Create an Advertisement With AI: Brief, Copy, Visuals and Checks

To create a finished advertisement with AI, write a one-page brief, have AI draft several copy angles, generate or choose visuals, then pass the draft through a claims check, an image-rights check and a platform-size check before a person approves it. The checks are what separate a usable ad from a risky one.

Most free ad-generator pages stop after the copy and the picture. That is the easy half, taking minutes. The half that protects the business, deciding whether every claim can be proved and whether you may use every image, still needs a human with the facts. Below is one ad followed from brief to final files.

Follow me on Instagram@sagnikteaches

The workflow in six gates

Treat the process as six gates in a fixed order. An ad does not move to the next gate until the current one is signed off.

Connect on LinkedInSagnik Bhattacharya

1. Brief: the offer, audience, proof and constraints, written by you.
2. Copy: AI drafts angles, you choose and edit.
3. Visuals: AI or your own photos, with a note of where each came from.
4. Claims check: every factual statement matched to evidence.
5. Rights check: every image, font, logo and music clip cleared.
6. Platform check: sizes, text limits and file weights.

Gate four is the one most people skip. It sits between creative and production on purpose, because it is cheaper to remove a bad claim from a sentence than from a live campaign.

Subscribe on YouTube@sagnikteaches

The main example: a kitchen fitter's free design visit

Everything below uses one illustrative business, a four-person kitchen fitting company that wants enquiries for a spring campaign. The figures are made up to show the method.

• Offer: free in-home design visit, fitted kitchens from a stated starting price.
• Audience: homeowners planning a refit in the next six months.
• Budget: $600 for the month on one platform.
• Goal: 20 booked design visits.

The sum that justifies the budget: if one in three visits becomes a job with a margin of $2,400, then 20 visits give roughly six or seven jobs, about $15,000 in margin against $600 of spend. A realistic first month will fall short of that, but the arithmetic tells you the ceiling and the break-even: the ad pays for itself if it produces a single job from 20 visits.

Gate one: the brief that stops generic output

AI invents whatever you leave blank. Fill in the brief fully and it has little room to invent. This is the brief for the kitchen fitter, ready to paste.

Business: four-person kitchen fitting company.
Offer: free in-home design visit; fitted kitchens from $9,500.
Audience: homeowners aged 35-60 planning a refit within six months.
Proof I can state: 11 years trading; 14 installers' hours logged per kitchen on average;
  workmanship guarantee in writing (5 years).
Proof I cannot state: "best", "cheapest", any star rating, any customer quote.
Tone: plain, practical, no hype, no exclamation marks.
Channel: paid social image ad plus a display banner.
Call to action: Book a free design visit.
Constraints: no before/after claims, no fake customer faces, no competitor names.

The "proof I cannot state" line does more work than any other. It gives the AI a list of banned claims and gives you a checklist to run later. A strong brief takes 15 minutes. If you cannot fill in the proof line, you do not yet have an advertisement, you have an adjective.

Gate two: drafting copy angles

Ask for distinct angles, not synonyms. Three versions of "Get your dream kitchen" are one angle. This prompt forces separation:

Using the brief above, write four ad concepts, each with a different promise:
1. Speed (design visit within a week)
2. Price clarity (starting price stated)
3. Guarantee (written five-year workmanship cover)
4. Planning help (one-page checklist for the visit)
For each give: a headline of up to 27 characters, primary text of 50 to 150 characters,
and a button label. Do not use any claim listed under "Proof I cannot state".
After the concepts, list any statement you made that needs evidence.

Those lengths follow Meta's current recommendation for Facebook Feed image ads, a headline of about 27 characters and primary text of 50 to 150 characters, which I checked in Meta's ads guide on 4 November 2026. Recommendations change, so look again on the day you export. Here is an illustrative output for concept three, and what a person would change.

The last row is the useful behaviour: because the prompt asked the model to list statements needing evidence, it surfaced a hidden statistic. Not every model will list them all, so you still read every line yourself.

For the broader craft of channel-specific wording, see how to write Facebook and Instagram ad copy with AI. This tutorial concentrates on the gates around it.

Gate three: visuals, and a rule for where each one came from

You have three sources: your own photographs, stock you have licensed, and AI-generated images. For a trades business the strongest creative is usually your own finished work. AI is best used for backgrounds, layout variations and tidying, not for inventing the thing you sell.

Examples of safe and unsafe uses in the same campaign:

• Safe: an AI-generated plain-colour backdrop with space for a headline, placed behind a real photo of a finished kitchen.
• Safe: using AI to extend the edges of a real photo so it fits a wide banner.
• Risky: an AI image of a perfect kitchen presented as your work. If a customer expects that finish, the ad misleads, whichever tool drew it.
• Unsafe: an AI-generated "happy homeowner" presented as a customer. It implies a testimonial that does not exist.

Keep a simple source log: one row per image with where it came from, the date, the tool and plan used, and the prompt. That log is the evidence you will want if an image is ever challenged.

