Shopping Tool System Prompt Analysis
This post breaks down the components of the ChatGPT shopping assistant released November 26th, 2025. Below you can find extrapolations from the tool’s system prompt. This info may give SEOs and others the tools to optimize their websites to be found in or increase their visibility in the ChatGPT shopping research tool. If you’re unfamiliar, you can find the OpenAI announcement for the assistant here. In the network tab, the tool defines itself as “A shopping tool that helps find the right product or gift faster by asking smart questions, scouring in internet for reviews and deals, and researching across trusted sources to deliver a personalized buyer’s guide.”.
- System Prompt Overview
- Supported and Non-Supported Categories
- Product Attributes
- Bounds of the Tool
- Shopping Tool Actions
- Potential SEO Strategies and Tactics for ChatGPT Shopping Tool
The tool appeared to be a rush job on OpenAI’s part. The week of Black Friday in the US means there’s a great opportunity to get large amounts of real-world data very quickly, in addition to being a prime time to drive high-volume adoption and a bunch of easy press. In OpenAI’s announcement they showed internal benchmarks which were, let’s say, “something less than performant” when it comes to accuracy.
The shopping assistant appears to be a tool called by an MCP. For me, this was the first time seeing OpenAI use an internal MCP to call a tool. At least, my first time seeing it appear explicitly in the network tab.
A little more snooping in the network tab led to me our find of the day: The Shopping Assistant system prompt.
Exploring the Shopping Assistant System Prompt
This prompt was about 17,700 characters long and contained a lot. Within the prompt you can see efforts made to reign in common idiosyncrasies of ChatGPT such as being aggressively thorough, or reiterating/repeating the user’s statements or feedback. Something I found particularly interesting is that when a user expresses a preference such as model, material, size, color, budget, region, brand, or retailer, that the shopping tool must “treat those as locked” and “never re-ask, confirm, or ask additional questions”. It also notes “You’ve been provided with the users’ entire conversation. If there are preferences already expressed in that conversation, DO NOT REASK OR CONFIRM”.
As a marketer and SEO, this appears to have some implications worth considering. First and foremost is that once users express a preference, a brand may lose the opportunity for product discovery if they don’t align with the expressed preference. This is a filtering of your funnel that you do not directly control and highlights a critical need for both audience data and educational/influencer campaigns to prompt users to mention key terms that favor brand inclusion.
Below you’ll find breakdowns of the rest of the system prompt. Have questions? DM me on LinkedIn.
ChatGPT System Prompt Components
The following are the components of the system prompt as they were, extracted on 11/26/2025.
Product Categories
This table includes categories explicitly mentioned as supported (Good SERP), unsupported/poor data (Bad SERP), prohibited, or used in examples.
| Category Type | Specific Categories Mentioned |
|---|---|
| Good SERP Coverage | Consumer electronics, Apparel & footwear, Home goods & décor, Home appliances & cleaning, Fitness & sports gear, Toys & baby products, Outdoor & camping gear. |
| Bad SERP Coverage | Vehicles (cars, motorcycles, ATVs, boats, RVs), Real estate (homes, rentals, land), Insurance (auto, home, life, health, pet), Financial products (credit cards, loans, mortgages), Local services (plumbers, cleaners, etc.), Hospitality & travel (hotels, flights), Professional/B2B equipment, Prescription medications, Event tickets, Digital content & licenses. |
| Prohibited / Restricted | Weapons (firearms, ammo, specific knives), Drugs (recreational, paraphernalia), Illicit financial/fraud items (fake IDs, skimmers), Cybercrime tools (malware, spyware), Sexual harm content, Self-harm items, Extremism/violence, Illegal medical/biological items, Contraband, Alcohol/tobacco/nicotine (where restricted), Gambling, Pirated media, Adult products/Sex toys. |
| Example Categories | Smart light switches/bulbs, Tesla Model Y vs Hyundai Ioniq 6 (Cars), College gifts, Designer vegan tote, Running shoes, Capsule outfit set, Downhill skis, 55″ 4K TV, Espresso machine (Breville Bambino). |
The “Example Categories” are preceded by the following text in the system prompt:
How to set \"show_rateable_products\"\n- True for product discovery queries where seeing examples helps, AND where we expect to have GOOD SERP product API coverage.\n- False for queries where seeing examples DOESN'T help, OR where we expect to have BAD SERP product API coverage.
