Shopify’s Shop MCP gives AI agents a new way to search product data, build carts and complete Shop Pay checkout through a single connection. That matters because product discovery in this environment is not simply a matter of where a merchant ranks in a catalog or in Google. An agent also interprets the available product information and decides what to recommend.
A hands-on experiment reported by Search Engine Journal illustrates how consequential that decision layer can be. Three Claude models received the same fixed set of 20 products retrieved through Shop MCP, yet they did not select the same top product in any of six searches. For merchants, the result is an early indication that catalog presentation, review signals and prompt fit may shape unpaid AI shopping discovery alongside established SEO and ecommerce optimization work.
What Shop MCP adds to agentic commerce
Shopify announced Shop MCP on October 13, 2026. According to Shopify’s Shop MCP announcement, the connection enables an agent to search the Shopify Catalog, create carts and use Shop Pay checkout. Shopify said catalog search does not require an API key and that the catalog covers more than one billion products.
The significance is not merely that an AI assistant can point a shopper to a product page. An agent can use structured product information during a shopping interaction, narrow the choices, make a recommendation and potentially take the shopper through checkout. This places a model-mediated selection process between the buyer’s request and the merchant’s listing.
That process is distinct from a conventional search result page. Traditional organic visibility still depends on crawling, indexing, ranking and search-result presentation. In an agent-enabled shopping flow, visibility can additionally depend on whether the agent receives the item in its candidate set, how the catalog describes it, and how a particular model evaluates those details against the shopper’s request.
Same products, different AI decisions
The Search Engine Journal test used three Claude models: Haiku 4.5, Sonnet 5 and Opus 5. Researchers gave each model the same 20 products obtained via Shop MCP and ran six searches. Each model made 50 selections per search, producing 1,980 total picks.
The models never agreed on the leading product across those searches. Their reported top-product agreement rates were 8% for Haiku, 0% for Sonnet and 20% for Opus. Those figures describe this specific experiment, not all AI shopping behaviour or every product category. Still, they show that access to the same catalog does not produce a single, stable answer.
The experiment also found that a product’s review count often appeared to influence the choice. In several cases, the product with the most reviews was selected most frequently. That was not a universal rule, however. Choices varied by model and search, suggesting that no single catalog signal can be treated as a guaranteed recommendation trigger.
A further test changed the ordering of an item in the catalog list, moving it from first to last. Model selections changed substantially. This reinforces an important limitation for merchants: agent outputs can be sensitive to factors beyond product quality or traditional rankings, including list position, prompt wording and model-specific interpretation.
| Layer | Core question | Potential merchant influence |
|---|---|---|
| Catalog discovery | Is the product available in the agent’s candidate set? | Accurate, complete catalog data and product availability |
| Model decision | Which candidate does the model consider the best fit? | Clear descriptions, relevant features, accurate claims and review signals |
| Checkout | Can the shopper progress from recommendation to purchase? | Operational readiness of the merchant’s commerce setup |
Why this changes the meaning of unpaid product discovery
The findings do not mean SEO is obsolete, nor do they establish that Shop MCP traffic will replace search-engine traffic. They do point to an additional form of unpaid discovery: a product can be surfaced because an agent considers its supplied data persuasive for a particular request.
This is different from paid placement. The reported experiment concerned non-advertising agent interactions and the selections made from catalog-provided products. A merchant may therefore need to think beyond visibility alone. Being present in an agent-accessible catalog is valuable, but it may not be sufficient if competitors communicate stronger evidence of fit in their titles, summaries, features or review profiles.
It also means that a product’s position in a traditional organic ranking cannot be assumed to predict an AI agent’s choice. In the experiment, models could prioritize review counts and other product attributes, while their preferences diverged despite identical inputs. The model is not a neutral display layer. It is an active recommender.
Practical steps for Shopify merchants
Merchants should avoid treating a single AI response as a definitive measurement of product quality or future demand. The observed variability makes repeated testing more useful than one-off checks. A practical starting point is to review the product information an agent can access and compare it with the language shoppers are likely to use.
Focus on the fundamentals that the experiment indicates agents may use:
- Verify how products appear through Shop MCP, including titles, summaries, feature descriptions and review information.
- Match catalog language to genuine buyer needs and category-specific prompts without adding unsupported claims.
- Keep product attributes, availability and claims accurate so an agent has reliable evidence to evaluate.
- Test representative customer questions across relevant product categories and note where recommendations differ.
- Compare the selected products with the attributes agents cite or appear to prioritize, then use those observations to improve clarity rather than manipulate facts.
The aim is not to find a universal wording formula. The test suggests there may not be one, especially when different models interpret the same products differently. Instead, merchants can reduce ambiguity, make genuine differentiators easy to understand and identify gaps between their catalog information and likely shopping questions.
If your ecommerce team needs a clearer view of how product information may support discovery in AI-driven shopping journeys, create a GEO Search Leads account. It is a practical first step toward assessing your website’s GEO readiness, identifying where content and entity signals may be unclear, and prioritising optimization opportunities without assuming that one AI response represents every shopper or platform.
Frequently Asked Questions
What is Shop MCP?
Shop MCP is Shopify’s connection for agents to search the Shopify Catalog, build carts and complete Shop Pay checkout. Shopify announced it on October 13, 2026, and said catalog search does not require an API key.
Did the tested Claude models receive different products?
No. The reported experiment gave Haiku 4.5, Sonnet 5 and Opus 5 the same fixed list of 20 products retrieved through Shop MCP. Their different top choices therefore reflect differences in model decisions rather than different initial product sets.
Do more reviews guarantee an AI shopping recommendation?
No. The experiment found that products with more reviews often aligned with top selections, but the pattern was not universal across models or searches. Review count appeared to be one influential signal, not a guarantee.
Does Shop MCP replace SEO for ecommerce stores?
No. The research does not show that it replaces traditional search visibility. It shows that agent-based shopping can add another discovery and recommendation layer, where catalog data and model interpretation may affect product selection.
Conclusion
Shop MCP makes agent-based product discovery a more concrete commerce channel within Shopify’s ecosystem. The reported Claude experiment shows why merchants should separate catalog inclusion from recommendation: the same products can lead to markedly different choices depending on the model and the request. Clear, accurate and buyer-relevant catalog information is therefore becoming important not only for shoppers and search engines, but also for the AI agents that may increasingly guide shopping decisions.
Top comments (0)