manifoldmcp/marketing-skills / skills/google-shopping
google-shopping
VerifiedWhen the user sells products online and wants Google Shopping research, product prices, sellers and reviews, price comparison, a product feed and Merchant Center check, or ChatGPT shopping results. Holds dated, sourced findings a model lacks (ChatGPT took most product picks from merchant feeds from July 2026 and dropped Instant Checkout, Google removed free listings from European search in September 2026 and turned native checkout on by default for Shopify stores, the review floors for stars, the 2026 feed attributes) and how to read live Shopping prices, sellers and reviews. Also use when the user mentions Google Shopping, Merchant Center, a product feed, free listings, who sells our product cheaper, price competitiveness, product reviews or star ratings, seller ratings, product structured data, ChatGPT shopping, the Agentic Commerce Protocol, UCP, agentic checkout, or a Shopping plan. Amazon goes to amazon, bids and campaigns to paid-ads, brand answers in AI to ai-search, web rankings to seo.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 12.6 KB |
| references/ai-shopping.md | markdown | 10.1 KB |
| references/feed-audit.md | markdown | 7.2 KB |
| references/field-notes.md | markdown | 11.0 KB |
| references/price-check.md | markdown | 6.6 KB |
| references/ranking.md | markdown | 7.9 KB |
| references/reviews.md | markdown | 5.9 KB |
| references/strategy.md | markdown | 4.8 KB |
references/field-notes.md
View on GitHub# Field notes
What Manifold has seen in its own runs on product searches, United States, English, desktop. Small samples: read them as evidence of how the surfaces behave and what the fields contain, not as market shares. Newest first.
## The Shopping tools on live data (16 October 2026)
Inputs: `shopping_search_products` on "espresso machine" (four fresh fetches within five minutes, at `depth` 40, 10, 20 and 39, and one cached repeat), "brooks ghost 17 running shoes" (40) and "running shoes" (80); `shopping_get_product`, `shopping_get_sellers` and `shopping_get_reviews` (100 and 50 reviews) on the Breville Bambino Plus and the Brooks Men's Ghost 17; `aeo_run_ai_answers` on "best espresso machine for beginners under $500" and "best running shoes for beginners" on `ai_mode` and `ai_overview`; five more `shopping_get_sellers` calls on ids taken from those answers. 166 fresh search rows, 2 product records, 48 seller rows, 150 reviews, 4 answer cells; 39 credits. Run after the rank fix merged that day.
**Rank after the fix.** Carousels no longer take rank numbers: the first "espresso machine" fetch ran 1 to 40 with no gap. Results the tool drops for having no product id still keep their place, though: "running shoes" at `depth` 80 returned 48 rows starting at rank 15, the Brooks query 12 rows spread over ranks 1 to 40, and "espresso machine" at `depth` 20 returned 17 rows ranked 1 to 39. What sat in the gaps (sponsored offers or offers with no product page) is not in the data. The tool now numbers the rows it returns 1 to N and fills a short `depth` from the whole billed page; the gaps showed that up to half a page can be such results, so fewer rows than `depth` still come back. Every row's `url` searched Google for "espresso machine sale", Google's own rewrite of the query; it now carries the searched keyword. No row marks Sponsored; `tags[]` held discount badges ("15% OFF"), "LOW PRICE", and once "360" and "3D".
**One fetch is one draw.** The 40- and 39-deep fetches of "espresso machine", five minutes apart, shared 19 of about 40 products and 2 of their top 10; the Bambino Plus was rank 1 in one and rank 20 in the other, under a different store. The 20-deep fetch shared 2 of its 17 products with the first. The cached repeat matched the first exactly and cost nothing.
**`gid` is the product; `product_id` follows the offer.** The same Lelit Bianca and Bambino Plus came back in later fetches under a new `product_id` and `data_docid` with a new headline store, and the same `gid`. Rows are offers: one fetch held the same product twice in a row (ranks 24 and 25) and the same `product_id` at Target and at Best Buy; the 80-deep fetch held one product on both of its 40-row pages (ranks 39 and 63). One product, the Fellow Series 1, appeared under two `gid`s.
