AI recommendations changed pension choices. That raises the stakes for AEO

Much of the AEO industry rests on an assumption that has been easier to state than to prove: AI recommendations matter because people act on them. A new controlled experiment provides unusually direct evidence that recommendation content can change subsequent economic choices. In a pension-allocation task, 81% of participants revised their portfolios after receiving AI-generated advice. Among those who revised, 95% moved toward the recommendation they had been assigned. [1]

The result does not show that consumers will follow every AI suggestion, nor does it prove that a recommendation in ChatGPT, Claude, Gemini or Perplexity translates automatically into a purchase. It does establish something narrower and important. When researchers changed the content of AI advice in a controlled setting, participants made materially different financial decisions afterwards. That makes recommendation a more consequential AEO metric than a simple brand mention.

<h2>This was a randomised experiment, not a survey</h2>

The working paper, <em>Do People Follow AI Advice? Evidence from a Pension Portfolio Choice Experiment</em>, enrolled 400 employed adults in South Korea who participate in workplace defined-contribution pension plans. Participants first allocated a hypothetical pension balance across 11 products. They were then randomly assigned one of two fixed AI-generated portfolio recommendations, one relatively aggressive and one relatively conservative, and were allowed to revise their choices. [1]

The design matters because it separates causal influence from correlation. The researchers did not merely compare people who use AI with people who do not. They changed the recommendation and measured how much of that experimentally induced difference appeared in the participants’ final portfolios. The task was hypothetical, but it included performance-based incentives and was designed around a consequential decision the sample already faces in practice. [1]

<h2>The advice changed decisions, but did not replace judgement</h2>

About 37% of the difference between the aggressive and conservative recommendations passed through to participants’ final portfolios. The changes affected expected return, volatility, risk-grade allocations and the number of products held. The authors found no detectable improvement in computed Sharpe ratios, a reminder that movement toward an AI recommendation is not the same thing as a demonstrably better outcome. [1]

The most useful interpretation is neither blind obedience nor irrelevance. Participants retained substantial weight on their original decisions. Among those who changed their portfolios, they implemented about half of the suggested adjustment on average. AI advice supplied a direction that people followed partially and selectively, which may be closer to how decision influence works outside a laboratory than an all-or-nothing account of trust. [1]

<h2>Recommendation is different from visibility</h2>

First-generation AEO measurement has understandably focused on inclusion: whether a brand appears, how often it is mentioned and which sources support the answer. Those are necessary signals, but they do not describe what happens once a person begins choosing among options. A brand can appear frequently without being recommended. Another can appear less often yet become the preferred option whenever the prompt reaches a decisive comparison.

That distinction is why the practical measurement hierarchy should extend beyond visibility. Visibility asks whether the brand entered the response. Consideration asks whether it became a viable option. Recommendation asks whether the system identified it as the preferred fit. Choice asks whether that recommendation changed human behaviour. The pension experiment does not complete that chain for consumer brands, but it offers causal evidence that the final link can exist. [1]

<h2>A rationale did not amplify the effect</h2>

The researchers also varied whether participants received a short explanation alongside the numerical recommendation. They found no detectable difference in average pass-through. That result should not be overgeneralised from one financial setting, but it challenges the simple assumption that more explanation necessarily creates more influence. In this experiment, the recommendation’s substantive content mattered more than whether it arrived with a short rationale. [1]

<h2>The limits are part of the finding</h2>

This is a preprint, not a universal law of AI behaviour. The sample comprised employed South Korean adults aged 35 to 55, the decision involved pension allocation rather than a retail purchase, and the researchers identify the effect of assignment to these two particular recommendations rather than a general advantage of one GPT model over another. The study is therefore best read as strong causal evidence within a defined setting, not as proof that 95% of consumers will do whatever an assistant recommends. [1]

<h2>AEO Updates Takeaway</h2>

The commercial question for AEO is becoming more precise. It is not only whether an AI system can retrieve a brand, but whether the system represents that brand as the right choice in a decision context and whether that representation affects the user. The evidence is still emerging, yet this experiment makes one conclusion harder to dismiss: AI recommendations can move choices without fully replacing human judgement. For brands, the contest is not ultimately for another mention. It is for defensible influence at the point of choice.

<h3>References</h3>

[1] Choi, H., Kim, J., Kovach, M., Lee, K.-M., Shin, E., & Tzavellas, H. (2026). <a href="https://arxiv.org/html/2608.11371" target="_blank" rel="noreferrer"><em>Do People Follow AI Advice? Evidence from a Pension Portfolio Choice Experiment</em></a>. arXiv:2608.11371v1, 11 August 2026.