AEO study. From unknown to agent-recommended.
AEO study · Internal groundwork
Summary
Before offering Answer Engine Optimization to clients, we wanted to understand what actually separates a merchant that can be indexed from one that can be found, described correctly and selected by a shopping agent. We used a controlled retail environment and observed the same shopping questions across multiple independent agent harnesses.
Why we ran the experiment
Search visibility and agent visibility are not the same. A page may be available to retrieval systems without entering an answer, and a merchant may be retrieved accurately without being recommended. We wanted to learn where those behaviors separated and how far legitimate first-party improvements could move a merchant toward the small set of options an agent presents to a shopper.
Methodology
We began with a negative baseline, then repeated a fixed set of branded and unbranded shopping prompts in fresh sessions. Branded prompts tested whether the merchant could be found after its name was supplied. Unbranded prompts tested whether it could enter an answer on the strength of the shopper's need alone. Repeating the same prompts at checkpoints helped us distinguish sustained movement from a single favorable response.
We changed four areas deliberately: structured product information, answer-ready content, technical discoverability and the evidence available beyond the merchant's own claims. For every observation, we kept retrieval, factual accuracy and recommendation as separate outcomes.
Method
- prompt
- branded / unbranded
- session
- fresh
- check
- repeated
- measure
- retrieval / accuracy / recommendation
What worked
A coherent catalog, explicit specifications, consistent policies and direct answers to real shopping questions gave the harnesses facts they could retrieve and repeat. Clear entity language reduced confusion, while crawlability and indexing signals helped accurate first-party information enter retrieval. Precise, long-tail shopping needs produced the earliest unbranded visibility; broad “best brand” questions continued to favor established merchants with stronger independent reputations.
- Unknown
- Retrievable
- Agent-recommended
Learning outcomes
The progression was real, but it was not uniform. Indexing, branded retrieval, factual understanding and unbranded recommendation are different states. A merchant may succeed at one and fail at the next. Availability also matters: an otherwise relevant product becomes a weaker answer when an agent cannot confidently complete the purchase.
We learned not to treat a successful screenshot as proof of durable visibility. Results vary with retrieval state, model behavior, timing and session context, so the same question must be repeated before movement can be considered meaningful. We also maintained a clear ethical boundary: no reviews, endorsements, discussions or product experiences were fabricated. Technical readiness can earn eligibility, but trustworthy external evidence is what earns consideration.
How this helps our clients
The experiment became a more disciplined assessment method. We now establish a controlled baseline, separate retrieval from recommendation, improve the evidence agents actually use and repeat the same questions over time. This helps us determine whether a merchant needs entity cleanup, catalog work, technical discovery fixes, stronger external evidence or simply more observation before investing further.