The PPC Manager's Guide to AI Ad Creative

The short answer

AI does not make ad creative good. It makes drafting fast. The quality of AI-assisted creative is set by three things a PPC manager controls: the brief the model works from, the range of angles you ask for, and the review gate that stops weak or risky assets going live. Get those right and AI lets you test far more ideas per month. Get them wrong and you ship forty versions of the same bland promise.

Step 1: write a brief the model cannot misread

Most generic AI copy comes from a thin prompt. Build one brief per offer and reuse it. It should contain:

• • Offer: what is sold, the price or pricing model if public, what happens after the click.
• • Audience: who buys, what triggers the search or scroll, what they already know.
• • Proof library: facts you can substantiate, each with its source (reviews count and platform, awards with year, test results, guarantees you actually offer in writing). The model may only use claims from this list.
• • Banned claims: superlatives you cannot prove, competitor names, regulated claims for your sector, anything legal has ruled out.
• • Mandatory items: trademark formatting, required disclaimers, geographic qualifiers.
• • Voice: three adjectives and two example lines that sound right, and one that sounds wrong.
• • Limits: the character limits of the format you are writing for.

A proof library is the single biggest improvement. Without it, models fill gaps with plausible claims such as "trusted by thousands", which may be untrue and may breach advertising rules.

Step 2: generate by angle, not by synonym

Ask for variety in the reason to act, not in wording. An angle matrix makes that explicit:

Illustrative angle matrix for a hypothetical accountancy firm offering fixed-fee bookkeeping

Angle	Buyer question it answers	Search headline draft	Social opening line draft
Price certainty	What will this cost me?	Fixed Monthly Bookkeeping Fee	Know your bookkeeping bill before the month starts.
Time saved	What do I get back?	Hand Over Your Receipts	Stop spending Sunday evenings on receipts.
Risk reduction	Will I get fined?	VAT Returns Filed On Time	Missed a VAT deadline once? It is avoidable.
Specific use case	Is this for businesses like mine?	Bookkeeping for Trades	Plumbers, electricians, builders: bookkeeping built around jobs, not spreadsheets.
Objection	Is switching painful?	We Move Your Books Across	Switching accountant takes less work than you think.
Proof	Why believe you?	ICAEW Chartered Accountants	Chartered accountants, not a chatbot.

Six angles with two or three drafts each gives the platform real choices. Twenty headlines on one angle does not.

Step 3: fit each format

Each platform has different units and limits, and they change. Keep a current table in the brief and check the help pages when building:

Creative units by channel and where to confirm limits

Channel	Creative unit	Key limits to respect	Where to confirm
Google Search	Responsive search ad	Up to 15 headlines of 30 characters and 4 descriptions of 90 characters; at least 3 headlines and 2 descriptions	Google Ads Help: About responsive search ads
Performance Max	Asset group	Headlines, long headlines, descriptions, images, logos and video, each with its own counts and limits	Google Ads Help: Performance Max specs and format requirements
Meta	Ad with primary text, headline, media	Text is truncated at different points by placement; Advantage+ creative may alter media	Meta Ads Manager previews and Meta Business Help Center
LinkedIn	Single image, document, video or carousel ad	LinkedIn recommends about 150 characters of introductory text and about 70 for the headline to avoid truncation	LinkedIn single image ad specifications

Ask the model to return a character count for each line, then check it with a formula in your sheet. Models count characters unreliably.

Step 4: the review gate

Every asset passes these checks before it goes live, whoever or whatever wrote it:

1. • Truth. Is every claim in the proof library? Is any price, offer or timescale current?
2. • Standalone sense. In responsive formats, any headline can appear next to any other. Read each alone and in random pairs.
3. • Duplication of meaning. Two assets that say the same thing waste slots and weaken Ad Strength signals.
4. • Policy and law. Platform policies, trademark use, and the UK advertising codes. The Committee of Advertising Practice is clear that the codes apply however an ad is made, and that saying AI was used does not fix an image that misleads (ASA).
5. • Voice. Would the client say this? Delete anything that sounds like every other advertiser in the auction.

If most drafts survive this gate, the gate is too loose. Cutting a large share of AI drafts here is the process working, not failing.

Step 5: test so that you learn something

Platform assembly makes clean A/B tests harder, because the system mixes assets. Practical rules:

• • Test one angle at a time where you can. Use Google Ads experiments or Meta's A/B test tool when a question matters, such as whether price-led copy beats time-saved copy.
• • Judge on conversions, not click-through rate. A curiosity headline can raise CTR and lower conversion rate. For lead generation, judge on qualified leads imported from the CRM.
• • Set a minimum volume before you decide. Decide in advance how many conversions or how much spend an asset needs before it can be judged, so you do not react to noise.
• • Pin sparingly. In responsive search ads, pinning reduces the combinations Google can test and can lower Ad Strength (Google Ads Help). Pin only what must appear, such as a regulated disclaimer.

Step 6: keep a creative log

Record each angle, where it ran, the dates and the result. After three months the log tells you which reasons to buy work for this audience, which is worth more than any single winning ad. Feed the winners back into the brief as the first examples the model sees.

For the Google build process end to end, see how to create Google Ads with AI. For creative aimed at people who have already visited, see the retargeting creative guide.

FAQ

Which AI model is best for ad copy?

The differences between capable models matter far less than the brief. A model given a proof library, banned claims and an angle list will outperform a stronger model given one line of instruction. Pick one you can use consistently and invest in the brief.

Should I use the platforms' own AI text suggestions?

They can add useful variety, but they write from your landing page and existing ads without knowing your banned claims. Where you can review suggestions before they serve, do so; where you cannot, decide whether the feature should be on at all.

How many new angles should I test each month?

Enough that each gets a fair amount of spend. For a small account that may be one or two angles a month; for a large one, several per campaign. Testing more angles than your budget can judge only produces noise.

Can AI write compliant ads for regulated sectors?

It can draft them, but the review gate matters more. Financial, health and legal advertising carry specific rules, and a model will not reliably know them. Keep a sector-specific banned list and have a qualified person approve.