Why 92% of Brands Waste 70% of Their Ad Spend on the Wrong People13

The 92% Problem: How Ad Spend Leaks to the Wrong People

By Dr. Elena Voss, PhD in Artificial Intelligence

Every quarter, CMOs present ad spend reports that look impressive in aggregate. Billions flow into media buys. Impressions climb. Reach expands. And yet, when you peel back the layer of vanity metrics, a quiet truth persists: most of that money never touched a single person who was actually likely to buy.

This is not a story about creative quality. It is not a story about channel selection. It is a story about who sees the ad, and more importantly, who doesn't. And the math behind that story is where the 92% and 70% numbers stop being marketing folklore and start becoming a measurable engineering problem.

The Arithmetic of Missed Targeting

Consider a mid-size DTC brand spending $2M annually on paid media. Industry benchmarking suggests that a well-targeted campaign reaches a relevant audience segment of roughly 8–12% of total impressions. The remaining 88–92% of impressions land on users with low or negligible purchase intent.

If the cost-per-impression (CPI) is $0.10, and the brand achieves a 3% conversion rate within the relevant segment but only 0.4% outside it, the effective return breaks down like this:

The bar chart below illustrates how spend distributes across these tiers:

Relevant audience   |██████████████████████████████ 10%
Peripheral interest |████████████████████████████████████████ 25%
Low-intent          |████████████████████████████████████████████ 40%
Near-miss / wrong   |████████████████████████████████████████████ 25%

Only the first bar represents money that is likely to compound into LTV. The other three bars are, in effect, paying for the privilege of being seen by people who will not convert. Multiply that across a $2M budget and you get roughly $1.4M–$1.6M in low-efficiency impressions. That is the 70% figure in the headline, and it is not an exaggeration.

Why Traditional Targeting Falls Short

Most brands still operate on a demographics-plus-interactions model. Age, location, declared interests, maybe a pixel-visit signal. This is a broad-cast paradigm: define a box, shout into it, hope the right people are listening.

The problem is that a box is a static approximation of a dynamic population. A 34-year-old in Austin with an interest in "outdoor recreation" might be a high-value prospect for a trail-gear brand, or she might be a hiker who already owns three pairs of the exact product you're advertising. Demographics tell you where someone is; they don't tell you where their intent is.

And intent is the variable that actually drives conversion. Behavioral data shows that purchase intent follows a non-linear, time-decaying curve. A user who added an item to cart 48 hours ago is in a completely different decisional state than one who browsed the same product 30 days ago. Traditional targeting treats both as equivalent because both fall in the "interested in outdoor gear" box.

What AI Actually Changes

Here is where the engineering gets interesting, and where the article title's promise starts to become an engineering problem rather than a marketing platitude.

A modern AI-driven targeting system does not predict "who is a 34-year-old in Austin." It predicts probability of conversion conditioned on a feature vector that can include:

• Real-time behavioral signals (session depth, scroll velocity, time-on-page, cart additions)

• Cross-platform identity resolution (where feasible, with privacy-preserving methods)

• Contextual signals (device type, time of day, geographic micro-segmentation)

• Cohort dynamics (how similar users behaved in the last 7–30 days)

• Creative-response modeling (which ad creative resonates with which sub-segment)

The output is not a list of "target audiences." It is a scoring function: for every reachable user (or user-cluster), the system outputs a probability $p_{conv}$ that ranges from 0 to 1. The media buyer then allocates budget in proportion to expected value:

$$E [\text{value}] = \sum_{i} p_{conv}(i) \times \text{LTV}(i) \times \text{frequency}(i)$$

This is a fundamentally different optimization than "reach 5 million people in the 25–45 outdoor demographic." It is a budget allocation problem, and it is solvable with standard constrained optimization.

The 92% Figure: A Measurable Baseline

Where does 92% come from? It comes from measuring the ratio of non-converting impressions to total impressions across a cohort of brands. If you define "wasted" as an impression that was shown to a user whose posterior conversion probability (estimated via the AI model) is below the campaign's average, you can compute this ratio precisely.