Image 1 | finished kitchen, shot on site | own photo | 14 Oct | owner's phone
Image 2 | plain slate-grey backdrop | AI-generated | 2 Nov | image tool, paid plan, prompt saved
Image 3 | fitter measuring worktop | own photo | 9 Oct | consent from fitter on file

Image 3 shows another detail people forget. A real person in a photo, even an employee, needs consent to appear in an ad, and that consent should be in writing.

Gate four: the claims check, line by line

List every factual statement in the ad, including numbers, comparatives, guarantees and prices, and next to each write the evidence you hold. A claim with no evidence in a drawer comes out or gets reworded.

Two of those claims depend on conditions: "from $9,500" and "free". Advertising rules in many places expect that an offer's main conditions are visible, not buried. If the starting price only applies to a small galley kitchen, the ad must not suggest it applies to every kitchen. For regulated products such as finance, health or investments, stop here and have an adviser read the copy; this tutorial is not legal advice.

A second AI pass can help as a challenger, not a judge. Paste the final copy and ask:

Act as a sceptical reviewer. List every statement in this ad that a customer could
reasonably read as a promise or a fact. For each, say what evidence a business would
need to hold. Then list wording that could mislead by omission, such as a price that
depends on conditions not shown.

Treat its output as extra eyes. It will miss things it cannot know, such as whether your guarantee has an exclusion list, so the evidence column is still yours to fill.

Gate five: rights for images, fonts, logos and music

AI tools make it easy to generate an image that echoes someone else's work. Four checks keep you out of trouble.

1. Read the tool's commercial-use terms for your plan. Free and paid tiers often differ. Adobe, for example, says its own Firefly models are trained on licensed Adobe Stock and public-domain content, but its Firefly page lists partner models from other companies separately from the models it calls commercially safe. Whether any intellectual-property protection comes with your output depends on your plan's terms, so read the ones you are on. Whether you can legally use AI-generated images in marketing covers the wider question.
2. Do not prompt with a living artist's name, a brand name or a character. Describe the look you want in plain terms instead.
3. Inspect the result. Look for garbled text, a watermark, a recognisable logo or a face that resembles a real person. Regenerate anything that does.
4. Keep the log from gate three. Rights questions are easiest to answer when you can show where each asset came from.

Fonts and music are the quiet traps. A font inside a design app may be licensed for use in that app's designs but not for a video you edit elsewhere, and a track from a free library may require attribution. Add both to the source log. Ownership of purely AI-generated output is unsettled in many places, so a design you want to protect, such as a logo, is better drawn or commissioned than generated.

Gate six: platform sizes and text limits

Do this last, because it is mechanical. Each platform publishes current specifications and changes them, so read the page on the day you export. For uploaded display banners on Google's display network, Google's help page lists these standard sizes and limits, which I checked on 4 November 2026.

Accepted file types for these uploaded ads are GIF, JPG and PNG, with a maximum size of 150 KB, and animations must run for no longer than 30 seconds. A kitchen photo exported at full resolution will be many times over 150 KB, so resize first, then compress, then test that the text is still readable at phone size. For social placements, check the platform's own ad specification page for the image ratios, text limits and safe areas it currently recommends, since the interface may crop the edges. Meta's guide for Facebook Feed image ads, for instance, currently recommends a 4:5 image at 1440 x 1800 pixels, in JPG or PNG up to 30 MB, which is a very different export from the 150 KB banner.

The practical consequence for the kitchen fitter is to design the headline short enough for the smallest banner first. "5-year written guarantee" fits a 320 x 50 banner; the longer line about fitters does not. Write the long and short versions together at gate two, so you are not trimming copy under time pressure at gate six.

The finished ad, written out

This is the kitchen fitter's final social ad after all six gates. It is a made-up example for illustration.

Notice what happened to the copy. The AI's "dream kitchen" language and the "best guarantee" line are gone. What remains is shorter, stated in specifics, and every phrase has a document behind it. That is also what tends to perform: specific offers are easier to believe than superlatives.

The same method on a second business: an online clothing store

To show the gates are not specific to trades, run a different brief through them. An online clothing store wants a banner for a spring sale on rain jackets. The first draft from the AI reads "Our waterproof jackets keep you 100% dry, guaranteed. Up to 60% off everything."

Gate four takes that apart in under five minutes.

• "100% dry, guaranteed": the product page says the fabric is water-resistant to a stated rating, not that it keeps you dry in any weather. Reword to the actual rating, or drop it.
• "Up to 60% off everything": only eleven lines carry 60%, and the sale excludes new arrivals. "Everything" is false. Reword to "Up to 60% off selected jackets" and state the end date.
• The AI image shows a model in a coat with a visible, slightly altered logo of a well-known outdoor brand. Gate five rejects it; regenerate without any lettering or use a real product shot.

Gate six then sets the size. The 300 x 250 version can carry the discount and a date, while the 320 x 50 mobile banner carries only "Rain jacket sale, ends Sunday". Two minutes of checking prevented an overclaim, a false "everything" and a lookalike logo. That is the value of running gates in order rather than polishing the picture first.

