Best Metrics to Judge AI Ad Creative, Ranked

Hook rate is the first honest read on an AI ad because most generated spots die before the product argument starts. Define it as 3-second video views divided by impressions (or the platform's closest thumb-stop column — name it in the sheet so the team stops arguing). A cold-traffic bar around 30%+ is a healthy conversation starter in many UGC accounts we work with; below that, do not bother defending the body. The file lost the scroll. Recast the first frame, not the offer.

Judge hooks against other hooks, not against CPA, until the open is alive. Swap only the first three seconds, freeze the body, and you will see this column move. AI workflows that fan out hook × avatar grids exist so this metric has something to say; a single melted take cannot tell you whether the face or the line failed. Read it after enough impressions that a cheap hour of delivery cannot fake a winner — a pre-committed impression floor, not 'tomorrow morning'. Put the definition in the column header every week so a new intern cannot mix 2-second views into a 3-second hook rate. If the platform renames the column, screenshot the mapping into the sheet; do not rely on memory.

It lies when the placement mix shifts. Reels and Stories inflate thumb-stops versus in-stream; compare like placements or you will promote a 9:16 cut because it 'hooks' and then watch it fail in feed. It also lies on retargeting, where people already know you. Keep a cold-only view. If hook is strong and hold is weak, the open is not the problem — stop iterating first frames. That is how this ranking earns its keep: it tells you which later metric to read next. On Advantage+ blended delivery, pull a breakdown by placement before you kill a hook. A file can look healthy in total and dead in Feed. That is a placement story, and recasting will not fix a Reels-only win you misread as a concept win.

Hold rate asks whether the argument after the hook is worth the remaining 20 seconds. Pick one definition and freeze it: 15-second views over 3-second views, or average-watch over length, or thru-play on a 15-second cut. In the accounts we work with, cold UGC that keeps 12–20% of hooked viewers into the mid-spot is doing a real job; a steep drop right after the open means the second sentence was a slogan. AI scripts fail here when they stall on a greeting.

Read hold only on ads that already cleared a hook floor. Otherwise you will 'fix pacing' on a file nobody stopped for. The production lever is structure: demo-first and objection-then-proof hold better than a feature list to camera. If you generate with a beat map, the hold curve is a test of whether the model followed it. A talking-head lecture with B-roll wallpaper will hook on a rude first line and then haemorrhage. Annotate the hold curve at the second where you planned the second demo. If the cliff is there, the beat map was ignored in production. If the cliff is immediately after the hook, the second sentence is the slogan you failed to ban.

Do not confuse hold with completion on a 6-second cut — everyone completes a bumper. Match length when you compare cells. Sound-off is most of the impressions; if captions do not carry the argument, hold will sag even when the VO is good. And a 12–20% band is a working range, not a trophy. Some categories sit lower and still buy well because intent is high. Use the range to catch cliffs, not to brag in a Slack screenshot. When you compare AI avatars, match length to the second or hold will crown the shorter file. A 16s cut is not 'higher hold' than a 28s cut with the same story. Normalise or do not compare.

CTR is the close. On cold paid social, an outbound (link) CTR around 1%+ is a bar we treat as 'the CTA and the promise are doing a job' — below that, with healthy hook and hold, you have an entertaining file that does not ask for the click. AI UGC often under-closes because the prompt forbade selling and then forgot to add an afterthought link. Ranked third because a click without hold can be curiosity; a click after hold is intent.

Use outbound clicks over impressions, or CTR on thru-plays if you want to judge only held viewers — but then say so in the sheet. Platform 'CTR' that includes likes and photo clicks will flatter a pretty avatar and hide a dead offer. Match placements. A 1%+ bar is a starting conversation; high-intent retargeting should clear it easily, and some cold prospecting in cheap auctions will sit under it and still CPA. The point is the diagnosis: strong hook, strong hold, weak CTR is a CTA or offer problem, not a first-frame problem.

