AI engines
Alexa for Shopping and Rufus: what brands can measure
Amazon folded Rufus into Alexa for Shopping on May 13, 2026, and 400 million customers used it in 12 months. No Amazon API returns the answers it gave.

By Mathias C.SEO consultant, and more recently GEO, at MentionBox
Amazon retired the Rufus name on May 13, 2026. The assistant was merged with Alexa+ and relaunched as Alexa for Shopping, in the main Amazon search bar, on amazon.com, in the Shopping app and on Echo Show, with no Prime membership or Echo device required in the United States. Amazon's third-quarter 2026 results, published on October 29, 2026, report that it "was used by 400 million+ customers in the last 12 months, with customer interactions up 570% year-over-year", with "100+ new capabilities" added. For a brand, the consequence is narrow and awkward: the assistant now sits between a shopper and the product page on every Amazon surface, and Amazon publishes no API that returns what it answered. The measurable part today is the advertising layer, where the Prompts report has carried per-prompt metrics in the United States since March 25, 2026, plus whatever a third-party vendor collects from the live product.
This page is written for an e-commerce or brand team that has been asked for an Alexa for Shopping number, and for the developer who went looking for an Alexa shopping API and found a skills reference instead. It covers what changed in May, what each official Amazon API around the assistant returns, the one first-party report that exists, what three vendors publish about collecting the surface, and the five figures worth building.
What changed on May 13, 2026
The rename was the small part of it. Rufus was a panel a shopper opened; Alexa for Shopping answers from the box people already type in, and in Amazon's words it "combines deep product knowledge, in-depth information from across the web, and powerful shopping capabilities" with a customer's preferences and history. Amazon's announcement lists the surfaces around it: AI-generated overviews "at the top of search results in the Amazon Shopping app", AI overviews on product detail pages, comparisons generated from search results, a year of price history, Scheduled Actions for routine purchases, and discovery of products "from stores across the web through Shop Direct", where "the Buy for Me agentic AI feature handles the entire purchase on your behalf".
Three of those change what a brand is optimising for.
- The overview above the results. The first thing a shopper reads about a category is now a paragraph written by the assistant, not a row of listings.
- The compressed set. A comparison or a recommendation names a handful of products out of a page that used to show dozens, and the attributes it compares come from listing data.
- The agent off Amazon. Through Shop Direct and Buy for Me, an Amazon agent reads a brand's own storefront to buy on a customer's behalf.
Amazon also sells the approach to other retailers: its AWS Agentic Shopping Assistant packages architecture guidance and starter code so a retailer can launch its own conversational experience "in weeks", and the same page credits Amazon's assistant with "nearly $12 billion in incremental sales for Amazon last year alone".
The official APIs around the assistant, and what each returns
Searching for an Alexa shopping API returns a real Amazon reference, and it answers a different question than the one most teams are asking.
| Amazon surface | What it is for | What it returns about the assistant |
|---|---|---|
| Alexa Shopping Actions for Alexa Skills | letting a skill you built add to a list or cart, buy, or recommend up to 10 products | nothing; the four actions are outbound, invoked through Connections.StartConnection with an array of ASINs |
| Alexa+ MCP Toolkit | connecting your own MCP server so the assistant can use your capabilities | nothing on recommendations; Alexa+ "handles natural language understanding, response generation, and user interface (UI) rendering" |
| Amazon Ads Prompts report | measuring the paid prompts that open an assistant conversation | prompt text, impressions, clicks, CTR, CPC, spend, sales, ACOS, ROAS, 7-day orders and units |
| Selling Partner API, Brand Analytics | search and purchase reporting on your ASINs | five report types, keyed on search terms and purchases, none on assistant answers |
| Creators API | programmatic catalogue data for publishers and affiliates | product data; it replaces Product Advertising API 5.0, which now answers 403 with an AccessDeniedException |
From each Amazon documentation page named in the first column, read on November 2, 2026.
The skills reference explains the confusion. It documents four actions, connection://AMAZON.AddToList/1, connection://AMAZON.AddToShoppingCart/1, connection://AMAZON.BuyShoppingProducts/1 and connection://AMAZON.RecommendShoppingProducts/1, each taking a products array of ASINs, with a one-item limit on the first three and ten on the last; two of them are marked a developer preview that "might change as Amazon receives feedback". The API lets your skill put products in front of a customer, and never reports which products the assistant put in front of anyone else.
