Eleven years spent learning what a good audience looks like doesn’t protect anyone from the moment when the top-performing segment flattens out in a single quarter, without anyone in the meeting being able to explain why. The feeling that follows has more to do with the volume of data than with expertise, and this volume has, imperceptibly, turned human pattern recognition into the slowest tool in the room.
The solution most teams arrive at is to hire yet another analyst. It would be more appropriate to change the very system that identifies the patterns. Modern AI marketing powered by deep learning algorithms learn from every ad the account has ever served and continuously update this understanding, without relying on assumptions about the buyer’s next step. The difference may seem small, but it changes the entire rhythm of planning.
What deep learning does that rule-based targeting cannot
The typical setup is easy to describe: someone on the team built audience rules (age group, previous purchases, three site visits), and these work until behavior changes—which it always does. A rule-based system can only detect patterns that a human already knew to look for. Deep learning reverses this order, because it takes the unordered behavioral data—with its rejections, half-abandoned shopping carts, and odd browsing sessions at 2 a.m.—and discovers structures within it that no one had previously noted. The platform thus forms its own hypotheses, rather than relying on those provided by the team.
The vendors’ claims, however, deserve to be challenged on one point. With small data volumes, deep learning isn’t automatically superior to a well-calibrated rules-based engine. A regional brand running 40,000 ad impressions per month will achieve better results with a careful media buyer who has clean conversion tracking than with a model that hasn’t seen enough results to learn from. The technology justifies its cost at scale or in accounts where behavior changes faster than the team can manually recalibrate.
The self-learning loop, explained without jargon
Every impression produces a result, whether it’s a click, a skipped ad, a full view, or a sale. Deep learning systems feed these results back into the model, so tomorrow’s prediction is built on today’s actual results, rather than assumptions from last quarter.
Repeated millions of times, this process causes the model to uncover segments that planning documents never mentioned. The working question shifts accordingly: instead of “who should we target,” it becomes “what did the model find that we missed?”
Why the Attribution Dashboard Is Misleading
Last-click attribution gives all the credit to the final touchpoint with the buyer. The model is convenient and familiar, but it systematically hides everything that happened before. If a deep learning model identifies an incremental audience that converts at a later touchpoint, the dashboard might never show it.
Academic research on consumer behavior and purchasing decisions has long pointed to the same problem: the path to a sale is rarely linear, and models that assume linearity misallocate the budget. The JSTOR archive contains decades of such studies, which are worth an afternoon’s reading for anyone who hasn’t gone through them yet.
The platforms themselves have moved away from simple models. In 2023, Google removed the first-click, linear, time-decay, and position-based attribution models from Google Ads and Google Analytics 4, retaining only data-driven, last-click, and external attribution. Data-driven attribution, in turn, remains a statistical estimate and does not replace an experiment. In a study published in *Marketing Science* in 2019, Brett Gordon, Florian Zettelmeyer, and their colleagues compared the results of fifteen randomized experiments conducted at Facebook—totaling 1.6 billion impressions—with estimates obtained using standard observational methods.
Observational methods generally overestimated the effect of ads, although in some cases they significantly underestimated it, and the authors concluded that even very comprehensive data do not guarantee reliable estimates. For an advertiser, this means that an attribution model, no matter how sophisticated, remains a hypothesis until it is experimentally confirmed.
The practical implication is simple to state but difficult to implement. A control group (holdout) is a portion of the audience that, by random selection, is not shown the campaign, so that the difference in conversion rates between the two groups measures the actual effect of the ads, including the effect on audiences that the model identified on its own. Without this comparison, any figure on the dashboard mixes sales driven by the campaign with those that would have occurred anyway.
Before judging any AI-based campaign, it is therefore advisable to first refine the measurement using control groups and incrementality tests. If the additional value generated by a model cannot be measured against a control group, the budget is simply paying for a more expensive assumption.
A quick test: Is the account ready?
Not all advertisers should jump straight to a deep learning platform, and the following check should be performed first.
- Does the account serve enough impressions each month for the model to have real results to learn from?
