AI Creative CPM Benchmarks for DTC Brands

Framework 1: The AI Vidia CPM Diagnostic

The CPM Diagnostic is the model AI Vidia runs on every new account before the first media plan gets signed off. Five checks, one verdict, and the output is the variant supply target written into the weekly brief. Every step uses a number from the method table above, so the verdict is checkable in the client's own Ads Manager.

1. Step 1. The variant supply audit. Pull the last 90 days of ad account data and count unique creatives with spend per ad set per week. Two reference points: enterprise advertisers in Motion's 2026 dataset launch 18.8 new creatives a week, and the Andy Okay account runs at 50 new ads a week. If your count sits in the low single digits, the auction is choosing from a set too small for a 5 percent winner rate to produce a winner at all, and no bid change fixes that.
2. Step 2. The frequency-to-CPM read. Chart frequency and CPM per placement, week by week, on the same axis. The week frequency turns up on a fixed audience is the week to expect CPM to follow, and the placement where it turns up first is the one to feed first. Write down your own account's turning point per placement; that number, not a vendor's, is your refresh cadence.
3. Step 3. The learning phase check. Read the Delivery column at ad set level. Meta's help centre states that an ad set needs about 50 optimisation events in a 7-day window to leave the learning phase, and every ad set that fails to get there pays unstable results while it waits. If most ad sets are still learning, consolidate before adding variants; more creative in a structure that cannot exit learning only spreads the same events thinner.
4. Step 4. The placement mix audit. Map spend against unique creatives per placement. A placement carrying a large share of spend but a small share of the weekly variant supply is the fastest way to a rising CPM, because it is the placement where frequency climbs first. Reels and Stories placements consume vertical variants faster than feed does, which is why AI Vidia delivers ad creative in 9:16.
5. Step 5. The CPM target. Set a written per-placement CPM target from your own last-90-day median and lower quartile, not from a vendor table, and put it in the weekly brief so the media buyer and the creative team share one number. Accounts that skip this step keep treating CPM as a media buying problem long after the diagnostic shows it is a supply problem.

Run the diagnostic once and the next 90-day plan writes itself. Accounts failing Step 1 do not need a new model or a new agency; they need a written batch cadence and enough variants to fill it. Accounts failing Step 3 need fewer, fuller ad sets before they need more creative. Accounts passing the first four checks but never setting a target in Step 5 need the Friday CPM read wired to a Monday rebrief, which is what the second framework covers.

Kevin's take

That distinction is the reason the CPM Diagnostic above starts with the variant supply audit and ends with a target from your own data, not the other way around. Media buying skill still matters on budget pacing and audience structure, but the mechanism shows a CPM gap that closes on the supply side. Fix the supply line first and the bid strategy conversation gets a lot shorter.

Framework 2: The AI Vidia Variant Supply Loop

The Variant Supply Loop is the execution model AI Vidia runs every week on a Performance Retainer. Five days, one batch, and a Friday read that sets the next week's supply target per placement. It is the cadence behind the 50 new ads a week AI Vidia ships for Andy Okay and the 142 ads shipped in 11 weeks for IndianBites, and every number it produces comes from the method table above, read in the client's own Ads Manager.

1. Step 1. Monday audit and brief. Pull frequency, CPM and unique creatives by placement for the prior week and compare each against the target set in the diagnostic. The brief sizes the coming batch to the placement whose frequency turned up first, names the concept families and variants, and the brand lead signs it against the existing brand lock before generation starts.
2. Step 2. Tuesday batch generation. The studio runs the brief through the active model stack, tags every raw variant to a placement and a hook family, and gates the raw batch against the brand lock checklist. What fails the lock never reaches the account.
3. Step 3. Wednesday brand lock pass. Selected variants are trimmed, graded, captioned for sound-off viewing and finished against the lock (palette, framing, product treatment, lighting register). Wednesday's output is the shippable batch, and this is the day the work stops looking AI-generated and starts looking like the brand.
4. Step 4. Thursday 9:16 cuts and staggered delivery. Every shippable variant is cut in 9:16 vertical, named to the ad account convention, and uploaded on a staggered schedule across the week rather than all at once, which is what keeps frequency from spiking the day after a large batch lands.
5. Step 5. Friday CPM and frequency read. The studio pulls CPM and frequency by placement for the week's batch, checks each placement against its target, computes cost per winner, and writes Monday's rebrief. Placements whose frequency is still climbing get first priority in the next batch; placements at or below target hold their supply level.