The first-party layer that does report: advertising
One Amazon product reports on the assistant prompt by prompt, and it is a paid one. Sponsored Products prompts and Sponsored Brands prompts are described as "a new AI-powered enhancement to your existing campaigns that automatically engages shoppers with relevant product information": they appear on search results pages and product pages, and a click either opens an assistant dialog or answers on the page. Open beta started in November 2025, general availability in the United States on March 25, 2026. Existing campaigns are enrolled automatically, the clicks are "charged as part of your bid and billing settings on a cost-per-click basis", and a prompt can be paused at any time in the Ads Console.
In the Ads Console, Reports, Create report, report category Sponsored Products, report type Prompts returns prompt text, the associated ad, impressions, clicks, click-through rate, cost per click, spend, sales, ACOS, ROAS and 7-day orders and units, and the same data is reachable through the Amazon Ads API. Two published figures frame what to expect: Amazon states that "nearly 20% of shoppers who interact with a prompt continue the conversation about that brand", a figure Andy Jassy repeated for brand prompts in the assistant in his first-quarter 2026 earnings commentary, and Amazon Ads puts the lift at 6% in conversions when prompts are added to a Sponsored Brands ad.
A second brand-facing lever is dated but not live: Branded Conversations, where a brand supplies the expertise the assistant uses after a shopper clicks a Sponsored Brands instruction, is set to launch in January 2027 in the United States. Until then, a brand shapes the words the assistant chooses through listing content, and the surface through bids.
What third-party vendors publish about collecting the surface
Three of the 55 APIs this site rates name the Amazon assistant in their own engine lists, under three different labels, and their published detail differs enough to matter.
| API | Label in its engine list | Collection method published | Published price | Main limit on this surface |
|---|---|---|---|---|
| Promptwatch | Amazon Alexa | yes: Alexa answers collected from the live product, runs made from the monitor's country | $15.83 per 1,000 answers, $95 Essential plan covering 50 prompts and 6,000 answers | analytics at the product-appearance level; no share of voice |
| Peec AI | Amazon Rufus UI, model amazon-rufus-scraper | partly: its pages state UI scraping for most engines, without naming this one | not published; API documented as Enterprise only | the channel carries an unsupported_country_codes list |
| LLM Pulse | Alexa for Shopping, next to Walmart Sparky | general statement only: it collects answers visible to the user | €30.67 per 1,000 answers, €299 Scale plan, add-on models from €8 per model per month | the assistant is a paid add-on, not an included engine |
Engine labels, methods and prices from each vendor's own pricing and documentation pages, checked on October 23, 2026.
Promptwatch is the one whose shopping payload is documented endpoint by endpoint. Its shopping product appearances endpoint is a GET /shopping/items, described as a "paginated list of product appearances captured in AI shopping answers", and its filters are what make it usable on one surface:
curl -G https://api.promptwatch.com/shopping/items \
-H "Authorization: Bearer $PROMPTWATCH_KEY" \
--data-urlencode "models=<alexa model id>" \
--data-urlencode "promptTypes=ORGANIC" \
--data-urlencode "sortBy=positionInResponse" \
--data-urlencode "from=2026-10-01T00:00:00Z"
Each item carries title, priceText, merchants, rating, numReviews, featuredTag, productUrl, productDomain, positionInResponse, cite and a priceStats object with average, high, low and offer count. One field decides whether the join to your catalogue is easy: foreignId, documented as the "exact (case-insensitive) product id in the source answer engine", which on this surface is the ASIN. The promptTypes filter separating ORGANIC, BRAND_SPECIFIC and COMPETITOR_COMPARISON keeps a branded trigger out of a category figure.
Peec AI publishes the plainest evidence that the surface exists as a channel. Its model channels endpoint names the Amazon entry and limits it by country:
{
"data": [
{
"id": "amazon-0",
"description": "Amazon Rufus UI",
"current_model": { "id": "amazon-rufus-scraper" },
"is_active": true,
"unsupported_country_codes": ["AE", "AR", "AT"]
}
]
}
That unsupported_country_codes array is the honest part of the picture, and every vendor has an equivalent whether or not it publishes one. The assistant is a United States product today, so a European brand measuring it is measuring the United States market.
