Why SEO foundations can still miss the AI shopping candidate set

Search visibility and AI-shopping visibility share foundations. Both depend on accessible pages, understandable language and credible information. But an AI shopping answer does more than retrieve a page. It has to associate a product with the buyer's situation, obtain usable commerce facts, weigh evidence from other sources and decide whether the product belongs in the candidate set.

That extra decision path explains why a technically sound storefront can still have a product-level discovery problem.

A useful distinction: retrieval is not selection

Traditional SEO is usually evaluated through indexed pages, search impressions, positions and clicks. These are not simple metrics, but they are tied to a ranked search result.

Generative systems produce a different output. They retrieve information, synthesize multiple sources and compose an answer in which products may be mentioned, compared, recommended or omitted. The KDD 2024 paper GEO: Generative Engine Optimization formalized this distinction and proposed visibility measures designed for generative answers rather than conventional result positions.

For product discovery, this creates at least four separate questions:

1. Could the system retrieve usable information about the product?
2. Did it treat the product as a plausible candidate for this buyer need?
3. If included, did it compare the product using accurate and relevant facts?
4. Where did it place the product in the final answer?

SEO contributes to the first question. It can also support the other three. But a page being available to retrieve does not establish that the product will enter the right candidate set.

THREE-LAYER EXPLANATION

MC Lab's three-layer explanation

MC Lab uses a three-layer diagnostic framework to organize that question. This is an operational synthesis, not a pre-existing academic theory with the same name. Each layer is supported by adjacent research, but the combined model is MC Lab's interpretation of the product-selection problem.

LAYER 1 · PRODUCT-SOURCE CLARITY

Layer 1: Product-source clarity

The first layer is the product's own public explanation: what it is, who it fits, which category it belongs to, how it differs and whether the landing product page supports the purchase situation.

A page can contain accurate nouns and still leave the important association implicit. “Sterling-silver necklace,” “porous stone” and “essential oils” describe the object. The buyer question, however, asks for a bridge between those facts and a situation: discreet workplace use, polished everyday appearance and local availability.

An AI system has to make that bridge before it can treat the product as a serious candidate. This is broader than adding keywords. It is category and buyer-fit clarity at the SKU level.

LAYER 2 · COMMERCE-DATA ELIGIBILITY

Layer 2: Commerce-data eligibility

The second layer is the observable commerce data needed to identify and compare the offer: product type, attributes, variants, price, availability, market eligibility and structured product information.

Research on ecommerce catalog systems supports the importance of this layer. Amazon's 2025 CatalogRAG study used product type, similar catalog entries and attribute-filled examples to improve missing-attribute prediction across US, German and French catalog data. The study concerns catalog completion rather than public AI-shopping recommendations, so it should not be treated as direct proof of shortlist inclusion. It does show that product understanding improves when category and attribute relationships are explicit and internally consistent.

This layer is also where diagnosis must remain disciplined. Structured data, feeds and crawl access are eligibility checks—not automatic explanations and not automatic paid fixes. In the jewelry observation, public price and availability were visible, but merchant-side feed coverage and platform eligibility were unknown. They remain alternative explanations, not established causes.

LAYER 3 · AUTHORITY ENVIRONMENT

Layer 3: Authority environment

The third layer is the information environment around the product: reviews, media, community references, retailer listings and other sources that may support, omit or contradict the merchant's own claims.

This matters because an AI shopping agent does not necessarily evaluate a product page in isolation. Wharton Generative AI Labs' 2026 technical report, Agentic Shopping is Complicated and Contingent, ran roughly 26,000 controlled tests and found that prior sources, competing sources, source order, injected user context and tool-call design could all shift product choices. It is a technical report rather than a peer-reviewed commerce-ranking study, but it provides unusually direct evidence that the surrounding information environment can change an agent's decision.

External evidence can also conflict. Chen, Zhang and Choi's EMNLP 2022 paper, Rich Knowledge Sources Bring Complex Knowledge Conflicts, showed that retrieved passages can disagree with one another and with a model's internal knowledge, while retrieval performance influences which information the model relies on.