Fix the close without rewriting the ad. Afterthought CTAs, a visible URL, a platform button that matches the spoken line. If the click is healthy and purchases are not, stop reading this column and go to the landing page. CTR cannot see the PDP. That is why it is not rank 1, and why ROAS is even worse as a creative score. If outbound CTR is strong only on retargeting, do not export that bar to cold cells. Split the sheet. A 1%+ cold conversation that you pollute with warm clicks will keep bad prospecting files alive and starve new hooks.

Once CTR is alive, the next creative-adjacent metric is conversion rate on the click — purchases (or the event you actually optimise to) divided by outbound clicks. It is not a pure creative score, but AI ads fail here in a specific way: the script promised a mechanism the PDP buries. Ranked fourth because killing the video for a page mismatch is how good generated creative gets a bad reputation. Ranked here because AI teams often regenerate the avatar when the PDP is the leak. Check this column before you spend another credit. A matched first screen is cheaper than a new batch, and it is still a creative decision in the only sense that matters: the sentence they clicked.

Compare cells that share a URL separately from cells that do not. If two hooks send to the same page and convert differently, that is still creative. If every hook dies on the page, that is the page or the offer. Message-match the hero to the winning line before you recast the avatar. Speed, trust badges and a broken coupon will also tank this number — so do not treat it as a script note until you have ruled those out once. Screenshot the first screen next to the ad's first three seconds in the test doc. If a stranger cannot see the same noun in both, you do not need more data. Fix the page or stop scaling that cell.

For AI UGC, keep a column for 'promise in the ad' versus 'first screen of the URL'. That gap list is more useful than a decimal. When you scale a winner, freeze the URL. Swapping the landing while you raise budget is two tests at once, and then everyone blames the generator. Checkout issues still hide here, so run one control: same ad, known-good URL. If that converts, the new page is guilty. If it does not, you are back in the creative stack. Do this once before a week of regenerating faces. Credits are cheaper than a bad diagnosis, but they are not free, and a control URL is the cheapest test in the stack.

Cost per purchase (or per the event you trust) is how the business keeps score. It is a terrible first read on AI creative because it bundles offer, audience, bid, frequency, landing page and the file. Ranked fifth as the verdict you take once cells have had matched spend and the diagnostic columns are not screaming. If CPA is high and hook is 18%, you already knew. Ranked fifth on purpose: it is the scoreboard after the diagnostics. Using it first is how AI UGC gets blamed for an offer problem, and how a great hook gets killed because the cell did not get even spend. Matched spend is part of the metric; without it, CPA is gossip.

Match spend and match the optimisation event. A $20/day cell on a $90 CPA product has not spoken after an afternoon. Write a floor (spend or conversions) before launch. Do not crown a winner that had a cheap CPM hour. When two files have similar hook, hold and CTR and different CPA, then you have a creative quality gap worth recasting — often avatar, proof device, or a claim that created returns rather than purchases. Log the spend each cell actually received, not the cap you set. Advantage+ will starve siblings; a starved cell with a pretty CPA is not a winner. If you cannot match spend, read hook and hold only and refuse a CPA crown.

Returns and chargebacks belong next to CPA if you sell physical goods. An AI ad that over-promises texture will look cheap on day two and expensive on day twenty. If you cannot see that, you will scale the liar. CPA is the verdict; contribution after returns is the grown-up verdict. For subscriptions or replenishment, a 7-day CPA will lie in the other direction. Note the window next to the number. AI ads that discount hard look cheap until the second cycle. Put that next to the verdict so finance and creative argue about the same period.

Frequency is how you avoid firing a still-good AI ad. When hook rate falls while frequency climbs in a 7-day window, you have fatigue or an audience ceiling — not a sudden failure of the generator. Ranked sixth because it is a maintenance metric. It tells you to refresh the first three seconds or to widen delivery, not to throw away the angle. Ranked sixth because it saves winners. AI volume makes it tempting to kill a file the day hook dips, when the audience simply saw the same first frame too often. Read the slope before you brief a new concept, or you will pay to rediscover the same angle.

