--- name: meta-creative-testing description: Judges new Meta ads the way a creative tester does, from Ads Manager exports. Runs the two winner tests (Meta keeps spending as the budget grows, most sales from people who never bought), picks what to fix first on a failing ad, counts concepts instead of ads, sets the test budget, works out how many new ads a week an account can judge and needs, and writes the weekly dead-idea notes. Use when someone shares Meta Ads Manager data and asks which ads are winning, what to fix, what to test next, or how many new ads they need. --- # Meta creative testing Judge ads on what Meta does with money and on who buys, never on cost per purchase alone. A cheap cost per purchase on the first few hundred dollars only shows the ad found the easiest buyers. Pick the job from what the user asks. If data is missing, ask for it once, in one message, naming the exact export (section "Exports"). ## Exports All from Ads Manager > Ads tab > Reports > Export table data (.csv). - **Columns**: Ad name, Ad set name, Amount spent, Purchases, Cost per purchase, Impressions, 3-second video plays, Video average play time, Link clicks. - **Before / after**: when judging a budget change, two exports: the 7 days before the change and the 7 days starting the day after it. Do not judge until the second week is complete. - **New vs past buyers**: the "after" export with Breakdown > By delivery > Audience segments. This only works if Existing customers are defined in the ad account's Advertising settings (Audience segments). If the user hasn't set it up, tell them how and judge the spend test only. ## 1. Winner check For every ad with at least 3 purchases in the "after" week: 1. **Spend test**: its share of its ad set's spend, before vs after. Passes if the share held or grew. 2. **New buyer test**: share of its purchases from "Existing customers". Passes if under 50%. New audience + engaged audience = people who never bought. Output one table sorted by spend: ad, cost per purchase (before), spend share before -> after, % purchases from existing customers, verdict. | Verdict | Rule | Action | |---|---|---| | WINNER | passes both | make versions: new face, format or edit, same reason to buy | | RESELLS TO PAST BUYERS | spend test only | leave it running, don't scale on it | | META WON'T SPEND | new buyer test only | keep it on test budget, don't build on it | | FAILS BOTH | neither | turn it off, keep what stopped people | | NOT ENOUGH DATA | under 3 purchases | one more week | Then list failed ads with the highest 3-second video plays / impressions: their openings worked and are worth reusing. ## 2. What to fix first (one failing ad, about 7 days on its own budget) Check in this order and stop at the first that applies: 1. **Meta barely spent on it** (far below its share of the test budget): the idea failed. New concept, often a new format. No hook will save it. 2. **Most people leave in 0 to 3 seconds** (3-second plays / impressions well below the account's own other ads, not an internet benchmark): the opening failed. Test new hooks, keep the rest. 3. **They watch but don't click** (average play time holds, link click-through rate low): the middle never made them want it. Keep the opening, rebuild the middle: the reason, the proof, the demo. 4. **They click but don't buy**: check the landing page, price and stock before touching the ad. 5. **Right reason, wrong person for the buyer** (comments, age breakdown): change the person and the place. That's a version. ## 3. Count concepts, not ads - A **version** keeps the reason to buy and changes the hook, the face, the format or the edit. - A **concept** is a different reason to buy, for a buyer in a different moment. - Quick test: ask the buyer in the ad "why did you buy?". Same answer = version. Group the export by concept. If ad names start with a concept code (e.g. `C07-nightshift-nurse-V02-new-face`), group by the code. If not, group by reading the names and ask the user to confirm the groups. Report: ads, concepts, and concepts that had a full week of test spend (only those count as tested). Suggest the naming pattern `C{n}-{buyer moment}-V{n}-{what changed}`. ## 4. Test budget - New ads get their own ad set or campaign with their own budget, never the ad set the winners live in (adding an ad there is a significant edit that resets learning). - All test spend together: at or under 20% of daily budget. - Each idea: about one sale's worth of spend a day (the account's cost per purchase), for 7 days. - More ideas than slots: the most different go first (new buyer, new reason, new format before a small twist). - A test that passes the winner check moves to the main budget; its slot goes to the next idea. ## 5. How many new ads a week ``` Ceiling = (daily budget x 0.2) / cost per purchase = most ideas judged at once, a week each Floor = concepts needed per winner / weeks a winner lasts ``` Ask for the account's hit rate (concepts that became winners). If unknown, show the floor for 1 in 10 and 1 in 20, and a winner life of 2 to 3 weeks. If the ceiling is below the floor, say so plainly: more ads won't help, the account needs a bigger test budget or a lower cost per purchase first. ## 6. Versions vs new concepts in the next batch | Winner's state | Mix | |---|---| | Cost per purchase steady | mostly versions, plus new concepts in every batch | | Cost per purchase climbing | shift toward new concepts | | Meta pulling spend off it | new concepts carry the batch, versions only buy time | ## 7. Dead-idea notes (weekly, before the next brief) An idea is dead when Meta stops spending on it, or its cost per purchase stays above target for a full week of real spend. For each one, fill: ``` Idea (concept code): Who it spoke to (person + moment): The line that carried it: Weeks live / total spend: How it died: never made them want it OR sold, then tired Already tried: new hook? rebuilt middle? new person? Next: a different reason for this buyer OR a successor (new buyer or new reason) ``` Every new concept in the next brief should point back to one note. ## 8. Picking between two concepts for one slot 1. Is it a person in a moment, not an age range? 2. Would a stranger get it, not only people who know the brand? 3. After one listen, do you know what it sells? Both pass: the one most different from what's running gets the slot. Every concept should fill: `{person} in {moment} buys {product} because {reason}.` ## Rules - Never call a winner on cost per purchase alone, and never within 7 days of a budget change. - Compare an ad against the same account's other ads, not against benchmarks from the internet. - Say which numbers you could not compute and which export would fix it. Don't guess missing columns.