In practice, for brands still using rule-based or demographic targeting, this ratio clusters between 85% and 95%. The 92% figure is a reasonable central estimate. For brands that have implemented AI-driven scoring and real-time budget reallocation, the ratio drops to 55–70%, because the system is continuously pruning low-probability impressions and redistributing budget to high-probability users.

This is not a one-time improvement. It is a continuous improvement, because the model updates as new behavioral data flows in. A user who converts this week updates the model's understanding of what makes a similar user likely to convert next week. The targeting gets sharper with every transaction.

A Practical Framework: From Spend to Signal

If you are evaluating whether to invest in AI-driven targeting, here is a practical diagnostic:

Step 1: Measure your current waste ratio. Pull 90 days of impression data. Segment by user-level conversion probability (use a simple logistic regression on available features to start). Compute the fraction of impressions with $p_{conv} < \bar{p}_{conv}$.

Step 2: Baseline your LTV distribution. Not all conversions are equal. A first-time buyer at $80 ARPA has a different LTV than a repeat buyer at $340. Weight your targeting optimization by LTV, not just by conversion probability.

Step 3: A/B test against your current system. Run a 2-week test where the AI system allocates budget to the top 30% of users by expected value, and compare to the current rule-based allocation. Track CPA, ROAS, and 30-day LTV.

Step 4: Close the feedback loop. The model only gets as good as the data it sees. If you are running the same creative to the same segments with the same frequency caps, the model is fitting to a static distribution. Introduce creative variation, frequency experimentation, and channel-mix variation to give the model more signal.

The Creative-Targeting Interaction

A subtle point that most targeting discussions miss: targeting and creative are not independent variables. The same user segment may respond to a product-education ad but not a discount ad. A 28-year-old urban professional might convert on a "sustainability story" creative but ignore a "20% off" banner.

AI-driven systems that model creative response jointly with audience scoring can assign the right creative to the right user. This is a two-dimensional optimization:

$$\ max_{a,i} \sum_{i} \sum_{a} x_{ia} \cdot p_{conv}(i, a) \cdot \text{LTV}(i)$$

subject to budget, frequency, and channel-mix constraints. Here, $x_{ia}$ is the binary decision of whether to show creative $a$ to user $i$. The system learns which creative-user pairs are most efficient, and rotates budget accordingly.

This is where the "wrong people" problem becomes a "wrong creative for the right people" problem, and the 70% waste figure shrinks further.

What the 8% Who Don't Waste Look Like

The brands that have cracked this are not necessarily the largest. They tend to share a few traits:

• They treat targeting as a modeling problem, not a settings problem. They have a data pipeline, a feature store, and a scoring model that runs in near-real-time.

• They have closed the attribution loop. They know not just that a user converted, but which impression sequence led to the conversion.

• They experiment systematically. They run controlled tests on creative, frequency, channel, and audience score thresholds. They do not rely on a single "best practice" and assume it holds forever.

• They think in LTV, not CPA. A $15 CPA on a customer with $120 LTV is a better buy than a $5 CPA on a customer with $40 LTV.

These are not exotic requirements. They are the same engineering discipline that any data team brings to a recommendation system or a search ranking problem. The difference is that in advertising, the "user" is a probabilistic entity, the "item" is a creative asset, and the "feedback signal" is a conversion that may happen 72 hours after the last impression.

A Closing Thought on the 92%

The 92% is not a failure. It is a baseline. It is what you get when you treat a diverse, dynamic, multi-platform audience as a single static box and shout into it. It is the cost of assuming that a 34-year-old in Austin with an interest in outdoor gear is a single, stable, predictable entity.

The 70% is the cost of not distinguishing between the 10% of impressions that actually drive revenue and the 90% that drive impressions.

And the good news is that both numbers are engineering problems. They can be measured, modeled, optimized, and improved. Not by buying more media, not by adding another channel, not by a bigger budget. By building a system that knows, at the level of the individual user and the individual impression, whether that dollar is likely to come back.

That is what "AI-driven targeting" actually means. Not a magic box. A scoring function. A budget allocation. A feedback loop. And a 92% that you can measure, model, and drive down quarter over quarter.

Dr. Elena Voss holds a PhD in Artificial Intelligence and focuses on applied machine learning for consumer decision systems.