Your AI Doesn't Know What Worked Last Week

Here's a number that should stop every marketer cold: only 19% of content marketers track AI-specific KPIs.

That means 81% of businesses using AI to produce marketing have no systematic way of knowing whether the AI output is getting better, getting worse, or just generating expensive noise. They're producing more content than ever. They have no idea if any of it works.

Software engineers figured this out months ago. The teams getting extraordinary results from AI code aren't the ones generating more code. They're the ones that built verification systems around the AI's output. Systems that catch problems, measure quality, and feed results back into the next cycle.

Marketing hasn't caught up yet. And the gap is costing real money.

The Software Engineers Got There First

In early 2026, a company called StrongDM published something that made the entire software world pay attention. Their AI team had a rule: "Code must not be reviewed by humans."

Not because they were reckless. Because they'd built something better than human review.

Instead of having engineers read AI-generated code line by line, they built verification systems. Automated scenarios that tested whether the software actually worked. Digital replicas of the services their code needed to talk to. Holdout tests that the AI couldn't see during development, so it couldn't cheat.

The result? A three-person team shipped production-quality security software in months. Software that would have taken a traditional team years.

The lesson wasn't "trust AI blindly." It was the opposite: build systems that verify AI output so rigorously that you don't need to review every line. The humans moved from doing the work to designing the system that checks the work.

Every, a technology media company, codified this approach into a framework they call Compound Engineering. Their finding: the teams getting 300-700% productivity gains from AI coding agents spend 80% of their time on planning and review, and only 20% on the actual work. The leverage isn't in generating faster. It's in evaluating better.

Marketing Skipped the Verification Step

Now look at how most businesses use AI for marketing.

The workflow looks like this: open ChatGPT (or Claude, or Jasper, or whatever tool you've chosen). Type a prompt. Get output. Maybe tweak it. Publish. Move on.

That's an open loop. Input goes in. Output comes out. Nothing about the result feeds back into the next cycle.

There's no system asking: did that email get replies? Did that ad produce leads that actually converted? Did that blog post get cited by anyone, or did it vanish into the ocean of content nobody reads?

The data confirms this is widespread. HubSpot's 2026 State of Marketing Report found that 33% of marketers say measuring ROI is their single biggest challenge. Not creating content. Not finding audiences. Measuring whether any of it worked.

Meanwhile, AI-generated content crossed a threshold in late 2024: it now accounts for over 50% of all online text. AI articles rose from 2.2% in January 2020 to 51.7% by May 2025. The internet is drowning in AI content. Most of it has no feedback loop. Most of it is, by any honest measure, average.

This is exactly what you'd predict from the research. A 2024 study published in Nature by Shumailov et al. demonstrated that when AI models are trained recursively on their own outputs, quality degrades by up to 40% in factual accuracy within just three generations. Rare patterns disappear. Distinctive ideas get smoothed out. Everything drifts toward bland averages.

The researchers call it model collapse. We might call it something more familiar: your marketing is starting to sound like everyone else's.

We've written about this before. AI made your marketing faster, cheaper, and identical to everyone else's. The feedback loop problem is the mechanism that explains why it happens.

The Three Feedback Loops That Actually Matter

Avinash Kaushik, one of the sharpest minds in marketing analytics, frames the problem with a hierarchy he calls Smart KPIs: Accountability over Outcomes over Activity.

Most AI marketing tools optimise for activity. Words produced. Ads created. Emails sent. The dashboard shows you did things. It doesn't show you whether those things mattered.

Kaushik's research puts a number on the stakes: 55-70% of the success of any campaign is the creative. Not the targeting. Not the bidding strategy. Not the platform selection. The creative. The actual message your audience sees.

If creative drives the majority of outcomes, and AI is producing most of your creative, and you have no system to measure which creative performs, you're optimising the wrong variables entirely.

Here are the three feedback loops that separate businesses compounding their AI marketing results from businesses just compounding their volume.

Loop 1: The Creative Testing Loop

At the large advertiser tier, the average account ships 11 ads per week and produces 1.75 winners per month. Top performers ship 31 ads per week and produce 5.99 winners per month.

The difference isn't just volume. Top performers have a system to identify winners faster and retire losers earlier. They run structured tests. They track performance at the creative level, not just the campaign level. They feed what works back into the next batch.

Without this loop, more AI-generated ads just means more expensive mediocrity. You're paying to discover that 92-96% of your creative doesn't work, but you're never feeding that discovery back into the process.

