# Muse Spark is not Muse: what model-based trackers miss about Meta’s shopping agent | Enrich Labs > We asked Meta’s Muse agent and the Muse Spark model the same 100 shopper questions in five categories. Across 200 agent answers, Muse never linked a product to Amazon, routed shoppers to different stores, and changed its brand picks far more between conversations than the model does. Research from Enrich Labs. _Source: https://www.enrichlabs.ai/research/muse-spark-is-not-muse_ --- ## Key findings 1. **The Muse agent never sent a shopper to Amazon.** Across 200 answers and roughly 436 recommended products in five categories, not one linked to Amazon. Muse Spark, asked the identical questions through Meta’s API with web search, linked Amazon in every category, from 2% of picks in running to 19% in cookware.1 2. **The agent favors brand sites and a different set of retailers.** A brand’s own site or an independent store received a larger share of agent picks than Spark picks in all five categories. In beauty the agent sent 36% of picks to Sephora and 3% to Ulta; Spark split them 25% and 25%. In protein, GNC took 10% of Spark picks and none of the agent’s. 3. **The model overstates how stable Muse is.** Two Spark answers to the same question share 60 to 69% of their brands in every category. Two agent answers share 57% in dog supplements and only 21% in beauty. A tracker built on Spark reports a steady ranking that real Muse shoppers do not get.4 4. **The category leaders mostly match; the rest of the list does not.** Zesty Paws leads dog supplements in 70% of answers from both systems, and HOKA, Brooks, CeraVe, Lodge and Optimum Nutrition top their categories under both. Below the leaders, the two systems share between 4 and 8 of their ten most-recommended brands. 5. **Measuring the agent is slow and failure-prone, which is why nobody does it.** An agent answer took three to fifteen minutes of live browsing, and about one attempt in three had to be retried; a Spark answer takes seconds. That economics explains why “Muse tracking” products query the model. It does not make the model the right instrument.5 One limit: the Williams Sonoma flat-bottom option couldn’t be verified. Its page returned an access error, so I didn’t include it. AI The Muse agent, explaining a pick it dropped Cookware, “a carbon steel wok for a gas stove”\* ## Overview Muse is Meta’s personal agent. Ask it for a product and it searches Meta’s product catalog, opens a browser to check store pages, and comes back with two or three specific products, each with a store and a price, next to a carousel of product cards. More than ten million people in the US and Canada set it up in its first six weeks, and Meta says shopping is its most common use.6 For a brand that changes two things at once: being on the short list is now binary, and the store is part of the recommendation, because **Muse decides where the sale happens**. Several AI visibility tools now report a brand’s standing “in Muse.” They do it by querying Muse Spark, the language model behind the agent, through Meta’s developer API. That is a reasonable way to learn which brands the model knows. It is not a measurement of what shoppers using the agent are told, because the agent adds a browser, Meta’s catalog, a set of shopping rules that favor well-known merchants, and an account with memory. We wanted to know how much those additions change the answer. So we ran both. We wrote 20 unbranded shopper questions in each of five categories (dog supplements, beauty, running shoes and gear, cookware, protein and sports nutrition), questions like “a sheet pan that doesn’t warp in a hot oven” or “a stability running shoe for overpronation,” never naming a brand or store. Each question went to the Muse agent in a fresh side chat on a US account, and to Muse Spark through the API with Meta’s web search tool on, with identical prompt text asking for three recommended options with store and price.1,2 Every question ran twice on each system. From each answer we extracted the products the system named as its picks, the brand, and the store it linked.3 100 unbranded shopper questions, 20 in each of five categories 200 answers from the Muse agent, in 27 US accounts, plus 200 from Muse Spark 0 agent recommendations linked to Amazon, out of about 436 21% brand overlap between two agent answers to the same beauty question; 63% for Spark ## The agent never links Amazon The clearest result in the data is an absence. In 200 agent answers, not one recommended product was linked to Amazon. Spark, with the same questions, linked Amazon routinely: one pick in five in cookware, one in ten in protein, and at least a few percent everywhere else. Amazon has reportedly blocked Muse’s browser since September, and the agent recommends from stores whose pages it can open and verify. Whatever the mechanism, the effect for brands is the same: **a product sold mainly on Amazon is invisible to the agent even when the model that powers it would recommend it**. Exhibit 1 * * * The agent linked Amazon in none of its recommendations; the model did in every category. **Share of recommended products linked to Amazon,1** Muse agent vs. Muse Spark, by category 1A recommendation is a product the answer names in its written picks, with the store it links. Products shown only in Muse’s product carousel are not counted. 