Mine Meta Ad Library and TikTok Creative Center, then use an LLM to extract structure — hooks, proof, offers — and rewrite in your voice instead of cloning claims.
Competitor ads are a structure library, not a script library. Open Meta Ad Library and TikTok Creative Center, save ads that actually stopped you, then use an LLM to extract the skeleton — hook type, first frame, proof beat, offer shape — and throw the original words away. Rewrite in your customer's language, with claims you can stand behind, and test three structures as hooks. If you cannot name the mechanic without quoting the rival, you do not have research — you have a copy-paste waiting to become a policy problem.
Takes 10 minutes
On the first line of the research doc: we steal structure (hook type, first two seconds of visual, proof order, offer math), never claims, never trademarks, never a rival's unique phrasing, never their talent. If a sentence would still make sense with your product name swapped in and it is a generic mechanic, it may be fair game to test. If it is their slogan, their study, or their face, it is not.
Structure looks like: 'confession hook, empty-fridge first frame, demo by day three, bundle CTA.' Copying looks like: their opening line with your URL. The first can become three original hooks. The second is how you inherit their legal risk and their fatigue. Your audience may have already seen their ad; running a paraphrase still feels like the same ad in the auction's creative clustering more often than people admit.
Add a second rule: every saved ad must become a testable hypothesis or it does not get saved. 'Nice UGC' is not a hypothesis. 'Question hook plus product-in-hand might lift hook rate on cold traffic for SKU 12' is. Research that does not specify the test is mood boarding. Cap the session before you start so you cannot collect forty near-duplicates of talking heads.
Pro tip:If you would not want a competitor to run your line tomorrow, do not run theirs today.
Takes 25 minutes
Pick five to eight advertisers: two direct rivals, two adjacent category winners, two brands you steal mechanics from even if they sell something else. In Ad Library, filter by the page, country, and platform. Sort for the ads that have been running a long time or that keep coming back — duration is a crude proxy for 'this was worth the spend,' not a guarantee.
For each ad, capture only what you need: a still of the first frame, the first spoken line, on-screen text, and a one-line guess at the offer. Do not archive the entire 45-second read unless the proof beat is the point. Note whether it is UGC, founder, green-screen, or B-roll-led. Ten complete cards beat fifty screenshots you will never reopen.
Ignore comments that say 'I want this' as proof of performance. Library does not give you CPA. A long-running ad is a clue, not a trophy. If a rival is still running a claim you know you cannot make, do not 'adapt' the claim. Adapt the shape and put your true proof in the middle. Their longevity might be an aggressive claim you will not copy on purpose.
Takes 25 minutes
Use Creative Center for top ads, formulas, and industry examples in your region. Same cap: 25 minutes, same five-to-eight brand list plus a few in-category top ads you do not recognize. Save first two seconds and the mechanic. TikTok's native patterns (text-first, stitch energy, voiceover on B-roll) are often cleaner to extract than Meta's because the platform already groups them as formulas.
Watch on mute first. If the open does not work without sound, tag it as sound-on dependent so you do not blindly port it to a Reels placement that needs captions. Then watch with sound for the hook line. Write the line down even if you will throw the words away — you need to see how short it is. Most winning opens are shorter than your current script's first sentence.
Do not download and re-upload anyone's file, ever. Do not clone a Spark-style organic post that is obviously a creator's likeness. You are taking notes, not taking assets. If a trend audio is the only thing that makes the ad work, you do not have a portable structure; you have a trend. Skip it unless you can name the visual mechanic without the audio.
Takes 20 minutes
Paste transcripts or your notes — not the video file as a thing to recreate — and ask for a table: hook type, first-frame visual, proof type, offer shape, CTA, what is probably a claim you must not copy. Ask for three original hook directions for your product that keep the mechanic and change the language. If the model returns a paraphrase of the rival's first line, reject the output and tighten the prompt.
A prompt that works is boring: 'Do not reuse more than two consecutive words from the source. Name the mechanic in eight words. Propose three hooks in our customer quotes below. Flag any health, income, or uniqueness claims.' Attach your voice-of-customer list and your banned-claims list. The LLM is a pattern extractor, not a ghostwriter for the competitor's legal team.
Calibrate on one example as a team. If two people cannot agree that the output is a new line, the prompt is still too loose. Save the good prompt next to the swipe file. Do not let juniors 'just ChatGPT the ad' without the claims list. That is how invented clinical results show up in a test grid.
Takes 20 minutes
Take the three hook directions and write them in the words your buyers already used in reviews. Then check every line against what you can substantiate. A structure borrowed from a rival who claims 'dermatologist-tested in 14 days' does not give you that claim. Your proof beat might be a 30-day return, a texture shot, or a real review quote you have permission to use.
Read the hooks out loud. If they still sound like the rival, they are still a copy. Change the object, the guilty pleasure, the specific mess. Specificity is how you leave their ad without leaving the mechanic. Pair each hook with a first frame you can actually produce: your product, your avatar, your B-roll. A mechanic you cannot shoot is not a mechanic you can test.
Log the source card as 'structure used, words discarded' so a later teammate does not 'improve' the line back toward the original. This log is also your defense if someone asks whether you cloned a competitor. You want to show a table of mechanics, not a folder of their ads renamed.
Takes launch window 48–72 hours
Lock one body script and one offer. Render 3 hooks × 2 avatars. Launch together, wait 48–72 hours or about 1,000 impressions per file, then mark each source mechanic won, grey, or killed. The competitor does not get credit if their shape dies in your account. Your audience is not their audience.
Name files by mechanic and avatar, not by competitor name. You do not want 'brandX-hook1' sitting in Ads Manager. After the window, write the result back on the research card. A mechanic that loses twice gets retired for this product. A mechanic that wins gets rewritten again next month, still in new words, because even borrowed shapes fatigue.
When you generate the grid, paste the three original hooks into your UGC renderer and keep the body identical so you are testing the open. That is the same discipline as any hook test; the only difference is the ideas started as library notes. If you change the offer on one variant, you are no longer testing the research.
Pro tip:Klip Kanvas can render the 3 × 2 from the three rewritten hooks in one sitting — still never paste the competitor's original lines into the script box.
AI makes competitor research faster and copying easier, so the fence belongs in the prompt and the brief. Libraries are intake. The LLM extracts structure. Voice-of-customer and the claims list do the rewrite. A 3 × 2 grid is the only grade that counts. Do not clone first lines, faces, or proof you do not own. A long-running rival ad is a hypothesis about a hook type, not a script waiting to paste.
Do not treat this as legal advice. Recreating their words, trademarks, or a recognizable likeness is the danger zone. Extract the mechanic, rewrite the language, use your own talent or licensed avatars, and only make claims you can support.
You can ask for structure. If you ask it to remake the script, you will get a paraphrase. Reject paraphrases. Require new wording and a claims flag on every line.
Eight to twelve complete cards is plenty. If you cannot extract a mechanic and a test hypothesis for each, you collected too many.
That is a format, not a mechanic. Push one level down: what is the first line doing, what is in the first frame, what proof appears by second ten. Format-only swipes produce identical tests.
Create your first AI UGC video ad in minutes — no filming, no actors, no editing.
Try Klip Kanvas freeFeed ChatGPT real customer language, lock a four-beat UGC structure, force spoken English, and strip the model tells before anyone films or renders the ad.
Use Gemini in Asset Studio, Demand Gen, and Performance Max as a draft engine: feed a real brief, rewrite every line against claims, and keep human UGC as the YouTube-native asset.
Paste a product link and Klip Kanvas writes the script, casts the creator and renders the ad — no filming, no actors, no editing.