Run the loop for eleven weeks and you have the IndianBites shape: 142 ads shipped with weekly testing throughout, 2.4x ROAS on the winning cohorts of tested variants, and a conversion rate up 32 percent in the client's own words. Run it for a year and you have the Andy Okay shape: 50 new ads a week and 1,000+ AI ads shipped. Neither is a CPM forecast for your account, and AI Vidia does not sell one; they are what the loop did on two accounts a reader can look up.

Proof from named client accounts

Four results, four named clients, nothing else. AI Vidia publishes no aggregate across them and no studio-wide CPM, because four accounts in four categories do not average into a benchmark; they show what the loop does to whichever metric each account is bought on.

• IndianBites (Copenhagen DTC food brand): 142 AI ads shipped in 11 weeks, 2.4x ROAS on the winning cohorts of tested variants, conversion rate up 32 percent. Full case at the IndianBites case study.
• Andy Okay (art prints): 50 new ads a week, 1,000+ AI ads shipped across statics, UGC-style video and story ads. Full case at the Andy Okay case study.
• Singh Law (legal services): cost per lead down 19 percent over three months on a scalable creative production plan, in the words of Sharandeep Singh. Quote on the AI Vidia home page.
• Sorvan Design: customer base up 93 percent year over year after testing new markets with persona-specific AI creative across channels, in the words of Ivan Asen. Quote on the AI Vidia home page.

CPM is not a bidding problem. It is a frequency problem wearing a media buying costume, and frequency only moves when the creative line ships more than the auction is used to seeing.

What the four have in common is the loop, not a model. A food brand bought on ROAS, a print brand bought on volume, a law firm bought on cost per lead and a design studio bought on new-market customers all got the same thing: enough tested, on-brand variants for the auction to find the winners. The Andy Okay result is the one closest to this page, because 50 new ads a week is a variant supply figure, and variant supply is the input the frequency row responds to.

When each CPM band wins

Pick an in-house design team when the concept count is low, the brand lock is still being written, and a senior designer with prompt experience can own both. Read the frequency row monthly, because a small team's batch cadence is set by its capacity rather than by the account's saturation curve, and CPM climbs quietly when the two drift apart.

Pick a DIY SaaS stack (Synthesia, Runway, Midjourney and their peers in self-serve) when the team owns a written brand lock, has someone to run QA against it, and the calendar can absorb revision cycles. The stack widens the variant surface on one or two placements but rarely sustains the weekly count that keeps frequency flat on every placement at once.

Pick a managed studio when the account needs weekly tested volume in the range Motion's enterprise advertisers run (18.8 new creatives a week) or above, without adding headcount. At Motion's roughly 5 percent winner rate, a line shipping ten creatives a month finds a winner about every other month; a line at 50 a week, the Andy Okay cadence, finds about ten a month. The full video surface sits at AI video ads.

Stay with traditional production only when the category requires hero film with a real face and voice and the brand can absorb a slower cut cadence, which is most of luxury, premium spirits and couture. For every other DTC vertical the method table above tells you where your CPM line actually sits.

The next step

The fastest way to get a CPM benchmark you can trust is to run one on your own account. AI Vidia's Control Test does exactly that: 10 new variants against your current best ad, measured for 14 days in your own Ads Manager on one metric agreed in advance, and CPM read next to frequency is a common choice. Book a 30 minute scoping call at book; bring the last 90 days of Meta and TikTok data and the diagnostic runs on the call. This works especially well for ecommerce brands scaling paid social faster than their creative line can keep up.