For a merchant, this means a clear product page may still compete with sparse, older or contradictory representations elsewhere. It does not mean that merchants should manufacture mentions or pursue generic “authority” at any cost. It means legitimate reviews, accurate retailer data and consistent public facts are part of the environment in which the product is interpreted.

PRACTICAL DISTINCTION

SEO is a foundation, not the opposing camp

The practical conclusion is not that GEO replaces SEO.

SEO helps make product information accessible, legible and discoverable. Product-level AI-shopping diagnosis asks what happens next:

• Does the system connect the SKU to the buyer's real situation?
• Are the category, attributes, price and availability usable in comparison?
• Do outside sources support or destabilize the product's representation?
• Does the product survive the comparison and appear in the answer?

Shopify's own industry guidance reaches a similar, though not academic, conclusion. Its 2026 GEO Playbook describes SEO, brand authority and product data as complementary foundations for AI visibility. MC Lab's model is more diagnostic and product-specific, but the overlap is useful: conventional search foundations are necessary inputs, not a complete account of AI-shopping selection.

EVIDENCE LIMITS

What this observation does—and does not—justify

The anonymous jewelry test justifies three claims:

1. A product can have clear public product and category content and still be absent from one natural AI-shopping candidate set.
2. Candidate-set inclusion should be analyzed separately from traditional search visibility and final ranking.
3. Product-source clarity, commerce data and the authority environment are plausible cause layers that should be tested separately.

It does not justify saying that the merchant's SEO was successful, that its GEO was poor, or that one identified layer caused the omission. A single captured answer is evidence of one observed outcome under recorded conditions—not a stable market ranking.

This caution is especially important because generated answers can look more evidentially complete than they are. Liu, Zhang and Liang's EMNLP 2023 study, Evaluating Verifiability in Generative Search Engines, found substantial gaps in both citation coverage and citation accuracy across the systems they evaluated. The exact products have since changed, but the research supports a durable testing principle: preserve the full answer, sources, timestamp and session conditions before interpreting the result.

NEXT DECISION

The smallest useful next step

When a relevant product is absent, the answer is not a generic GEO checklist. The next step is to identify the earliest unsupported transition in the decision:

• If the product cannot be associated with the need, clarify the product type, category bridge and buyer-fit evidence.
• If comparison facts are missing or unstable, improve the smallest relevant set of product and commerce attributes.
• If outside sources contradict the offer, reconcile the controllable facts and document what cannot be controlled.
• If the product is included and fairly rejected, there may be no visibility problem to fix.

Then retest under controlled conditions.

The goal is not to promise a recommendation. It is to move from “the product did not appear” to a narrower, evidence-based explanation of where it left the decision—and what, if anything, the merchant can responsibly change.

SOURCES

References

1. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative Engine Optimization. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining.
2. Zhang, B., Khan, S., & Walter, S. (2025). CatalogRAG: Retrieval-guided LLM prediction for multilingual e-commerce product attributes. KDD 2025 Workshop on LLM4ECommerce.
3. Kumar, A., Meincke, L., Shapiro, D., Mollick, L., Puntoni, S., & Mollick, E. R. (2026). Agentic Shopping is Complicated and Contingent. Wharton Generative AI Labs, Prompting Science Report 6.
4. Chen, H.-T., Zhang, M., & Choi, E. (2022). Rich Knowledge Sources Bring Complex Knowledge Conflicts: Recalibrating Models to Reflect Conflicting Evidence. Proceedings of EMNLP 2022, 2292–2307.
5. Liu, N. F., Zhang, T., & Liang, P. (2023). Evaluating Verifiability in Generative Search Engines. Findings of EMNLP 2023, 7001–7025.
6. Shopify. (2026). The GEO Playbook: How (& Why) to Optimize for AI Discovery. Industry guidance; not peer reviewed.