a reason to keep reading.
4. The problem, in their words. Straight from your research. If they do not recognise themselves here, nothing below matters.
5. How it works — three steps, concretely. Not features; the shape of what happens.
6. The objections, answered in order. Price, effort to switch, "will this work for my case", what happens if it does not work. Your support inbox already ranked these for you.
7. One call to action, repeated where it is natural, naming what happens next.

Two habits make this page better than the sum of its sections. Write the objections section first — it is the part that does the persuading, and knowing what you are overcoming shapes everything above it. And after the draft, read only the headings in order: if they do not tell a coherent story on their own, a skim-reader gets nothing, and most readers skim.

When not to use it

• You do not know the audience yet. Generated copy for a guessed persona is a confident guess. Go and read twenty reviews first.
• Very short, very high-stakes copy — a tagline, a company name, the one line on the homepage. Worth a person's full attention; the model is useful for exploring options, not for the decision.
• Anything requiring a claim you cannot evidence. The problem is the claim, and no phrasing fixes it.
• When it needs to sound like you specifically — a founder's note, an apology, a personal email. The register is the message.

Common mistakes

• Generic copy. "The best solution" says nothing. Specific benefit, specific reader.
• Skipping the research. The words are already in your reviews and tickets. Writing from imagination when evidence exists is the most common unforced error.
• Writing to the wrong awareness stage. Usually the reason a competent page does not convert.
• Letting the model invent numbers. A fabricated statistic in an advert is a legal problem, not a style one.
• Falling in love with a version. Test it; the market votes and you do not.
• No call to action, or three. One, and name what happens when they click.