The landing page is part of the advertisement

A compliant ad that sends people to a page contradicting it is still a problem. Before launch, open the destination on a phone and compare it with the ad line by line. The price, the conditions, the guarantee wording and the end date must match. If the ad says "free design visit", the page should say what happens at the visit and whether there is any obligation. If the sale ends Sunday, the page should not still say "ends soon" the following Tuesday.

Ask the AI to compare, giving it the ad text and the page text:

Here is an ad and its landing page text. List every place where the page contradicts,
softens or omits something the ad promises, and every claim on the page that has no
supporting evidence shown. Do not rewrite anything yet.

A typical useful reply is a short list such as: "The ad says from $9,500; the page lists no price", "The ad promises a visit within a week; the page says two to three weeks". Fix the page or the ad, whichever is wrong, and record the date you checked.

Who signs off, and what they are signing

Name one approver. In a four-person company that is the owner; in a larger firm it may be a manager plus someone from the team that fulfils the offer, because the person who delivers the job knows whether the promise is realistic. The approver does not judge the design. They confirm three things: every claim has evidence, every asset has a log entry, and the destination matches. Put the date and their initials on the evidence table and file it with the final artwork.

Keep the file for as long as the ad runs and a while afterwards. If a customer or a platform queries a claim six months later, the table answers the question in a minute instead of a week of searching. A shared folder with one sub-folder per campaign, holding the brief, copy versions, source log, evidence table and exported files, is enough.

Realistic failures and how they show up

• The price that quietly changed. The ad still says $9,500 after a supplier's costs rose. Symptom: enquiries arrive and the first quote does not match. Fix: put a review date in the evidence table and set a calendar reminder.
• The image that looks too good. An AI kitchen with impossible reflections draws clicks, then complaints that the finish was not what the ad showed. Fix: use real photos of real work for anything that stands for the product.
• The invented stat. A model adds "9 in 10 customers recommend us". Symptom: you cannot find the source. Fix: delete it, and add "no statistics unless supplied by me" to the brief.
• The cropped banner. The button disappears at one size. Fix: export each size from the master and view it on a phone before launch.

Testing after launch, and when to stop an ad

With $600 across four concepts, each version gets roughly $150, which will not produce statistically tidy results. Treat the first fortnight as a way to eliminate the weakest concept, not to crown a winner. Note for each version the cost per enquiry and, more important, how many enquiries became booked visits. An ad that gets cheap clicks from people who never book is the expensive ad.

Write one line per week in a log: what ran, what it cost, what came back, and what you changed. Change one thing at a time, the promise, the image or the offer, so you know why a result moved. A local shop's complete AI ad set shows the same cycle with a different business, and ad copy variants with client sign-off is useful if someone else has to approve your ads.

Stop an ad when the claim in it stops being true, when a platform flags it, or when its booked-visit rate stays well below the other versions after the first couple of weeks. If you would rather design this workflow with someone, that is what my AI implementation consultation covers.

A one-page checklist to keep beside the brief

• The brief names claims I cannot make.
• Copy has at least three genuinely different angles.
• Every number, comparative and guarantee has a document attached.
• Conditions on any price or "free" offer are visible near the claim.
• Each image appears in the source log with the tool, plan and prompt.
• No fake customers, fake reviews or invented statistics.
• Every person shown has consented in writing.
• Sizes, text and file weight checked against the platform's current page.
• A named person approved the final version, with a date.

Questions before you publish an AI-made ad

Do I have to say an ad was made with AI?

There is no single rule that covers every ad. Google, for example, requires verified election advertisers to disclose synthetic content that realistically depicts people or events, and advertising rules generally treat misleading imagery as a breach whatever tool made it. Read each platform's current ad policy and, for regulated products, ask your adviser.

Can I use AI images of people in an ad?

You can, but a realistic person who does not exist can mislead if the ad implies a real customer or a real result. Avoid fake testimonials and before-and-after pictures. Use real photos of real customers with written permission for anything that implies an experience.

Who owns an image an AI tool makes for me?

It depends on the tool's terms, and ownership of AI output is unsettled in many places. Read the commercial-use clause of the plan you are on, keep the prompt and the date, and avoid asking for another brand's logos, characters or named styles.

How many ad versions should I make?

Three to five distinct angles is enough to learn something on a small budget. Changing only a colour or one word rarely teaches you anything. Change the promise, the image or the offer one at a time and note which version won.

Further reads

• AI for Creating Advertisements: One Ad Brief, Five Formats — Turn one ad brief into five formats.
• AI Website Copy for Financial Advisers: Service Pages and Bios — Copy for a regulated service with a compliance read.
• Best AI Tools for Advertising Agencies: Copy Variants, Mock-Ups, Reporting — Tools agencies use for mock-ups and reporting.
• Roofing Facebook Ads: Copy and Audiences Tested With AI on a Small Budget — Audiences and copy tested on a small budget.
• AI Ad Copy Generator for PPC Agencies: Responsive Search Ads at Scale — Writing responsive search ads at scale.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Google Ads Help (uploaded display ad specifications), Google Ads political content policy (synthetic content disclosure), Meta Ads Guide (Facebook Feed image specifications) and the Adobe Firefly product page, checked 4 November 2026. Check each platform's current specification page before exporting.