Product Attributes and Descriptors
This table outlines the specific attributes the system tracks, asks about, or uses to filter products. If you work in e-commerce, start checking your PDPs for these attributes.
| Attribute Type | Descriptors & Specifics |
|---|---|
| Financial | Budget/Price (Must use “Up to $X” and “$X+”, never ranges), Cheapest price, Best deals. |
| Physical/Technical | Size/Fit, Material, Color, Hardware (silver vs gold), Smart platform (Google Home, Alexa, etc.), Specs (HDMI 2.1, 4K), Laptop accommodation (16″). |
| User-Centric (Innate) | Height, Weight range, Foot length, Gender (Men’s, Women’s, Unisex, Youth), Ability/Skill level (Beginner, Expert). |
| Usage & Context | Living space, Usage frequency, Primary settings (Office, Travel, Gym), Terrain mix, Milk preferences (for coffee). |
| Style & Vibe | Style, Brand, Feel, Vibes (Cute, Minimal, Edgy), Brand preferences. |
| Forbidden/Expert Attributes | The system explicitly avoids asking for: Ski length, Bike geometry, Lens flange distance. |
User Attributes & System Sources
This table details the sources of information the system uses to construct its understanding of the user outside of the quiz. As I mentioned before, the tool has access to the user’s ChatGPT profile, custom instructions, and entire conversation history. This will further challenge marketers who don’t have fine details on their audience or audience’s preferences.
| Source Type | Specific Information / Data Points |
|---|---|
| Explicit User Input | The user’s current prompt/request, User-specified requirements, Locked preferences (Model, material, size, color, budget, region, brand, retailer) provided in conversation. |
| Implicit Context | Inferred preferences such as likes, dislikes, style, region, household, habits, or past purchase patterns, “User context” blocks provided in examples (e.g., “Product designer, tech-savvy, lives in Oakland”). |
| Conversation History | The user’s entire conversation history is provided to the system; previously expressed preferences must not be re-asked. |
| Innate Qualities | Gender (if not explicit in data, it must be requested), Height, weight, and foot length. |
Bounds of the Tool
Based on the system prompt, the tool operates within the following specific boundaries and operational logic:
Primary Function: Clarification, Not Execution
- The ChatGPT Shopping tool is designed to be a “Shopping Assistant” that generates structured multiple-choice quizzes.
- It does not explain its choices or provide commentary; it outputs a strict JSON object containing questions or requirements.
Scope of Inquiry (The “Quiz” Constraints)
- Question Volume: It is limited to asking 0–5 clarifying questions. It asks 0 questions if the request is highly specific or a direct comparison.
- Question Content: It avoids expert-level criteria (e.g., “Ski length”) in favor of innate user attributes (e.g., “Height”).
- Redundancy: It is strictly forbidden from re-asking about preferences already stated in the current prompt or previous conversation history.
NOTE: Once the initial prompt is given to the assistant, the shopping process is automated. If the user does not answer the quiz questions, each question times out and the tool proceeds to the next step without the clarifying information.
Product Display Logic (“Show Rateable Products”)
The tool has a binary decision-making process for showing products:
- Shows Products (True): Only for “Product discovery” queries where visual examples help AND the category has good API coverage (e.g., Apparel, Electronics, Home Goods).
- Hides Products (False):
- When the user wants the “cheapest” item or specific specs (sorting/lookup tasks).
- Direct comparisons (X vs Y).
- Categories with bad data coverage (Vehicles, Real Estate, Insurance, Tickets, etc.).
When the tool shows products for the user to rate, the options are acting as filters using inferred attributes (or possibly explicit, taking from the image alt text or corresponding URL’s page text/metadata) and visual cues from the product images it’s displaying. Liking one product image card may be adding weight to it’s associated attributes. Conversely, disliking a product image card may add negative weights to some or all of its attributes. All this is to say that we should be thinking very carefully about what product attributes are conveyed via product images and what attributes matter the most to our target audience.
