How to Make AI Ads for a Local Shop: One Ad Set From Start to Finish

Photograph the real product, use AI only to restyle the background and draft the words, add the offer text yourself in a design tool, then check the result against the platforms' misrepresentation and AI-label rules before you spend anything. Run one image ad and one 15-second video ad side by side for a week on a small, counted test budget.

Ads fail for plain reasons: the picture shows something you do not sell, the offer on the page differs from the offer in the ad, or nobody counted what the ad produced. The example below is a made-up homeware shop with a mug-set offer, and the checks are written as lists you can copy.

Follow me on Instagram@sagnikteaches

Fix one offer and one counted result before opening any tool

Start with the offer, because everything else, including every AI prompt, depends on it. The shop in this example sells candles, mugs and plants, and its offer is 20% off any boxed mug set, in store only, until a stated end date. Six designs are in stock. There is no delivery. Write those facts down in one place, since they become the rules every prompt and every ad must obey.

Connect on LinkedInSagnik Bhattacharya

OFFER: 20% off any boxed mug set, in store only, until [end date]
STOCK: six designs (list them), all in boxes, all in the shop now
NOT OFFERED: delivery, personalisation, gift wrapping
CLAIM: show the ad at the till and say the code word
CODES: IMG20 for the image ad, VID20 for the video ad
SHOP: [address], open [hours]
GOAL: 10 claimed codes in the first 7 days, from a test budget of $84

Notice two things. The goal is something the shop can count at the till, not clicks, and each ad has its own code word, so you will know which one brought the customer in. Without different codes you will see only a total, and you will argue about which ad deserves the credit. A counted result also means you can check tracking before you spend, the point of setting up conversion tracking before an AI ad spends your money.

Subscribe on YouTube@sagnikteaches

The page the ad points to can be a single screen: the offer, the end date, the code word, the address and the opening hours. If the page says something different from the ad, Google's misrepresentation policy lists "inconsistencies between the ad and the landing page" as a problem, so write the page from the facts block above, not from memory.

Shoot the product first: the real photo is the ad's evidence

Take six photos of the mug sets on a plain white board near a window, one design each, plus a close shot of the printed pattern and the box. A phone is enough. Daylight, no filters, the product filling most of the frame, and the label or pattern in sharp focus. These photos do two jobs. They are the starting point for the AI image, and they are your proof of what the product really looks like when someone says the ad showed something different.

Keep the originals untouched in a folder called "real", and put every AI result in another folder called "styled". Compare them side by side before anything is published. The shop that skips this step ends up with an ad nobody can check against the product.

Image ad: let AI change the table, not the mug

The safest use of an image generator in a product ad is to restyle everything around the product and leave the product alone. Upload one real photo and tell the tool exactly what must not change.

Use the uploaded photo of our boxed mug set. Keep the mugs, the box and the printed
pattern exactly as photographed: do not change shape, colour, handle or pattern, and
do not add any products or logos. Replace the white background with a wooden kitchen
table beside a window in soft morning light. Leave the top third of the frame empty
for text. Make a 4:5 portrait version.

Here is what a first result can look like, and what to do about it. The review below is illustrative.

Result review (side by side with the real photo at 100%)
- Table, window light, empty top third: good
- Mug handle: thicker than the real one              FAIL
- Pattern colour: blue shifted towards teal           FAIL
- A third mug appeared at the edge of the frame       FAIL
- Text on the box: gibberish letters                  FAIL
Decision: reject. Re-run with "show only the objects in the photograph;
do not redraw the box or its text".

Four failures in one image is a realistic result, and each is a misrepresentation risk. A handle that is the wrong shape, or a colour that is not the colour on the shelf, turns the picture into a product you do not sell. Google's misrepresentation policy bars offering "products or services that you don't have or can't deliver" and says manipulating media to deceive or mislead is not allowed. Whether the manipulation was deliberate matters less than whether a customer would feel misled in the shop. If three attempts do not fix the product, stop: use the real photo with a simple background removed and a styled background added in a design tool.

Make a checklist and run every image through it before it goes anywhere.

• Product unchanged: shape, colour, pattern, handle, count. Compare at 100%, not in a thumbnail.
• No invented text: image generators often garble lettering. Remove any text the tool drew on the product or box.
• No extra products or brands: a third mug or a stray logo is a promise you cannot keep.
• No AI-generated people. A photorealistic person made by AI changes how Meta labels the ad, as the rules section below explains.
• Looks like your shop: if a customer walks in expecting the table in the picture, would they find a similar display?

Offer text goes on afterwards, in a design tool such as Canva, not in the AI prompt. Generators can misspell words and invent letters, and a spelling mistake in "20% off" is a cheap way to look careless. Add the offer, the end date and the code word as live text in the design tool, then export a 4:5 portrait and a 1:1 square, and check each platform's current ad specs on the day. For a broader view of free tools you could use at this stage, the tutorial on free AI advertisement tools for local shops compares five.

Video ad: three real clips and one animated still

Fifteen seconds is long enough to show an offer and short enough to finish. The same rule applies as for the image: let real footage carry the product, and use AI for the one shot you cannot film cheaply. Here is a shot list for the mug-set ad.

Film all three real shots in one ten-minute session, holding the phone upright, and keep each clip a few seconds longer than you need so you can trim. For the animated shot, upload the checked still to an image-to-video feature. Canva lists "turn images into videos" as a Premium AI tool, available on its Free and Pro plans, while its "create a video with Canva AI" feature is an Ultra tool, not on Free and included on Pro. Runway is an alternative: its Free plan gives a one-time 125 credits and its Standard plan lists 625 credits a month with no watermarks, from $15 a month, or $12 a month billed annually. Both pages are summarised in the plan table later on, and both change, so check before you subscribe.

Check the animated clip before you use it. Play it at full screen with the sound off, and watch for four things: the mugs drifting or changing shape, the pattern warping, the steam appearing from the wrong place, and any object that was not in the still. Reject the clip if any of them shows, even for a moment. Then add captions, because some viewers watch with the sound off, and keep the end card on screen long enough to read twice. If you add an AI voiceover, record the script you wrote, not one the tool improvised.

A short video built from your own footage plus one animated still is cheaper to make safely than a fully generated ad, and it is far more honest about what the shop sells.

Copy: three variants, then strike the claims you cannot prove

Give the assistant the offer block and ask for three short variants. Tell it what it may not claim, because it will otherwise reach for the usual retail phrases.

Write three Facebook and Instagram ad variants for the offer below. Each needs a
primary text of two short sentences and a headline of five words or fewer. Use only
the facts in the offer block. Do not claim best-selling, handmade, free delivery,
limited stock or anything not in the block. Write plainly and warmly, in the spelling my customers use.

[paste the offer block]

A first answer might read like this (illustrative).

A  Our best-selling mug sets are 20% off this weekend only! Handmade designs, free
   delivery over $30. Show this ad in store.   Headline: Weekend mug sale
B  Six boxed mug sets, one lovely offer: 20% off in store until [end date]. Show this
   ad at the till and say IMG20.   Headline: 20% off mug sets
C  Treat yourself, or someone else. Pick a boxed mug set and save 20% in store.
   Headline: Boxed mug sets, 20% off

Variant A breaks the rules four times: "best-selling" is unprovable, "handmade" is false for a factory-made range, "free delivery" is not offered and "this weekend only" contradicts an offer that runs until a stated date. Google's policy lists "promising products, services, or promotional offers in the ad that are unavailable", and an ad that promises free delivery you do not provide is exactly that. Variant B is nearly ready. Variant C is safe but leaves out the end date and the code, so a customer cannot act on it. Use B for the image ad, let C become a second variant once it has the end date and a code, and drop A. Keep the edited versions next to the originals: the differences are your claim checklist for the next ad. For a longer list of prompts for this stage, see writing Facebook and Instagram ad copy with AI.

The rule check, in one table

Run this table on every ad before it goes live. It reflects what each platform says in its own pages, and a final column tells you what to do. Policies change, so treat the table as a prompt to read the current page, not a substitute for it.

Beyond platform rules, decide your own line on disclosure and write it down. My line would be this. A background restyled by AI, around a real product, needs no announcement, though I would add "image styled with AI" to the caption anyway because it costs nothing. An animated still needs the same line. Anything an AI generated to resemble a real thing a customer might rely on, such as a shop front, a result or a person, should not be in a shop ad at all. If you want a fuller view of what customers expect, the tutorial on telling customers when you use AI in marketing sets out the question from the customer's side.

If you advertise to people in the EU, the AI Act's transparency rule (Article 50) has applied since 2 August 2026 and covers realistic AI-generated or manipulated image and video. How it applies to a shop's ad depends on the detail, so ask a qualified adviser rather than assuming. Keeping to real products and no AI people avoids most of the problem anyway.

Tools and plans, as the pricing pages read on 28 October 2026

You need very little paid software for this build, but the allowances matter, because every failed attempt uses some of them. The table says what each page states. Plans and limits change often, so read the page again on the day you start.

For this ad set, free tiers are enough to start. Canva's 20 uses on Free have to cover every attempt, so plan them: a handful for the image and a few for the clip, and do the rest of the editing by hand. If the animated shot keeps failing, drop it and use a fourth real clip instead. A video does not need AI to work.

Set the test: radius, budget and Meta's 50-results rule

Choose the result nearest a customer that the platform can count cheaply, set a radius around the shop that matches how far people really travel, and keep the audience broad to begin with. The budget is the part most owners guess. Meta's own help page says an ad set usually leaves its learning phase "after about 50 results in the week after the ad set's last significant edit", and that an ad set which cannot reach enough results shows as "Learning limited". So work backwards from the result you optimise for. The costs below are assumptions for illustration, not benchmarks; replace them with the cost per result from your own account.

This shop will spend $12 a day, $84 over seven days. That affords 50 results a week only if a result costs $1.68 or less ($84 ÷ 50), so the first two rows are in reach and the last two are not. The practical answer is to optimise for landing page views and judge the ads by the code words claimed at the till. Raise the budget to about $29 a day if you want the platform to optimise for claim forms.

Put both ads in one ad set, name everything so you can read it in a report, and tag the links so the page counts visits per ad.

Campaign:  Mugs_Oct_Offer
Ad set:    Mugs_Oct_Radius_Views
Ads:       Mugs_Oct_Image-A   (code IMG20)
           Mugs_Oct_Video-A   (code VID20)
Link tags: ?utm_source=meta&utm_campaign=mugs_oct&utm_content=image-a
           ?utm_source=meta&utm_campaign=mugs_oct&utm_content=video-a

Then leave the ad set alone for the full week. Meta treats edits that meaningfully change how an ad set might perform as significant, and lists frequent budget changes among them; a significant edit restarts learning, and with it the week Meta uses to count results. If you later move to Google's Performance Max instead, Google's guidance is to run it for at least six weeks before comparing performance, and changes reset learning there too. The tutorial on testing ad variations with AI on a small budget covers how much difference a small budget can and cannot detect.

Read the first week against what a claim is worth

At the end of day seven, put the platform's numbers beside the till's. This is an illustrative result for the mug-set test.

Start with what a claim is worth. A boxed set that lists at $36, costs the shop $14.40 and sells at $28.80 after the discount earns about $14.40 before the ad cost, so the break-even cost per claim is $14.40. The image ad, at $7.60, is comfortably under it. The video ad, at $23.00, is over it.

Now be honest about the sample. Seven claims across both ads is below the goal of ten, and a 5-to-2 split could easily be luck. A workable rule is to act only when one ad's cost per claim is at least double the other's and the ad set has produced at least ten claims. Here the ratio is 3 to 1, but there are only seven claims, so the right move is to extend the test a second week with nothing edited. Meanwhile ask the staff what customers said when they claimed: "I saw it on Instagram" and "my friend showed me" are data the report cannot give you. The tutorial on how much to spend on AI-run ads before judging results explains why patience matters at this scale.

If the second week confirms the first, pause the video ad, keep the image ad running, and spend the saved budget on a second image variation built from a different mug set. If the video ad catches up, you have learned that both formats work and can keep both. Either way, write down the numbers: cost per view, claims, cost per claim, and what a claim is worth. The next ad set starts from those, not from a guess.

Here is how the work adds up for this shop, with time estimates for a first attempt.

Nothing in that list needs a designer or an agency. What it does need is the habit of treating every AI output as a draft to be compared with the real product and the real offer, and of writing down what each ad earned.

Questions shop owners ask about AI ads

Do I have to label an ad as AI-made?

It depends on the platform and the rules keep moving. Meta says it applies an AI info label to ads made or materially edited with its own generative tools, and it requires self-disclosure for AI-made media in ads about social issues, elections or politics in certain cases. Read each platform's current page before launch, and say so yourself when a customer could mistake the image for a real photo of your product.

Can I use AI-generated people in a shop ad?

I would avoid it. On Meta, an ad with an AI-generated photorealistic person gets its AI info label next to the Sponsored tag, and a made-up person can look like a made-up customer. Show real staff or customers who have agreed in writing, or show hands and products, which say something true about your shop.

How long should the test run?

Seven days is a sensible first test, because Meta counts about 50 results in the week after an ad set's last significant edit before it leaves its learning phase. Avoid editing the ad set during that week. If you use Google's Performance Max instead, Google's guidance is to give it at least six weeks before comparing results.

Further reads

• AI for Creating Advertisements: One Ad Brief, Five Formats — One ad brief turned into five formats.
• Free AI Video Generator for Small Business Ads: Limits to Check First — Free video generators and the limits to check first.
• How Meta and Google Use AI to Run Your Ads: A Plain-English Guide — How Meta and Google use AI to run ads once you pay.
• Can You Put Jewellery on an AI Model? Disclosure Rules for Sellers — Disclosure questions when AI shows your product on a model.
• AI Tools for Ecommerce Photography: Backgrounds, Shadows, Consistent Sets — Backgrounds and shadows for product photos in consistent sets.
• Local SEO for a Retail Shop: Profile, Stock Posts and Reviews With AI — Profile, stock posts and reviews that work alongside paid ads.
• Art Supply Store Marketing: Workshops and Student Offers With AI — A worked month for an independent art supply shop: two capped workshops, a student-offer email and four brand-of-the-month posts, with the maths done first.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Meta help page on how AI-generated images in ads are labelled and Meta's June 2026 update on AI in ads; Google Ads misrepresentation policy; Google Merchant Center image requirements; Meta help page on the learning phase; Canva, ChatGPT and Runway pricing pages; Google's Gemini help page on image downloads (all checked 28 October 2026). Costs per result and sample results are illustrative.