Why are AI insights often inaccurate in marketing?

AI models work with the data they receive. If the data is incomplete, outdated, or biased, the output will reflect those gaps. The difference from traditional analytics is that AI presents results with a confidence level that discourages verification.

How do you define an effective ICP for a marketing team?

A good ICP is defined at the product level, not just the company level. It combines account fit (industry, size, budget), contact fit (role, decision authority), and engagement signals. The fit score provides context for behavior.

How do you avoid personalization based on stale signals?

Ask yourself "is this relevant right now?" before every automated message. Set expiration rules for behavioral signals and don't be afraid to send nothing if you don't have a current signal.

Why Marketing Decisions Can't Keep Up with Data

Every week brings new tools, new data, new possibilities. But the gap between what we can measure and what we actually understand isn’t closing. It’s widening.

Over the past few days, I’ve read several articles that seem unrelated at first glance: AI-generated insights that backfire, leads without a clear definition, personalization running on autopilot, and an ad format viewers can skip. But underneath, the story is the same: the tools evolve, the questions we ask them stay the same.

We see this in Romania too. Marketing budgets grow, data access is better than ever, but decisions still rely on instinct, reflex, or reports that nobody double-checks. The tools get smarter. The questions we ask them don’t keep up. Here’s what caught my attention this week and why it matters for anyone running campaigns right now.

AI insights have a credibility problem

A GWI study published by Marketing Dive shows that 80% of industry professionals frequently encounter inaccurate or misleading AI insights. And yet, only 36% say they verify those insights most of the time.

The numbers pile up: 72% agree AI can make something seem more relevant than it actually is, 63% have accidentally used insights that turned out to be outdated, and over half of decision-makers have made a business decision based on incorrect AI data. Of those, 27% report direct financial losses.

What strikes me here isn’t that AI gets things wrong. Any model working with imperfect data will. The problem is that people don’t check. They trust the output without questioning the input. And that’s the exact same mistake they made before AI, with Google Analytics data or market reports. AI just amplifies it.

Seventy-five percent of respondents believe AI will never replace direct human research. That’s an important statement, because it suggests the industry already knows where the problem is. It just doesn’t act on it. The gap isn’t in awareness. It’s in process.

For any marketing team, the lesson is straightforward: no automatically generated insight is a validated truth. It’s a hypothesis. And hypotheses get tested, not dropped straight into your content strategy.

Before you ask for better leads, define what that means

An article on MarTech.org asks an uncomfortably simple question: before you ask for better leads, have you defined what better actually means?

Most sales teams complain about lead quality. Most marketing teams complain that sales doesn’t follow up. And yet, surprisingly few have an ICP (Ideal Customer Profile) defined at the product level, not just the company level.

The article draws an important distinction between three signals: how well an account fits (account fit), how well a contact fits (contact fit), and how engaged they are (engagement). A highly engaged contact at a company that doesn’t fit your profile isn’t a good lead. It’s a noisy one. An account scoring 90 with moderate engagement deserves attention. One scoring 45 with the same activity belongs in nurturing, not on a sales call.

At difrnt. we see this consistently: the clients who grow fastest are the ones who clearly defined who they’re looking for before investing in traffic. The rest generate high volumes with low conversions and then wonder what went wrong. The definition comes before the performance, not after. Without it, scaling traffic only scales confusion.

Personalization works. Until it expires.

Also on MarTech.org, an article about relevance raises a problem few people discuss: relevance has an expiration date.

The numbers are compelling: 81% of consumers ignore messages that aren’t relevant to them, and 1 in 4 says a generic message makes them less likely to buy. On the other hand, 71% of consumers want brands to learn from their behavior over time, not just remember what they bought last month.

The problem? Marketing automations often run on stale signals. You searched for a vacation three months ago, and the newsletter still sends you destination deals. You bought a laptop, and retargeting follows you with laptops for another month. Automation is excellent at execution, but it doesn’t know when to stop. And the longer a signal sits in the system without being questioned, the more confidently the machine acts on it.

Only 15% of organizations manage to deliver experiences that surprise consumers, according to Adobe. And 75% of practitioners say real-time personalization remains a major challenge. Not because they lack data, but because they lack temporal context. The right question isn’t what did the customer do, but what’s relevant to the customer right now. Sometimes the answer is: nothing. And that’s OK.

Twitch tests ads you can skip. And it works.

Marketing Dive reports that Twitch has launched two new ad formats: shoppable ads (add products to your Amazon cart directly from a stream) and skippable ads (brands only pay if the viewer watches more than 30 seconds).

The testing results are interesting: Razer achieved a 2.44x higher ROAS with skippable formats compared to standard Twitch campaigns. A third of the customers who engaged were new to the brand. And nearly 70% of Twitch’s audience is Gen Z or millennial, a segment many brands chase but don’t know how to reach without looking forced.

Why does this matter beyond Twitch? Because it validates a simple principle: forced attention isn’t attention. When you give users control, you actually get more engagement, not less. It’s a lesson Google learned with YouTube TrueView a decade ago, and Twitch is now confirming it. The skippable format doesn’t lower the bar for quality. It demands it. And for brands willing to create ads people actually want to watch, the economics are significantly better.

Four different topics, one common thread: the gap between what’s available to us and what we do with it keeps growing. We don’t lack tools. We lack the right questions, clear definitions, and the courage to stop an automation running on expired data. That’s not a tech problem. It’s a leadership problem. And it starts with asking whether the question you’re optimizing for is still the right one. The tools will keep getting better. The real competitive advantage is in adapting the campaigns and the thinking behind them to the reality in front of us.