40 answers per category per system. Source: Enrich Labs Muse visibility benchmark, November 2026 Enrich Labs * * * In the cookware set, Spark recommended a Kitsure dish rack and a Utopia Kitchen skillet on Amazon; the agent recommended neither brand in any answer. The finding is also the most robust one in the study. It held in both agent runs, on every account, in all five categories, and it is the one measurement a model-based tracker cannot make, because the model has no browser and is not subject to which stores admit the agent. ## Where the agent sends shoppers Amazon is the sharpest difference in store routing, not the only one. In every category, a larger share of the agent’s picks pointed to a brand’s own site or an independent store than Spark’s did: 60% against 47% in cookware, 36% against 28% in beauty, 58% against 50% in protein. Exhibit 2 * * * The agent sends more shoppers to brand sites in every category, and to different retailers. **Share of recommended products linked to a brand’s own site or an independent store,1** Muse agent vs. Muse Spark, by category 1Brand sites and independent stores means any linked store other than a named national retailer (Amazon, Target, Walmart, Chewy, Petco, Sephora, Ulta, REI, Dick’s, Williams Sonoma, GNC and similar). Source: Enrich Labs Muse visibility benchmark, November 2026 Enrich Labs * * * The named retailers swap too, and the swaps are category-specific. In beauty, Sephora received 36% of agent picks and Ulta 3%, where Spark gave each about a quarter; with Ulta went the drugstore brands sold there and at Target (Maybelline, Revlon, L’Oréal), replaced by brands Sephora carries (Clinique, The INKEY List, MAC, Fenty). In running, Dick’s Sporting Goods rose from 3% of Spark picks to 10% of the agent’s while REI fell from 13% to 4%. In cookware, Williams Sonoma went from 1% to 11%. In protein, GNC went from 10% to nothing. Dog supplements, where Chewy, Petco and brand sites lead under both systems, is the closest match, and even there Amazon is the difference. These are not small shifts for a retailer, and they are invisible to anyone measuring Muse through the model. **Store routing is also the most stable thing the agent does.** Brand picks changed a great deal between our two agent runs, as the next section shows. The Sephora, Dick’s and Williams Sonoma patterns held in both. ## Same leaders, much less consistency Ask Spark the same question twice and you get mostly the same brands: the two Spark runs shared 60 to 69% of their recommended brands per question in every category. Ask the agent twice and the overlap drops to 45% in running, 37% in protein, 28% in cookware and 21% in beauty. Only dog supplements, a category with a few dominant brands, keeps the agent close to the model.4 Exhibit 3 * * * The model gives nearly the same brands every time. The agent does not. **Average share of recommended brands that two answers to the same question have in common,1** by category 1Overlap is the brands both answers recommend divided by the brands either recommends (Jaccard), averaged over the 20 questions in a category and every pair of runs. Agent runs used different Muse accounts, so agent-vs-agent variation includes account history as well as chance. Source: Enrich Labs Muse visibility benchmark, November 2026 Enrich Labs * * * This variability is not noise in our measurement. It is what shoppers experience. Two people who asked Muse for an eye cream for dark circles were, in our runs, told about Clinique and CeraVe in one conversation and The INKEY List and Sunday Riley in the other. Spark named The Ordinary and CeraVe both times. Two people asking for a long-wearing foundation for oily skin both heard Estée Lauder from the agent, then Fenty Beauty in one chat and MAC and Too Faced in the other, all at Sephora; Spark paired Estée Lauder with Maybelline at Ulta and Revlon at Target or Amazon. Part of this is Muse remembering. Each agent run used a different Muse account, and Muse carries memory across conversations, so some of the gap between runs is account history rather than chance. For a brand, that distinction does not help: your visibility in the agent is a rate across many conversations and many accounts, not a yes-or-no answer, and a single check of “does Muse recommend us” says little in most categories. Exhibit 4 * * * At the top of each