One vendor sells this single surface on its own. Pangolinfo's Amazon Alexa page documents an API that submits a prompt and returns the assistant's answer with recommended ASINs, prices, ratings and the two to five follow-up questions it suggests, at a latency it puts at roughly 30 seconds per call. Its plans are $19, $99 and $369 a month at 6 credits a conversation, so about 1,600, 10,000 and 40,000 conversations, which is $11.88, $9.90 and $9.23 per 1,000 conversations by our arithmetic, read on November 2, 2026.
For the assistants outside Amazon, the computed layer is further along. As a concrete example, API for GEO answers POST /v1/shopping/products with the product as the engine named it, your own externalId, visibility, mean price, rating, the merchants credited and a share of voice, at $13 per 1,000 answers down to $7 at its largest tier, over ChatGPT, Perplexity, Google AI Mode, Claude, Copilot and Grok as published on its pages, with Gemini, Google AI Overviews, DeepSeek and Meta AI confirmed in our hands-on test of October 26, 2026 rather than documented on apiforgeo.com. It does not list the Amazon assistant, which is the pattern across the market: shopping metrics get built where answers are reachable without a retail login. MentionBox, which publishes this site, has a commercial link with API for GEO. Our guide on tracking products in ChatGPT Shopping and Copilot covers that side in full.
Five figures a brand can actually build
Collection gives conversations, and a report needs counts. Each of these needs one field from the sections above.
| Figure | Formula | Worked on 320 collected conversations |
|---|---|---|
| ASIN appearance rate | conversations naming one of your ASINs ÷ conversations collected | 61 of 320: 19.1% |
| Share of the named set | your product mentions ÷ all product mentions in those conversations | 74 of 1,160: 6.4% |
| Mean position in the set | average positionInResponse of your ASIN when it appears | 3.4 |
| Follow-up coverage | follow-up questions naming your brand or a differentiating attribute ÷ follow-up questions returned | 18 of 912: 2.0% |
| Quoted price gap | mean price shown − your listed price, over your listed price | $148 against $139: 6.5% |
Formulas ours, worked on an illustrative month of 40 prompts run eight times. The field each one needs is named in the vendor references above.
The first two are the headline and the context: an appearance rate alone moves with the prompt list, while the share of the named set moves when a competitor's listing improves. The third one carries more weight here than on a web surface, because an assistant in a search bar names a short set rather than a page of results, and sliding from third to sixth inside that set can halve the clicks without changing the appearance rate at all.
The last two are specific to this surface. Follow-up questions are the assistant's own fan-out, and one it keeps asking about compatibility, sizing or material points at a gap in your attributes; our explainer on query fan-out in AI search covers how those secondary questions work on other engines. The quoted price gap catches a stale offer or a third-party seller undercutting the brand on the same ASIN.
Mistakes that produce a number nobody can defend
- Buying the skills API. The Alexa Shopping Actions reference is an outbound interface for a skill you built, with two of its four actions in developer preview. No request in it reports an assistant recommendation.
- Reading paid prompt metrics as visibility. The Prompts report counts clicks on an ad that opens a conversation. It says nothing about the organic answer a shopper gets when they type the same question and ignore the ad.
- Expecting Brand Analytics to cover it. Five report types, all keyed on search terms, ASINs and purchases. Assistant-led demand lands in them stripped of the question that produced it.
- Mixing the two names in one series. A vendor column labelled Rufus and one labelled Alexa for Shopping describe the same surface before and after May 13, 2026. Merged without a note, the rename looks like a traffic event.
- Comparing an Amazon number with a ChatGPT number. The assistant answers from a catalogue Amazon controls, inside a store, to a logged-in shopper with a purchase history, while a general assistant answers from the open web. Our guide on GEO metrics sets out the definitions that survive a comparison.
- Measuring a country you do not sell in. A monitor pointed at the wrong market returns thin data that reads as a decline, and at least one vendor publishes an unsupported-country list on its Amazon channel.