Safety and Prohibitions
- The tool has a hard boundary on safety. If a request involves weapons, drugs, cybercrime tools, self-harm, or other illicit items, it must return a specific “Prohibited request” JSON structure with no questions or results.
Interpretation of User Intent
- It prioritizes implicit preferences (e.g., inferred style or past habits) to guide the options it presents, but it never mentions these implicit traits directly in the text.
- It assumes the user knows innate details (height, gender) but protects the user from technical jargon.
Shopping Assistant: Network Tab Findings
The network tab had a few other goodies for us. One of the first things you may note is the list of actions the shopping tool contains:
- get_all_products
- get_state
- set_nux_viewed
- start
- steer
- subscribe
Start and get_state are pretty self explanatory. Set_nux_viewed appears to reference the initial tool dialog that explains the tool’s purpose and how to use it. Presumably this exists to avoid showing users the intro text each time. “Steer” refers to the portions of the quiz asking for additional user input on attributes like brand, style, price, etc. The quiz shows several products it selects and asks the user if they like this product. A yes or no response steers the tool in one direction vs another.
Shopping Tool Actions (Technical + Example Phrases)
Search
Executes a structured query over product catalogs.
User-facing examples:
| Action | Technical Definition | User-Facing Example Phrases |
|---|---|---|
| Search | Executes a structured query across product catalogs; returns ranked product objects. Appears to use Google Search (see image below) | “Find…”, “Show me…”, “Look for…” |
| Filter | Applies hard constraints (price, size, color, materials, shipping origin). | “Only show under $75.”, “Filter to US shipping.”, “Exclude slim fit.” |
| Sort | Reorders result sets by a metric (price, discount, relevance). | “Sort by lowest price.”, “Show the biggest discounts first.” |
| Personalize | Applies soft ranking boosts via user preference vectors. | “I prefer natural fibers.”, “I like relaxed fits.”, “I’m usually a size large.” |
| Inspect Product | Retrieves full structured attribute data for a product/SKU. | “Open details.”, “Tell me more about the second one.” |
| Compare | Produces a normalized comparison matrix aligning product attributes. | “Compare these.”, “Which one is heavier?” |
| Explain Recommendation | Reveals why an item was included or ranked (filters matched, boosts applied). | “Why was this recommended?”, “Why did you include this?” |
| Adjust Preferences | Updates long-term preference vectors or constraints. | “I don’t like cropped fits.”, “Prefer black in the future.” |
| Refine | Re-runs search + filter + sort after constraints change. | “Actually make it under $50.”, “Include long sleeves too.” |
| Explore Similar | Runs embedding/attribute-based similarity from a seed product. | “Show similar items.”, “Find alternatives to this one.” |
“srsltid=” param, so results are still pulled from Google Search
Potential SEO Strategies for the ChatGPT Shopping Tool
The key to optimizing for this LLM Shopping Tool is to recognize that ChatGPT acts as a structured filter or a guided funnel. Your content must be structured to preemptively and precisely answer the LLM’s diagnostic quiz questions, ensuring your products are the high-confidence recommendations.
Optimization based on Product Attributes and Descriptors
This is the most critical area for e-commerce SEO. The shopping tool’s quiz questions are entirely built on these attributes, which must be clearly crawlable. The good news is that these optimizations are the exact optimizations you would typically make for any e-commerce store.