Loop 2: The Conversion Signal Loop

Here's where Les Binet and Peter Field's research becomes critical for any business running paid ads.

Binet and Field's analysis of the IPA effectiveness database shows that the market has inverted their recommended 60/40 split between brand and activation. Brands now allocate 68.8% to performance and 31.2% to brand, with documented declines in long-term effectiveness as a result.

Why? Because performance metrics are the easiest to measure. Clicks, impressions, cost per lead. These are the numbers that show up in dashboards. AI tools optimise for them automatically.

But these are activity metrics, not outcome metrics. A click doesn't mean a customer. An impression doesn't mean someone remembers you. A lead form submission doesn't mean revenue.

The conversion signal loop means tracking what happens after the click. Did that lead answer the phone? Did they show up to the meeting? Did they sign a contract? How much was it worth?

When you feed real conversion data back into your ad platforms, the improvement compounds. As we've explained in our guide on why the post-click gap is where your budget disappears, the better the conversion data you provide, the better the algorithms optimise. Better optimisation produces better leads. Better leads produce better conversion data. The loop accelerates.

Without it, the AI is optimising for the cheapest click, which is almost never the most valuable customer.

Loop 3: The Distinctiveness Loop

Byron Sharp's research at the Ehrenberg-Bass Institute established that brands grow through mental availability: being the brand that comes to mind when someone enters the buying situation. And mental availability is built through distinctive brand assets used consistently across every touchpoint.

AI, left to its own devices, erodes distinctiveness. It's trained on the average of everything. It will naturally drift toward the generic. The phrases every competitor uses. The stock-image aesthetic. The safe, forgettable middle.

The distinctiveness loop means periodically auditing your AI output against your brand standards. Does the copy sound like you or like every other business in your category? Are the visual elements maintaining your brand's distinctive codes? Would a customer who knows you recognise this as coming from your business?

This is harder to automate than the other two loops. But it's the one that prevents what the model collapse research warns about: the slow drift toward indistinguishable average.

We've explored this dynamic in depth. Your marketing used to be expensive. That was the point. When everyone can produce the same volume of content at the same low cost, the signal value of that content drops to zero.

Why Activity Metrics Are the Wrong Speedometer

Sam Tomlinson's work on creative effectiveness and portfolio theory gives this problem a useful frame. He argues that most marketers evaluate all creative through the same lens, regardless of its role. They measure brand stories with the same KPIs as landing pages.

This is like measuring a savings account and a growth stock by the same criteria. One is supposed to be stable and reliable. The other is supposed to take calculated risks for outsized returns. Judging both by last quarter's return tells you nothing useful about either.

Tomlinson applies financial portfolio theory directly: your marketing needs core holdings (proven creative that reliably converts), growth plays (new concepts being tested), and speculative bets (bold creative that might fail but might also redefine your brand).

AI is excellent at producing all three categories. But without portfolio-level tracking, you can't tell which category each piece of creative belongs to, and you can't make informed decisions about what to produce next.

Kaushik puts it more bluntly. His Smart KPIs framework draws a clear hierarchy:

What This Means for Your Business

The fix isn't complicated. It's just not automatic.

The software engineers at StrongDM spent their time building "holdout tests" that the AI couldn't see during development. The marketing equivalent is simple: hold some of your measurement separate from the AI's optimisation loop. Track metrics the AI isn't optimising for. That's where you'll spot the drift before it costs you.

AI didn't create the feedback loop problem. Businesses have been running marketing without proper measurement for decades. But AI made the problem urgent, because the volume of unmeasured output went from manageable to overwhelming.

The businesses that build the loops now won't just produce better AI marketing. They'll compound their advantage every month while their competitors produce more of the same.

Further Reading

• Creative Benchmarks 2026 (Motion) - Data from 550,000+ ads and $1.3B in spend on what separates top performers
• Smart KPIs: Accountability Over Outcomes Over Activity (Avinash Kaushik) - The measurement hierarchy every marketer should adopt
• Compound Engineering: The Complete Guide (Every) - How software teams build verification systems around AI output
• Effectiveness in the Digital Age: Insights from Les Binet (WARC) - Why the shift to performance metrics is undermining long-term growth
• The Curse of Recursion: Training on Generated Data Makes Models Forget (Nature, 2024) - The model collapse research that explains why AI content drifts toward average