- Conversion tracking is clean, free of duplicates, and reviewed at least once a quarter.
- A control group test can be run without the finance department objecting to the “wasted” expense.
- Someone on the team is responsible for the model’s performance, in addition to campaign performance.
Three or more “yes” answers indicate a ready account. Fewer than three means the data foundation needs to be sorted out first, since a model fed with the wrong signals will only make the wrong decisions faster.
The first criterion is the hardest to quantify, because the volume threshold depends more on the number of conversions than on the number of impressions, and on how rarely the event the model must predict occurs.
An online store with hundreds of orders per week provides the model with a continuous stream of results, whereas a B2B company that closes ten contracts per quarter provides too few examples, no matter how high the traffic. In the latter case, optimization is usually based on an intermediate event, such as a request for a quote, and the link between this event and the actual sale must be verified separately.
The Decision That Matters
Observing teams that adopt this category of tools leads to an uncomfortable realization: most lose money in the first ninety days, and the cause usually lies in the fact that the model was not given room to prove them wrong, rather than in a failure of the technology. Teams feed it existing audience lists, cap spending on any unfamiliar segment, and check the dashboard weekly in search of the old numbers—only to then declare that AI is overhyped. The model did its job; people did not.
For a medium-sized account, the prudent approach would be to allocate a fixed budget, entrusted to the model with minimum limits and measured against a strict control group, over the course of an entire quarter, before any intervention. The system must be given room to surprise; otherwise, it isn’t worth allocating a budget to it at all.
What Research Says About Aversion to Algorithms
The behavior described above has a name in decision psychology. Berkeley Dietvorst, Joseph Simmons, and Cade Massey, from Wharton, published in 2015, in the *Journal of Experimental Psychology: General*, five studies in which participants—paid based on accuracy—had to choose between human predictions and those of a statistical model in tasks such as estimating the performance of MBA graduates or ranking U.S. states by the number of airline passengers. The model was more accurate in all five studies, and human predictions contained between 15 and 29% more errors in the MBA task and between 90 and 97% more in the one regarding air passengers. However, participants who had seen the model make mistakes became less willing to choose it, even when they had seen it outperform humans.
The authors called this phenomenon “algorithm aversion” and summarized its mechanism in a quote worth noting: people lose trust in an algorithmic predictor more quickly than in a human one after seeing them make the same mistake. Teams that check the dashboard weekly in search of past figures find themselves in exactly this situation, since every visible error in the model carries more weight than the equally real but less visible errors of the human planner. A ninety-day campaign provides ample opportunity for such visible errors, and the decision to abandon the model is usually made before the control group has accumulated sufficient data.
The same group of researchers showed, in a study published in *Management Science* in 2018, that aversion decreases considerably when people can modify—even slightly—the algorithm’s output. This finding qualifies the above recommendation regarding minimum limits. A test in which the team retains a few minor controls—such as a list of exclusions or a daily spending cap—is more likely to be carried through to completion than one in which everything is handed over to the model, provided that those controls are established in advance and do not change over the course of the quarter.
The framework has limitations that must be stated. The 2015 studies used two forecasting tasks, involving students and participants recruited online, and applying the conclusions to average buyers who are risking their own budgets remains an extrapolation. Moreover, none of the studies tracked repeated decisions over several months, such as those made in an advertising account. Jennifer Logg, Julia Minson, and Don Moore obtained, in six experiments published in 2019, the seemingly opposite result: non-experts were more likely to follow advice when they believed it came from an algorithm—a phenomenon known as “algorithm bias.” The notable exception was professionals who routinely make predictions; they relied less on algorithmic advice and thus saw their accuracy decline. The literature thus describes a reaction that varies depending on who is judging and what they have seen, and the marketer with eleven years of experience belongs precisely to the most skeptical group.
What the research community keeps repeating
The gap between what models can detect and what marketers actually do is a recurring theme in applied research. Articles published in journals such as MIT Sloan Management Review keep returning to the same point of friction: organizations adopt the tool, but not the decision-making structure it requires.