Chart hook rate against frequency for the winning file. A parallel drop in CTR with stable hold often means the open is burned. Fix with a new hook on a frozen body — the cheapest AI trick you have — before you brief a new concept. If frequency is high and hook is flat, you may simply be out of cheap people; scaling tactics are the play, not a new avatar. Set a frequency trigger that creates a task, not a panic: at 2.0–2.5 in a 7-day window, queue a hook recut. If you wait until CPA breaks, you are late. The recut should reuse the body so you still know what you tested.

AI volume can fake freshness. Ten near-duplicate takes do not reset frequency if the first frame is the same face in the same room. Change the open, the setting, or the proof device. Track unique first frames, not unique file names. Count unique first frames in the live mix, not ads in the campaign. Six ads with the same hoodie and room are one frame. AI makes this failure cheap and common; the metric only works if you are honest about sameness. If you cannot tell the thumbnails apart at a glance, Advantage+ cannot either, and frequency will climb on a ghost of one ad.

AI creative has a hidden metric: dollars (media plus credits plus founder hours) per unique concept that reached a spend floor. If you spent $400 to learn nothing because twelve files were the same hook in different shirts, your generator is not cheap. Ranked seventh because it is an ops number, but it is the number that decides whether batch tools are actually saving you money. Ranked seventh as an ops diagnostic for people who think generated ads are free. Media is still the bill, but credits and founder hours belong on the same line or you will 'save money' by producing a folder you never launch. Unique concepts that reach a spend floor are the denominator that matters.

Count unique first-frame-plus-angle pairs, not exports. A 2×2 that finishes is one cheap experiment. A folder of 30 unlabelled renders that never launch is a credit fire. Put credits and hours on the same sheet as media; founders under-count their Tuesday. The goal is to drive this cost down without collapsing isolation — cheaper attempts at the same messy test is not progress. Review this number in the same meeting as CPA, not in a tools meeting. If cost per concept is high because nothing launches, the fix is the Friday kill list, not a cheaper generator. If it is high because every concept is a twin, the fix is the matrix prompt.

This metric also kills tool sprawl. Three subscriptions that each need a learning week are more expensive than one ad-first generator and a 2×2. If cost per concept is rising while CPA is flat, you are producing for the timeline, not the auction. Include the hour you spent prompting. Founders hide that cost and then think a $29 tool is the line item. Two hours at founder rate plus unused renders is often the real price of a 'cheap' week. Put that number next to CPA once; it usually ends the tool-shopping.

ROAS is an account outcome. Comments are a social object. Neither tells you whether to recast the avatar. Ranked last because AI ads that look 'alive' in the comments can still miss the 1% CTR bar, and a blended ROAS number will hide a hero file behind a prospecting loser. Use ROAS in finance reviews. Do not use it in the creative stand-up. Ranked last because both numbers are real and still the wrong creative lever. A blended ROAS screenshot cannot tell you to recast, and a comment thread full of fire emojis cannot tell you the close is shy. Use them in the meetings they belong to, then return to hook, hold and CTR to decide the file's fate.

If you must glance at ROAS, split prospecting from retargeting and split offer from creative, or you will scale a branded-search effect and call it UGC. Engagement can be a leading clue — a comment that repeats the objection is next week's script — but volume of likes is not a kill rule. Watch time without clicks is a nice video. You are not running a nice-video budget. If leadership demands a ROAS slide, show prospecting versus retargeting as two rows and refuse a single blended crown. That is the minimum honesty. Creative still does not get judged on that slide.

The salvage: steal qualitative signal from comments, then go back to hook, hold and CTR to decide. Never promote an ad because the avatar got compliments. Complements are not conversions, and generated faces collect compliments from people who will never buy the SKU. Mine comments for objections and quotes; then put those sentences into rank 1's brief. That is the only engagement loop that pays. Do not raise budget because an avatar was called cute. Compliments are not a conversion event, and generated faces collect them from people who will never see the checkout.