category the two systems mostly agree. Below it, they drift apart. **Most-recommended brands and the share of answers naming them,1** Muse agent vs. Muse Spark Category Muse agent Muse Spark Top-10 brands in common **Dog supplements** Zesty Paws 70%, Nutramax (Cosequin, Dasuquin) 35%, PetHonesty 30% Zesty Paws 70%, Nutramax (Cosequin, Dasuquin) 48%, PetHonesty 38% 7 of 10 **Beauty (skin, makeup, hair)** La Roche-Posay 12%, CeraVe 12%, Clinique 10% CeraVe 25%, La Roche-Posay 22%, The Ordinary 20% 4 of 10 **Running shoes & gear** HOKA 28%, Brooks 25%, Nike 15% Brooks 35%, HOKA 28%, ASICS 18% 8 of 10 **Cookware & kitchen** Lodge 15%, Cuisinart 12%, GreenPan 10% OXO 28%, Cuisinart 25%, Tramontina 20% 6 of 10 **Protein & sports nutrition** Optimum Nutrition 22%, Naked Nutrition 12%, Dymatize 12% Transparent Labs 32%, Optimum Nutrition 30%, Orgain 25% 7 of 10 1Share of the 40 answers per system that name the brand among written picks. Agent answers name about 2.2 products and Spark answers 3, so Spark shares run higher for the same brand; compare ranks within a column. Product lines are rolled up to their brand (Cosequin and Dasuquin count as Nutramax). Source: Enrich Labs Muse visibility benchmark, November 2026 Enrich Labs * * * The leaders are where the two systems agree. Zesty Paws tops dog supplements in 70% of answers from both, HOKA and Brooks lead running under both, CeraVe and La Roche-Posay lead beauty, Lodge and Cuisinart lead cookware, and Optimum Nutrition leads protein for the agent while sitting second for Spark. If all you want to know is which two or three brands own a category in Muse, the model is a fair proxy. If you want to know whether your challenger brand is one of the next seven, or where the shopper gets sent to buy it, it is not. ## One question, four answers A single cookware question shows the whole pattern. We asked both systems for “a carbon steel wok for a gas stove.” All four answers agreed on the brands that matter: Craft Wok and Yosukata appear in every one. They disagreed on where to buy. Exhibit 5 * * * One question, four answers: the same brands, sold through different stores. **“A carbon steel wok for a gas stove,”** every pick and its linked store Answer Picks Linked stores **Agent, run 1** Joyce Chen, Craft Wok kitchensupply.com, craftwok.com **Agent, run 2** Craft Wok, Yosukata craftwok.com, us.yosukata.com **Spark, run 1** Craft Wok, Yosukata, Made In amazon.com, amazon.com, madeincookware.com **Spark, run 2** Yosukata, Craft Wok, Made In us.yosukata.com, amazon.com, madeincookware.com Picks are the products each answer named in its written recommendation. The agent’s first answer also explained a product it left out: “the Williams Sonoma flat-bottom option couldn’t be verified, its page returned an access error, so I didn’t include it.” Source: Enrich Labs Muse visibility benchmark, November 2026 Enrich Labs * * * Spark linked Amazon three times across its two answers. The agent linked it zero times, sent the shopper to the brands’ own sites and a specialty kitchen store, and explained in its own words that it had dropped a Williams Sonoma option because the page would not load for it. That sentence is the study in miniature: **the agent recommends from the stores it can open**, and the model does not know which those are. ## What should brands do? If you sell through Muse’s shoppers, or are deciding whether to care, this study suggests a short list: 1. **Measure the agent, not just the model.** A model-based score tells you whether Muse knows your brand. Only agent runs tell you whether Muse’s shoppers are sent to you, and from which store. Treat any “Muse visibility” number without a store breakdown as a model number. 2. **Make sure the agent can open your product pages.** The agent dropped products it could not verify and never linked the one retailer that blocks it. Pages that load slowly, hide price or stock behind interaction, or challenge automated browsers are, to Muse, pages that do not exist. 3. **Do not rely on Amazon as your Muse storefront.** In every category, the agent routed to brand sites, specialty stores and the retailers it could reach. A brand whose only strong listing is on Amazon has no path into an agent recommendation today. 4. **Expect variance and measure it over time.** Muse’s picks change between conversations and accounts. Track your share of answers across many runs, repeated monthly, rather than celebrating or panicking over one chat. 5. **Watch retailer routing in your category.