Key takeaways
- ChatGPT Shopping and Google AI Mode need their own prompt sets. Evaluation questions like "best running shoes for a marathon" behave differently from short, purchase-ready prompts like "where to buy [product] online."
- Promptwatch records every product card (price, merchant, rating, review count, carousel position) from tracked ChatGPT and Amazon Alexa responses, then matches cards to your own catalog by exact product ID.
- Shopping card placement is inconsistent response to response, so you judge it over weeks using appearance rate and average position, not a single screenshot.
- Shopping Insights, the feature that unlocks product-card tracking, is included from the Business plan ($579/month) upward.
- Run separate monitors per market and language if you sell internationally. Position and merchant data differ by country.
Why ChatGPT Shopping and AI Mode need a different setup than regular brand tracking
Most AI visibility tracking starts with brand mentions: does ChatGPT name your company when someone asks an unbranded question. That's step one, and it still matters. But shopping is a different animal. When OpenAI introduced shopping research in ChatGPT, it built a flow where a shopping question triggers a suggested "shopping research" mode, the user answers a few clarifying questions about budget and preferences, and ChatGPT surfaces product cards, not just prose.
That means two separate things are happening in the same conversation. There's the written answer, which names brands and cites sources the way any other ChatGPT response does. And there's the product carousel, which pulls structured data (price, merchant, rating) the same way a shopping ad network would. Tracking only the written answer misses half the picture. Tracking only the carousel misses the brand-level context that got you there in the first place.
Google AI Mode behaves similarly to AI Overviews in that it pulls from indexed pages and increasingly layers in shopping-style modules for commercial queries, though its product card mechanics are less uniform than ChatGPT's. The setup principle is the same either way: separate your prompts by intent, track both the text and the card, and give it time before you draw conclusions.
Step 1: Build two prompt sets, not one
The mistake most teams make is running the same prompt list for brand tracking and shopping tracking. Don't. Promptwatch's own ecommerce tracking research found that shopping cards attach far more reliably to short, purchase-completion prompts than to evaluation questions. So you want:
Commercial evaluation prompts. These are the "best X for Y" style questions, e.g. "best shoes to run a marathon." They surface brand recommendations and written citations. Keep 15-30 of these per category.
Shopping completion prompts. These are one product, one buying intent, roughly 5-15 words. Think "where to buy [marathon shoe model] online" or "[marathon shoe model] on sale with free shipping." These are the prompts that actually trigger product cards.
If you're not sure where to start, Promptwatch can generate suggested shopping prompts directly from your unmatched products, up to 50 products per run, in the language and country of the monitor you're setting up. That saves you from guessing at phrasing that shoppers in, say, Germany or Brazil would actually type.
Step 2: Set up your monitor and pick the right models
Inside Promptwatch, go to your Brand Hub and configure the basics first. Fill in your target personas, since this shapes how Promptwatch generates prompt suggestions that match how your actual buyers search, not generic category terms.
From there, create a monitor and choose your models. For shopping tracking specifically, ChatGPT is the priority, since it's the platform with a dedicated shopping research flow and product cards. Amazon Alexa responses are also parsed for product cards in Promptwatch. For broader commercial-intent tracking, add Google AI Mode and Google AI Overviews alongside Perplexity and Gemini if your buyers use them.
One monitor per market is the right structure if you sell in more than one country. Merchant names, prices, and even which products show up at all can differ by region, so mixing US and UK shopping prompts into one monitor just muddies the data.
Step 3: Upload your catalog so cards actually match your products
This is the step that makes or breaks product-level tracking. Promptwatch extracts product cards from tracked ChatGPT and Amazon Alexa responses and stores the product name, price, merchant, rating, review count, and carousel position for each one. But none of that is useful until it's tied back to your own SKUs.
You upload your catalog as a CSV, and Promptwatch matches cards to your products by exact product ID, a Shopify product ID or Amazon ASIN, for instance. If the ID in your CSV doesn't match the ID the AI platform surfaces, the card won't link to your catalog even if it's visually your product. Double-check your export before uploading; a mismatched ID format is the single most common reason teams think "tracking isn't working" when it's really a data format issue.
Once matched, Promptwatch flags unmatched products too, which is what feeds the automatic shopping prompt suggestions mentioned in step 1.