**Store mix.** The first "espresso machine" fetch: 24 stores in 40 rows, Target 5, Best Buy 4, Sur La Table, Home Depot and Seattle Coffee Gear 3 each, Walmart 2, brands' own stores (Breville, SharkNinja, Nespresso, Fellow, Gevi) 5; prices $54.99 to $2,999.95, top-10 median about $700. Amazon was in none of the 166 search rows or 48 seller rows; eBay sellers held 2 search rows and 12 seller rows.
**Likely fake shops rank and list.** A store with no rating, corretticatering.com, held search ranks 2 and 14 with a Breville Barista Express Impress at $299.99 and a Barista Pro at $288 (big retailers asked $799.95 for the first). Seller lists held a Bambino Plus cluster of three unknown shops at $314.30 (retail $499.95) and Gaggia Classic Pro offers at $104.65 from a catering site and a jewellery site (retail about $500), all without `seller_rating`.
**Seller lists are short.** Seven `shopping_get_sellers` calls returned 5, 5, 3, 10, 10, 10 and 5 stores: the list stopped at 10, the vendor's default, and is often 5. The tool now asks for 20, which the vendor documents and the price covers; not yet seen live. The Bambino Plus list lacked Breville's own store, Walmart and Amazon; the Ghost 17 list lacked Running Warehouse, the store on that product's search row. Tax was empty on 48 of 48 rows and condition on 48 of 48 (eBay listings included), so the tool no longer returns either; `shipping` was filled on 8, where `total_price` added it, and elsewhere `total_price` equalled `price`, so an empty `shipping` is free or unknown. `note` mixed Google's labels ("Best price", "Most popular", "Usually $110", "26% off $150") with store promotions ("30% Off FALL", "$20 Off $100+ YAYFALL"); "Best price" went to Best Buy when all five Bambino Plus stores asked $499.95. Seller URLs carry each store's own tracking (`utm_source=feed`, "FreePLA", a paid-social tag), which says how the store tags its feed, not whether the offer was an ad.
**Product records.** `specs[]` had 44 rows (espresso machine) and 76 (shoe), every group empty (specs now carry only name and value); the shoe's specs included the variant shown ("Men's Size 10.5"), two dozen "No" flags (golf shoes, wedges) and one name with an HTML entity, which the tool now decodes. `rating` and `ratings_count` were exact (4.38 from 4,153) where search rows carry Google's rounded display (4.4, 4200). `description` and `url` came from one store, for the Ghost 17 a Brooks resale storefront. `offers[]` listed the same stores as `shopping_get_sellers` for both products, with `regular_price` where a store showed a strikethrough, and no shipping.
**Reviews.** `stars[]` summed to 4,122 against `ratings_count` 4,153 on the Bambino Plus but to 2,688 against 2,156 on the Ghost 17 (25% more). The rows over-represent low stars: one-star was 12% of the Bambino Plus rows against 7% in `stars[]`, and 10% against 5% on the Ghost 17. The 100 Bambino Plus reviews came from 16 stores, about 30 of them in Canada, Australia, Switzerland and the UK; half the Ghost 17 reviews came from Brooks' Australian and New Zealand stores, others from Canada, the UK, Malaysia and a store demo site. Every `published_at` carried the crawl's day of month and time: month-precise, and a quarter of the Bambino Plus rows sat exactly one year back, the bucket for reviews a year old or more. 50 reviews were billed 3 credits.
**Ids from AI answers work.** AI Mode and AI Overview product links carry `catalogid`, `gpcid`, `headlineOfferDocid` and `productDocid` (links in the text) or `catalogid`, `gpcid` and `productid` (cited sources). On the one product also in a Shopping fetch (Ninja Luxe Café Mini), `catalogid` was its `product_id`, `gpcid` its `gid`, and `productDocid` its `data_docid` (the tool's own `url` files that docid under `headlineOfferDocid`). All five `shopping_get_sellers` calls on AI-answer ids returned stores, including one with `product_id` alone. The two docids gave different lists for the Gaggia Classic Pro: 3 stores with `productDocid`, 10 with `headlineOfferDocid`. The AI Mode link for the Bambino Plus pointed at another listing than Shopping's own (other `catalogid` and `gpcid`), the cluster of unknown shops at $314.30, one more out of stock, and Newegg at $597. AI Mode linked 4 to 8 Shopping pages an answer; AI Overview answered the espresso prompt with five product cards and did not answer the running shoes one.