Our view
For a brand on Amazon, in our view the honest reading of November 2026 is that the surface is large and the instrumentation is small: one first-party report exists and it is an advertising report, the lever that shapes the answer itself has a date in January 2027 and no reporting described yet, and everything organic runs through vendors collecting from the live product, at entry tickets between $95 a month and a €299 plan plus an add-on, on a single country. We would treat the Prompts report as the weekly number, because it is attributable, and collected conversations as a quarterly listing audit rather than a dashboard. A few hundred conversations will not move a forecast; they will tell you which attributes the assistant keeps asking about, which competitor ASINs sit beside yours in a compressed set, and whether the price it quotes is yours.
Frequently asked questions
Is there an official Alexa for Shopping or Rufus API?
There is no Amazon API that returns what the assistant answered a shopper. The Alexa Shopping Actions for Alexa Skills reference documents four actions a skill can invoke, add to list, add to shopping cart, buy products and recommend products, with a maximum of one item per request for the first three and ten for the last, and two of the four are still labelled a developer preview. The Alexa+ MCP Toolkit works the other way round, letting a brand expose its own capabilities to the assistant, and the Selling Partner API's five Brand Analytics reports cover search queries, catalogue performance, market baskets, search terms and repeat purchases, none of them assistant answers.
What happened to Amazon Rufus?
Amazon merged it with Alexa+ and renamed it Alexa for Shopping on May 13, 2026, moving it into the main Amazon search bar on the Shopping app and on amazon.com, with no Prime membership or Echo device required in the United States. Amazon's announcement credits Rufus with helping over 300 million customers in 2025, and its October 29, 2026 filing reports 400 million customers for Alexa for Shopping over the following twelve months with interactions up 570% year over year. Dashboards and vendor engine lists still carry the old name, so a column labelled Rufus and a column labelled Alexa for Shopping usually describe the same surface.
Can I see which products Alexa for Shopping recommends in my category?
Only by paying a vendor to collect the answers from the live product, because Amazon publishes nothing of the sort. Three of the 55 APIs this site rates list the Amazon assistant among their engines: Promptwatch at $15.83 per 1,000 answers inside a $95 Essential plan that covers 6,000 answers a month, Peec AI with an `amazon-rufus-scraper` channel whose API access its documentation limits to Enterprise customers, and LLM Pulse where the assistant is a paid add-on model from €8 per model per month on top of a €299 Scale plan. Prices checked on October 23, 2026.
Does Amazon Brand Analytics report AI assistant answers?
No. The Selling Partner API documents five Brand Analytics report types, `GET_BRAND_ANALYTICS_SEARCH_QUERY_PERFORMANCE_REPORT`, `GET_BRAND_ANALYTICS_SEARCH_CATALOG_PERFORMANCE_REPORT`, `GET_BRAND_ANALYTICS_MARKET_BASKET_REPORT`, `GET_BRAND_ANALYTICS_SEARCH_TERMS_REPORT` and `GET_BRAND_ANALYTICS_REPEAT_PURCHASE_REPORT`, and they are keyed on search terms, ASINs and purchases. A rise in assistant-led discovery shows up in them as impressions and clicks with no query attached, which is a signal that something moved rather than a measurement of it.
Do ads appear inside Alexa for Shopping, and can they be measured?
Yes, and that layer is the one with real reporting. Sponsored Products prompts and Sponsored Brands prompts opened in beta in November 2025 and went generally available in the United States on March 25, 2026, charged on a cost-per-click basis inside existing campaigns, which are enrolled automatically and can be paused prompt by prompt. The Prompts report carries prompt text, the associated ad, impressions, clicks, click-through rate, cost per click, spend, sales, ACOS, ROAS and 7-day orders and units, and Amazon reports that nearly 20% of shoppers who interact with a brand prompt continue the conversation about that brand.
How much does a month of Alexa for Shopping tracking cost?
Forty category prompts run twice a week on that one surface is 320 collected conversations a month, which is small enough to fit inside every entry plan published. On the rates checked on October 23, 2026 that volume uses 320 of the 6,000 answers in Promptwatch's $95 Essential plan, and on LLM Pulse it is a €8 add-on model on top of a €299 Scale plan. A vendor selling the Amazon assistant on its own bills about 6 credits per conversation, which works out between $9.23 and $11.88 per 1,000 conversations on its three published plans.