Step 4: Let the monitor run for at least a week before judging anything
This is the part people get impatient about. Shopping cards attach to individual responses, and the same prompt can return a card one day and plain text the next. That's not a bug, it's how the system works. Promptwatch's own guidance on ecommerce tracking is blunt about this: implement the monitor, then check it after eight days, not eight hours.
When you do check in, read two fields: appearance rate (how often the card shows up across repeated runs of the same prompt) and average position (where your product sits in the carousel when it does appear). A single screenshot showing your product in position two means nothing if the appearance rate over two weeks is 10%.
Step 5: Read the merchant field like it matters, because it does
The merchant field tells you who actually captures the click when your product appears in a shopping card. If your product shows up but a marketplace reseller is listed as the merchant instead of your own store, you're winning visibility and losing the sale. This is worth checking per product, not just per prompt, since the same item might route to your site in one response and to a reseller in another.
| What to track | Where it lives in Promptwatch | Why it matters |
|---|---|---|
| Product, price, rating, review count | Shopping Insights (product cards) | Confirms your listing data is accurate in AI's eyes |
| Merchant field | Shopping Insights (product cards) | Shows who gets the click, you or a reseller |
| Carousel position | Shopping Insights (product cards) | Tracks whether you're first or buried |
| Written brand mentions and citations | Brand/citation monitor | Shows why AI picked the brands it did |
| AI crawler activity on product pages | Agent Analytics (crawler logs) | Confirms AI can actually read your product data |
Step 6: Add crawler logs to see the "why" behind the numbers
Shopping and brand numbers tell you what's happening. Crawler logs tell you why. If ChatGPTBot or GoogleOther isn't hitting your product pages, no amount of prompt tuning fixes a card that never had a chance to appear. Promptwatch's Agent Analytics shows real-time crawler visits, which pages got read, and which hit errors, so you can tell the difference between "AI doesn't rank us" and "AI can't even reach the page." This connects directly to Promptwatch Data's research on average sources per response, which is a useful baseline for how many sources AI models typically pull per answer across ChatGPT, Claude, Perplexity, and Gemini: if your product page isn't even in the crawl pool, it was never competing for one of those slots.
Step 7: Check the fan-outs, not just the final answer
Promptwatch records the fan-out, the actual sub-queries ChatGPT runs, behind every tracked prompt. For ecommerce prompts, this often reveals ChatGPT running site: searches against individual vendors' own pricing pages, or appending the current year and words like "pricing" and "reviews" to otherwise generic category queries. Promptwatch's query fanout research tracks how many searches ChatGPT triggers per response and how that's trended, which is useful context for how aggressively it's checking sources before it builds a shopping answer. If your pricing page returns a 404 or blocks crawlers, that fan-out step is where you lose the placement, long before any card gets rendered.
A realistic rollout plan for your first month
Week 1: build your two prompt sets, upload the catalog, launch one monitor per market.
Week 2: resist the urge to react to day-to-day swings. Let appearance rate data accumulate.
Week 3: pull the crawler logs and fix any product pages that are blocked or erroring for AI bots.
Week 4: compare merchant data against your own storefront. If a reseller is consistently winning the click on your top products, that's your first action item.
Where this fits if you sell outside the US
If you run international stores, remember that shopping card behavior, merchant names, and even whether ChatGPT shows cards at all can vary sharply by country. Running a single global monitor will average away real differences. Set up separate Promptwatch monitors per market from the start, generate market-specific shopping prompts rather than translating your US list, and compare appearance rates across markets rather than assuming what works in the US automatically works in the UK or Germany.
A note on where Shopping Insights fits in pricing
Product card tracking through Shopping Insights is included starting on Promptwatch's Business plan, at $579 a month, which also includes five tracked sites, 350 prompts, and 42,000 responses. If you're earlier stage and mainly need brand-level AI visibility without the shopping card layer, the Essential ($95/month) or Professional ($245/month) tiers cover ChatGPT, Gemini, Perplexity, and AI Overviews tracking without the product-card parsing. It's worth starting with brand tracking on a lower tier and upgrading once you have catalog data ready to upload, rather than paying for Shopping Insights before you've built your prompt sets.
If you want a wider view of how other AI visibility tools handle ecommerce tracking specifically, or you're comparing options before committing, the GEO software directory at bestgeosoftware.com is a reasonable place to see how the field lines up. For agency teams managing this across multiple clients, 1001 SEO Media builds out GEO and AI search monitoring programs using this exact stack, prompt sets, crawler logs, and content fixes together, rather than treating tracking as a standalone report nobody acts on.