Cuttable's 280,000 Ads Show Why Volume Is the Wrong Goal

AI is making it much easier to produce, place and optimise marketing creative. Cuttable's analysis of 280,000 ads suggests that output volume is a poor measure of progress: 1% of the ads generated half of attributed revenue, while more than a third generated nothing.¹

That number makes a useful lens for almost everything that happened in AI in marketing this week. The tools are getting better at making assets, assembling media buys, moving campaign work between systems and recommending what a team should do next. The result is not less need for marketers, but less excuse to measure productivity by how much stuff the machine produced.

Volume is losing its halo

Cuttable's dataset covered more than US$260 million in Meta spend and found that roughly 42 ads were needed to identify one that materially improved business results.¹ That does not make creative volume useless, because testing needs variation, but it changes what the volume is for. The output only earns its keep when the system learns from it, which means performance history, customer response and brand assets matter as much as the generation step.

This is where a lot of AI content marketing still gets the economics backwards. A tool that turns one brief into 100 pieces can look productive on a dashboard while creating more review work, more weak variants and more opportunities for the brand to become forgettable. If 99 of those pieces teach the system nothing, the team has automated production without improving the decision about what deserves to be produced.

The UK CMO data published this week points in the same direction from a different angle. Censuswide found reported generative AI use among 500 UK CMOs had fallen from 89% last year to 62% this year, while 64% of those still using it said it exceeded expectations, up from 58%.² A drop in usage alongside a rise in reported value looks less like retreat and more like teams removing AI from jobs where it adds little, then concentrating it where the gains are real.

That is a more useful maturity signal than adoption moving endlessly towards 100%. Nobody gets extra points because every task contains an AI step, and a marketing team should be perfectly happy to stop using a tool where the human version is quicker, clearer or more accurate. The better question is whether the work that remains automated is becoming measurably better.

There is also a hidden cost in creative abundance: selection. If a team can produce 20 variants in the time it previously took to make three, somebody still has to decide which ones are faithful to the brief, which claims are supportable and which differences are meaningful enough to test. Generation can become cheaper while review becomes the new bottleneck, especially when the variants differ cosmetically rather than strategically.

That is why experimentation needs a stopping rule as much as it needs scale. A useful system should be able to retire weak patterns, preserve the ingredients that repeatedly work and distinguish a real performance signal from random variation. Otherwise the team ends up paying for cheap generation twice, first in compute and then in human attention.

Machines are moving downstream

The week's product launches show why the volume question is getting more urgent. Google is letting Gemini curate personalised media bundles inside Display & Video 360, while Search Ads 360 uses Gemini to consolidate insights and recommend actions.³ Sprinklr launched Marketing AI Agents designed to move from live customer signals into governed campaign execution, so the model is no longer sitting outside the work waiting for somebody to copy and paste its answer.⁴

Peec AI is making a similar move in AI search. Its Agent and Actions products can investigate why visibility changed, inspect cited sources, identify technical or content problems, then turn a finding into a brief or task, with approval required by default.⁵ Adclear is putting compliance checks inside Claude and ChatGPT for regulated marketing teams, while named human reviewers retain approval and the audit trail records what happened.⁶

These are all more consequential than another caption generator because they remove hand-offs. The machine is increasingly doing the assembly work between signal and action, which can save marketers from moving data, recreating context and configuring the same steps repeatedly. It also means a poor instruction, a bad data source or an incorrect assumption can travel further before somebody notices it.

Human approval therefore matters more when the system can act, not less. A weak draft in a chat window is cheap to ignore, but an agent that changes targeting, creates a task, modifies campaign settings or routes a customer response has a wider blast radius. The sensible design pattern in this week's launches was not total autonomy; it was context, action and a visible point where a person still decides.

For smaller teams, this is where AI should become genuinely useful. Time spent on repetitive production steps can fall without handing over the choice of audience, offer, claim or brand position. The best automation removes the mechanical distance between an idea and execution while leaving the consequential decision where the business can still own it.

The governance challenge also changes when more steps are connected. Teams need to know which actions are reversible, which require approval and what happens when the model is uncertain, because "human in the loop" is too vague if nobody can say where the loop actually sits. A content draft can tolerate experimentation; a compliance decision, budget change or customer-facing reply needs a much clearer boundary.

This is also why cost controls matter once agents become routine infrastructure rather than occasional assistants. A system that uses an expensive model for every trivial operation can waste money even when its outputs are correct, while deterministic software may handle some tasks better and more predictably. Good automation should reduce unnecessary work, including unnecessary AI work.

Better inputs beat more outputs

Meta's holiday guidance gives this principle a blunt numerical form. The company says an ad set needs roughly 50 conversions a week to learn effectively, and that purchases cost about 28% more when the system is below that threshold.⁷ More automation does not compensate for weak signals; it makes the quality and quantity of those signals more important.