**Cost and speed.** Billing matched the descriptions: `depth` 80 was 2 credits and 10 to 40 was 1, a cached repeat and a refused call 0, an AI Overview cell with no answer was charged. `shopping_get_sellers` returned data 15 to 22 seconds after the call. A location code ("2840") was refused as an invalid location name; the tools now send a code as a code, and names work as before.
## Product prompts on five AI engines and Google's results page (15 October 2026)
Inputs: `aeo_run_ai_answers` on four prompts ("best cordless vacuum for pet hair", "best espresso machine for beginners under $500", "best carry on luggage", "where can I buy a standing desk cheapest") on `chatgpt`, `gemini`, `perplexity`, `ai_overview` and `ai_mode`, with ten `brands` (Amazon, Walmart, Target, Best Buy, Costco, Home Depot, Wayfair, Dyson, Breville, FlexiSpot); `seo_get_serp` with `ai_overview: true` on "best cordless vacuum for pet hair", "espresso machine", "standing desk" and "best carry on luggage". 20 answer cells (19 answered) and 4 result pages, 64 credits. Models reported: gpt-6 (ChatGPT app), Gemini 3.5 Flash-Lite (app), sonar-reasoning-pro.
**Google's AI answers link products to Shopping listings.** All three AI Overviews and all four AI Mode answers linked product names to Google Shopping product pages (2 to 13 links an answer) and cited those pages as sources; the Gemini app did on two of four. ChatGPT and Perplexity never linked a Google Shopping page. The links carry `catalogid`, `gpcid` and `headlineOfferDocid`, the parameter names under which `shopping_search_products` rows carry `product_id`, `gid` and `data_docid` in their Google Shopping URLs. Seller calls made from such ids work; see the 16 October notes for which docid to pass.
**A third of ChatGPT's links came through feeds.** Of 33 ChatGPT citations, 10 carried a store's catalogue tags: Walmart and Best Buy links with `utm_medium=feed`, Travelpro and a coffee store with `utm_campaign=openai_catalog`, Target with `afid=chatgpt`, Wayfair with a `refid` code. ChatGPT linked store product pages (Walmart, Best Buy, Target, Home Depot, Macy's, Wayfair, the brand's own store) far more than reviews sites; Perplexity linked Best Buy's search and category pages and editorial reviews (RTINGS seven times on one answer); Gemini linked only Google Shopping and search pages on two answers and review sites on the other two.
**Stores named and linked.** Across 19 answers Walmart was named 6 times, Amazon 4, Best Buy 3, Wayfair 3, Target 2, Home Depot 2, Costco never. Amazon was named in four answers but linked by ChatGPT in none; its only links were an Amazon search page and a best-sellers page in AI Overviews and one Perplexity link. On "where can I buy a standing desk cheapest" ChatGPT linked three Walmart product pages, one Target and one Home Depot; AI Mode named FlexiSpot first, then Amazon and Walmart.
**`named_brands[]` holds products on ChatGPT.** On these prompts it listed product names ("Dyson V15 Detect", "BISSELL PowerClean FurFinder", three long Walmart standing desk titles) and once a category phrase ("cordless stick vacuum"), never a store. Count stores with `brands` and `mentions[]`.
**Which queries get which block.** Head terms got product blocks ("standing desk": Popular products and Deals on Standing Desks, no AI Overview; "espresso machine": both product blocks and an AI Overview). The two "best X" queries got an AI Overview and no product block. AI Overview was absent for "standing desk" and for the "cheapest" prompt. Product block rows in `seo_get_serp` carry no products (no url, no title beyond the block's name). No paid rows appeared on any page, as the paid-ads skill's notes also found.
**ChatGPT's fan-out names its picks.** Each ChatGPT answer ran one search, and three of four named the products or sources it already had in mind ("best espresso machines under $500 beginners 2026 reviews Breville Bambino Plus Gaggia Classic Pro"; "cheapest standing desks under $150 2026 Walmart Amazon IKEA Home Depot").
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |