Copy the Ads That Already Work: A Meta Ads Library Teardown Before Your First $10

Search your niche, sort by impressions

1. Open facebook.com/ads/library. No account needed.
2. Pick the country you sell in, set All ads, and type the words your buyer would search for: "ai headshot", not "generative portrait platform".
3. Leave Active status: Active ads on. Open Sort by and check that Impressions: high to low is picked. It was already picked on my search; the other option, Most recent, buries the ads with reach under this morning's tests.
4. Scroll and write down the advertisers that show up again and again. In the first 66 cards on my search, Career Pic AI had 22, GIO 13 and Roomvu 7. Those are the ones spending.

Keep only the ads that have run 30+ days

Every card shows Started running on. Subtract it from today. A ten-day-old ad proves nothing, because everybody tests. An ad still live after 30 days has survived a month of the advertiser looking at its numbers. Of the first 66 cards on my search, 26 had run for 30 days or more, and once repeats of the same ad are removed, a handful of real long runners are left.

Then look for the second signal: "N ads use this creative and text". That means the advertiser copied the same ad into more ad sets, which is what you do with a winner. Career Pic AI's "dated photo" ad below has run since August 2, and two of its versions that started on September 11 each show 24 ads using the same creative and text.

Ask four questions of every ad

Open each long runner, play the video if there is one, and write down four answers:

1. Hook. What stops the scroll: the first line of text and the first two seconds of the video. Who is it about, the viewer or the product?
2. Framework. In what order does the ad tell its story? Problem first, result first, a person talking, a list?
3. Call to action. What exactly does it ask for, and what takes the risk away: a price, a free preview, a refund?
4. How the product is shown. A screen recording, a before and after, a person using it, a static image?

Cravotta's observation from subscription apps is that most winners follow one frame: hook, value proposition, problem, solution, then a before and after. The before and after is the step people skip, and it is the one that makes the result believable.

The teardown: AI headshot ads

Here are the four questions on five ads from the search, four long runners and one new ad for contrast. Ages are as of October 19, 2026.

Career Pic AI, 78 days. The hook is a fear the viewer already has: the photo is costing them the callback. One sentence of how ("upload one selfie and get 10 professional headshots"), three checkmarks that answer the objections before anyone asks (no studio and no $500 photographer, full commercial rights, a refund on request), then the price: $4.99 one-time. The video is just the result, a grid of finished headshots scrolling past.

The same advertiser, a second hook. Career Pic AI's LinkedIn version keeps every part of the frame and changes only the first line: "Your LinkedIn photo is the first thing a recruiter sees." That is what a winner looks like from the outside: one frame, several hooks, each one aimed at a different moment.

Roomvu, 193 days. The oldest ad on the search names the buyer in the first three words: "For Mortgage Brokers". Roomvu runs the same ad for real estate agents and for insurance agents, one ad per profession. If your product serves a few clear groups, this is the cheapest hook there is: say their name.

Kevin Titus Photo with Evoto AI, 49 days. This is the user-generated frame done straight: a working photographer, in first person, describes the chore (glasses glare), says what the tool does about it, and hands over an offer through his own link. The ad runs under his name with the brand's next to it, so it reads like a recommendation, not a banner.

Headshot Photos, 17 days. A strong counter-position: "A cheap AI tool can fake the suit. It can't fake the person clients are deciding to trust." Write it down, but do not model yours on it yet. Come back in two weeks. If it is still running, it earned a place in your table.

Turn ten screenshots into one pattern table

Screenshot each long runner's card, the whole card, so the start date and the copy count are in the picture. For video ads, note what happens in the first three seconds, because a screenshot only shows the cover frame. Then hand the pile to a chatbot that reads images:

I'm attaching [N] screenshots of ads from the Meta Ad Library. Every one has run for 30 days or more in my niche: [niche]. Today is [date].

For each ad, fill one row of a table with these columns:
- Advertiser
- Days running (from "Started running on")
- Copies (the number in "N ads use this creative and text", or 1)
- Hook: the first line, word for word
- Who the hook is about: the viewer, the product or the advertiser
- Framework: the order of the parts, using only these labels: hook, problem, solution, proof, before/after, objection, offer, CTA
- Call to action and risk reversal (price, trial, refund)
- How the product is shown: screen recording, before/after, person talking, static image, other

Then, under the table:
1. The three things at least half of the ads do.
2. One thing none of them do.
3. Which ad I should model mine on, and why, in two sentences.

Do not guess. If a screenshot cuts off text, write "not visible".

Write yours on the same frame

Now keep the structure and swap in your facts. Three ads, one frame, three different hooks, so the test in step 7 tells you which first line works.

Here is the pattern table from my teardown:
[paste it]

My product: [name], [one sentence: what it does]
Who pays: [the person]
The painful moment they feel the problem: [one sentence]
Proof I actually have: [numbers, results, reviews; only real ones]
Price and risk reversal: [price, trial, refund]

Write 3 ads on the frame of [the ad you picked]. Same order of parts, my facts only, none of their words. For each ad:
- primary text, with the hook inside the first 125 characters
- a headline under 40 characters
- what the first 3 seconds of the video, or the image, must show
- the before and after, in one line

Give each ad a different hook: one about the pain, one that names the buyer's job, one about the result.

1. Rewrite the drafts in your own voice. Keep the numbers real: if you cannot prove "in 2 minutes", do not say it.
2. Make the creative show the product working. A 15 to 30 second screen recording of the real result beats a polished logo animation. Every long runner above shows the result.
3. Pick the problem before the ad. Cravotta's rule: when the problem really hurts, the ad only has to show the viewer it is about them. If your three hooks feel weak, the problem may be the thing to fix.

Run it at $10 a day, read it after a week

1. One campaign, one ad set, your three ads. Let Meta choose the placements.
2. $10 a day for seven days, and no edits for the first few days. Every change restarts Meta's learning, and at this budget you will not leave the learning phase anyway (Meta wants about 50 results a week for that). So read the ads, not the algorithm.
3. After seven days, read three numbers per ad: how many people clicked out of those who saw it, what a click cost, and how many clicks turned into a signup or a sale on your page.
4. Switch off the hook with the worst click rate, write two new hooks on the same frame, and run another week. When one ad keeps bringing signups at a cost you can afford, put it in more ad sets. That is the "N ads use this creative" you saw in the library.

I run Nick Launches' own Meta ads at about 10 Canadian dollars a day. It is small enough to keep running while you learn, and big enough to show which of three hooks people stop for.