Creative systems are starting to reflect the same idea. Canva's new Craft model was trained on human-designed templates and plans elements such as scale, hierarchy and spatial relationships before generating an image, a choice that makes sense for a product whose job is design rather than photorealistic novelty.⁸ The useful benchmark is not whether a model can make one impressive image, but whether it can make the fiftieth asset behave like part of the same brand.

Two campaigns this week made the contrast even clearer. Plusnet built its new social-only campaign around Plusso, a reusable pink character that can travel across Meta, TikTok, YouTube, Reddit and Snapchat, while RACV's agency started by digging through archives and listening to customer-service calls before creating its new brand platform.⁹¹⁰ Both approaches give faster production something distinctive to work from: a recognisable visual asset in one case, and organisational memory and customer language in the other.

That distinction matters for any Instagram content strategy using generative tools. A business can automate cropping, resizing, caption drafts, scheduling and variation, but its own photography, product details, customer language and visual habits are what stop the feed from looking interchangeable. Instagram marketing AI is most useful when it compounds source material the business already owns, rather than replacing that material with a generic approximation.

The same test should apply to Instagram AI content management and broader content systems. Ask what information the tool remembers from one post to the next, what performance evidence it uses, and which brand choices stay fixed when the format changes. Counting generated posts tells you almost nothing about whether the next post will be more recognisable or more effective than the last one.

There is a second reason the source material matters: it gives the system something stable to preserve while formats keep changing. Plusnet can make ten versions of Plusso without inventing ten different brand personalities, and RACV can adapt customer language without losing the organisational memory that made the campaign specific. That is a much stronger foundation for Instagram content generation than asking a model to rediscover the brand from a short prompt every Monday.

The feedback loop should also connect creative performance back to those source ingredients. If a product close-up repeatedly outperforms a lifestyle composition, or a particular customer phrase consistently earns attention, the system should learn that pattern without flattening the entire feed into one repeated formula. The aim is accumulated evidence with enough human judgement to know when a pattern is useful and when it has become stale.

How do brands keep their voice consistent with AI?

Brand consistency is becoming a systems question because AI now touches more of the path between source material and customer. The easiest answer is also the least glamorous: define the business facts, approved assets, audience, tone and boundaries once, then make those inputs available wherever generation or automation happens. A prompt can imitate a tone for one task, but a repeatable workflow needs context that survives beyond a single chat.

That becomes more important as new distribution surfaces open. OpenAI expanded ChatGPT advertising into 26 more countries this month, taking the product to 89 markets according to Reuters.¹¹ A conversational ad can sit beside advice that feels personal and useful, so the creative is entering an environment where trust is already part of the experience.

The consumer research published alongside that expansion should make marketers cautious about treating synthetic creative as a free production shortcut. Cint reported that 74% of US consumers surveyed believed brands and creators should disclose AI use in advertising, 44% said undisclosed AI-generated advertising would reduce trust, and 46% specifically disliked AI-generated people in ads.¹² Those figures do not mean brands should avoid AI production, but they do make provenance part of creative performance.

That point is easy to miss when the production metric is cost per asset. A synthetic spokesperson may be cheaper than a shoot, and a generated product scene may be faster than photographing the real thing, but the saving disappears quickly if the customer doubts what they are seeing. The practical dividing line is whether AI is helping the brand express something true more efficiently or manufacturing evidence that the customer is expected to believe.

For small businesses, reality is often an underused asset. The owner, shop, kitchen, product, treatment room, packaging, customers and local setting are all difficult for a competitor to copy exactly. An Instagram AI content tool should help turn those specific things into more usable formats, not make them look like everyone else's stock imagery.

The metric worth keeping

The most useful consequence of cheaper generation is that teams can test more ideas without spending more on every iteration. Cuttable's numbers show why that capability needs a feedback loop rather than a production target: if only a small fraction of creative drives a large share of results, the job is to identify what worked, understand why, and carry that evidence into the next round.¹ A system that produces twice as much but learns nothing has increased throughput, not capability.

The same logic applies beyond advertising. If Gemini assembles a media buy, Peec turns a visibility drop into a task, Sprinklr moves from signal to campaign action and Canva makes design production easier, marketers can spend less time on assembly. That creates room for the work software still cannot infer safely from a blank prompt: deciding what the business should promise, which customer matters, what evidence is credible and what should never be automated.

So the better weekly metric is not assets generated, prompts sent or hours spent inside an AI tool. Measure the percentage of recommendations accepted, the cost per useful task, the lift from the creative that survived testing, the consistency of the brand across formats and the number of mechanical hand-offs removed. When production gets cheap, judgement becomes visible in the difference between what can be made and what is worth making.

Sources