** Sephora over Ulta, Dick’s over REI, Williams Sonoma in cookware: if the agent favors a retailer you are not in, your brand loses visibility it would have under the model. If you want the category tables, the question set, or to compare notes on measuring agents rather than models, write to [seijin@enrichlabs.ai](mailto:seijin@enrichlabs.ai). ## Method and limits We wrote 20 unbranded shopper questions for each of five categories and sent each one, with identical prompt text, to the Muse agent and to Muse Spark.1,2 Agent runs opened a new side chat per question in a US Muse account, approved any browsing step the agent requested, discarded Muse’s interim progress notes, and waited for the answer to stop changing. Spark runs called the muse-spark-1.3 model through api.meta.ai with Meta’s web search tool enabled. Each question ran twice on each system; agent runs used different accounts for the two runs. A language model extracted each answer’s written picks with brand and linked store under a fixed instruction, and product lines were rolled up to their brand.3 We report share of answers naming a brand, share of picks linked to each store, and the brand overlap between pairs of answers to the same question.4 The honest limits: five categories and 20 questions each is a first edition, so brand rankings below the top few are directional in four of the five categories. The Muse accounts used were research accounts with earlier shopping history, and because Muse remembers past conversations, part of the run-to-run variation may come from account history. Roughly one agent attempt in three failed before producing a final answer (the page did not load, the message was not delivered, or the agent had not finished) and was retried on another account; two questions needed three attempts.5 All agent runs took place on November 9 to 11, 2026, from US accounts. The Amazon explanation rests on press reports that Amazon blocks Muse’s browser; our data shows only that the agent never linked it. ##### Footnotes 1. Prompt text for both systems: “I’m looking for \[question\]. Find me the three options you’d recommend. For each one, tell me the store or business it’s from, the price, and why you picked it. Don’t ask me questions first: if a detail is missing, assume something reasonable and tell me what you assumed,” followed by a research-only instruction not to buy, add to cart, contact anyone, submit forms or create accounts. Categories: dog supplements, beauty (skin, makeup, hair), running shoes and gear, cookware and kitchen, protein and sports nutrition. 2. Spark setup: model muse-spark-1.3 via api.meta.ai/v1/responses with the web\_search tool, no system prompt, default settings. It typically made 6 to 15 searches per answer. The agent setup is the consumer Muse product at muse.ai in a signed-in US account, reached through residential mobile connections, at most two concurrent sessions per connection. 3. Extraction used DeepSeek V4.1 Flash with a fixed instruction to list each product the answer names as a written pick (brand, product, store) and separately any product shown only in the carousel. Carousel-only products are excluded from every figure here. Brand normalization merges product lines into brands (Cosequin, Dasuquin and Denamarin into Nutramax; Ascent Protein into Ascent) and ignores retailer names used as brands. 4. Overlap between two answers is the number of brands both recommend divided by the number either recommends, averaged over the 20 questions in a category and over every pair of runs (agent 1 with agent 2; Spark 1 with Spark 2; each agent run with each Spark run). Top-10 overlap counts brands that appear in both systems’ ten most-recommended lists for the category. 5. Of the agent attempts we logged, about a third ended without a final answer: Muse’s page did not load within 90 seconds, the message was never delivered because the chat was reconnecting, or Muse posted only a progress note. Unfinished attempts were retried on a different account up to three times, and all 200 question-runs eventually produced a final answer. One mobile connection accounted for most load failures. 6. Adoption figures are Meta’s, as reported in press coverage of Muse’s launch (September 2026) and first weeks. Descriptions of how Muse selects and ranks products draw on published analyses of its shopping instructions and on our own observation of its answers. \* Quotes attributed to the agent are taken verbatim from its answers in our runs, with screenshots retained for every answer. They reflect Muse as tested on November 9 to 11, 2026. Written by Enrich Labs, from 200 Muse agent answers and 200 Muse Spark answers collected in November 2026. The question set, per-category tables and extracted picks are available on request. Previously in this series: [Can AI agents buy from the Fortune 500?](/research/can-ai-agents-buy-from-the-fortune-500), [When marketing runs itself](/research/when-marketing-runs-itself) and [Project FSD: Can an AI agent run a Shopify store?](/research/project-fsd-can-ai-run-shopify-store) Want to know whether Muse’s shoppers are sent to your brand, and from which store? [Get on the Agent Conversion Optimization waitlist](#waitlist).