The problem is a management issue, disguised as a technological one. Purchasing a platform is the easy part; the part that determines whether the investment will pay off is reorganizing the authority to approve a budget reallocation and the speed at which this can be done.
The framework used to evaluate any AI marketing proposal
The framework is called the Three-Question Test and takes about ten minutes during a sales call.
- What does the model learn that a rule engine couldn’t learn? A vague answer likely indicates a rule engine with better branding.
- How is the incremental gain measured relative to a control group? The lack of a clear answer means the claim cannot be verified.
- Who is responsible for the model after implementation? If no one internally takes responsibility, the agency is selling results it cannot explain.
Vendors who can answer all three questions earn a longer discussion, while the rest receive a polite follow-up email. Teams seeking a more solid technical foundation before evaluating any proposal can find the basics—without the sales pitch—in resources such as Data Science Central.
Before the three questions: finding vendors
The test assumes that vendors are already on the shortlist, but compiling this list raises a prior issue of discovery and verification. The AI-powered marketing platforms category is crowded, and some of the products are—as the very wording of the first question suggests—rule engines with better branding. Those looking for a partner typically start with search results and recommendations—that is, two sources where visibility is either purchased or inherited through relationships. A curated directory offers a different starting point, because in the online marketing category of the Jasmine Directory, agencies, consultants, and software providers are grouped by field, and each entry has been reviewed by an editor before publication.
Such a listing confirms that the company exists, that it operates in the stated field, that the contact information displayed belongs to it, and that it can be found later. It does not verify claims about model performance and does not answer any of the three questions. The listing also does not constitute a recommendation. Its role is to pave the way for a conversation in which questions can be asked, and the answers remain to be verified by the control group.
The difference from rankings such as “the top ten AI marketing platforms” lies in the funding model. Many such rankings are supported by affiliate commissions paid by the featured providers, so their order reflects, at least in part, commercial terms. An editorially curated list does not rank providers by performance and does not claim to do so; it defines the scope within which the buyer will apply their own criteria, and this modest claim makes it easier to use as a starting point.
For buyers in the European Union, the third question also has a legal dimension. A supplier that trains models using the behavioral data of an advertiser’s customers processes personal data on the advertiser’s behalf, and Article 28 of the GDPR requires, in this case, a written contract specifying what data is processed, for what purpose, and with what safeguards. The question of who is responsible for the model after implementation is therefore complemented by the question of who is responsible for the data, and a provider that cannot present such an agreement should not proceed to the testing phase. The same agreement should specify whether the advertiser’s data may be used to train models offered to other clients of the provider.
The reasoning regarding attribution also applies, in fact, to the directories themselves. A buyer who discovers a company in a directory and returns two weeks later via a name search appears in reports as organic or direct traffic, so last-click attribution erases the listing’s contribution just as it erases the contribution of the incremental audience found by the model. The remedies are similar: UTM parameters on every listed link, dedicated phone numbers, and source fields tracked in the CRM throughout the campaign—methods described in detail in the guide on directory listing performance. Unlike a paid campaign, a directory listing is difficult to compare with a control group, so its measurement remains approximate, and this limitation should be openly acknowledged.
What’s Left for the Next Quarter
The instinct that made someone good at this job isn’t outdated, but it’s better suited to judgment than to pattern detection. The model can find the audience, and the human decides whether it’s worth serving.
Thus, the distinction between detection and judgment runs through the entire problem. The model uncovers buyers the planner didn’t know about; buyers, in turn, discover companies they weren’t looking for; and in both directions, the discovery is only as valuable as the data it’s based on: a clean conversion signal in one case, a verified listing in the other. One can conclude that intuition has been moved to a more appropriate place—where decisions are made about what is worth believing from everything the systems have found.
For the next quarter, only one campaign may be selected. The three-question test is applied to any AI platform under consideration, a control group is set up, and the decision is then made based on the figures obtained, setting aside the sales pitch.
What would change in planning if the model were fully trusted to find buyers whose existence was previously unknown?