| LLM Attribute Type | SEO Action for E-commerce | Implication for Discovery |
|---|---|---|
| Physical/Technical Specs | Structured Data & Faceting: Use Schema.org properties (Product, Offer) to explicitly define all specs (e.g., dimensions, color, material). Technical specs (like “HDMI 2.1” or “4K”) must be prominent in product titles, H-tags, and rich snippets. | The LLM converts user needs (“TV for PS5”) into technical specs (“HDMI 2.1”). Your page must clearly state this spec in a crawlable, structured way. |
| User-Centric (Innate) | Dedicated Content & Filtering: Create robust internal linking and category pages for innate attributes like Gender and Skill Level (e.g., /running-shoes/womens/, /downhill-skis/beginner/). | If the LLM asks about Skill Level and the user answers “beginner,” the LLM needs a high-authority page or clearly tagged product to recommend. |
| Usage & Context | Descriptive Product Copy: Use explicit product descriptions that state the intended use and environment (e.g., “Perfect for Mixed Terrain running,” “Ideal for Office and Travel“). | The LLM converts “I need something for work” into the **Usage Context** attribute. Match your copy to these functional phrases. |
| Style & Vibe | Aesthetic Keyword Optimization: Optimize product descriptions and metadata with specific aesthetic keywords (e.g., “minimalist,” “edgy,” “cute,” “mid-century modern”). | The LLM uses inferred style preferences to filter products. Describing your products using these “vibe” words helps the LLM make the connection. |
| Financial/Budget | Clear Pricing Tiers: Structure your product data and category pages around simple price boundaries (e.g., “$100-$200,” “Premium $500+”). This matches the LLM’s required “Up to $X” and “$X+” format. | Ensure price filters on your site use simple, clear budget buckets that the LLM can easily map to. |
Optimization based on User Attributes and System Sources
This recognizes the LLM’s reliance on context and conversation history.
| LLM Source Type | SEO Action for E-commerce | Implication for Conversion |
|---|---|---|
| Implicit Context & Style | Build Category and Landing Page Authority: Ensure your top-of-funnel content and category pages for a certain aesthetic (e.g., “Scandi Minimalist Home Decor”) are exceptionally high-quality and authoritative. | If the user is flagged as “Minimalist,” the LLM will prioritize showing products from the sites it knows are authoritative for “Minimalist” styles. |
| Explicit User Input | FAQ/Q&A Schema: Implement Q&A or FAQ schema on product pages that directly address common constraints (size, fit, brand preference). | This helps the LLM confirm that the product meets the user’s specific, stated requirements, reducing the need for clarifying questions. |
| Conversation History | Consistency: Ensure product attributes are consistent across all parts of your site (product title, description, structured data, and reviews). Any inconsistency in naming conventions could lead the LLM to ask a redundant question. | A clean data hierarchy prevents the LLM from asking about something it already “knows” from a previous step in the conversation. |
Optimization based on Product Categories and Bounds
This dictates where an SEO’s effort will yield the highest return based on the LLM’s operational status.
| LLM Category Status | SEO Action for E-commerce | Rationale |
|---|---|---|
| Good SERP Coverage | Invest Heavily in High-Fidelity Data: For Apparel, Electronics, Home Goods, and Fitness, the LLM will show rateable products. Focus on perfect structured data and rich visuals. | The LLM is designed to guide users to visual examples in these categories, meaning your product is likely to be displayed to the user for immediate rating/selection. |
| Bad SERP Coverage (e.g., Vehicles) | Shift Focus to Authority and Comparison Content: Since the LLM *avoids* showing products here, focus on generating high-authority comparison content (e.g., “Tesla Model Y vs. Hyundai Ioniq 6: A Full Review”). | The user is more likely to be completing research *before* making an external purchase, so the LLM acts as an informational source, not a direct product funnel. |
| Prohibited / Restricted | Verify and Clarify Intent: For any products near the “edge case” line (e.g., knives, legal consumables), use product copy to explicitly clarify the product’s legal and safe intent (e.g., “Intended for culinary preparation only”). | This helps the LLM distinguish your product from prohibited items. |
Summary of Optimization Opportunities
The core optimization opportunity is to transform your e-commerce content from being broadly descriptive to being diagnostically precise.
- Structure Data to Match the Quiz: Ensure every product page acts like a well-structured answer key to the LLM’s potential questions (e.g., Material:
Vegan Leather, Fit:True to Size, Usage:Day-to-day Commute).Know what attributes are relevant to your audience (or which ones definitely are not relevant) and add them to your product descriptions. For clothing, include accessible Size Guides in Metric and Imperial measurements. - Optimize for Implicit Signals: Use aesthetic and lifestyle keywords to help the LLM link its inferred user profile to your product categories. What are more conceptual attributes you could use on your PDPs? “Moody”, “Warm”, “Dark Academia”.
- Prioritize High-Coverage Categories: Focus data integrity efforts on Apparel, Electronics, and Home Goods, as these are the categories where the LLM is most likely to display your product directly to the user.