Key Takeaways
- Amazon’s Project Starfish scrapes over 200,000 brand websites and uses AI to rewrite titles, bullets, images, and descriptions across millions of listings without seller approval.
- Amazon’s own AI has generated listing content that violated Amazon’s own policies, resulting in listing removals that sellers could not reverse through standard support channels.
- Brand Catalog Lock is the single most effective defense available: it prevents unauthorized edits to titles, images, bullets, and descriptions, and most sellers have not activated it.
- Incomplete listings are the primary vulnerability: Amazon’s AI was designed to fill gaps, and empty fields are treated as an invitation to overwrite.
- Sellers who maintain complete, consistent listing data across Seller Central, flat files, and their brand website are far less likely to be affected by AI-driven rewrites.
- Amazon’s internal estimate projects Project Starfish will generate an additional $7.5 billion in gross merchandise sales, signaling this program is a permanent fixture, not a test.
General Summary
Amazon is running an AI initiative called Project Starfish that scrapes brand websites, third-party retailer pages, and product databases to rewrite listing content at scale. Reported by Business Insider in July 2025, the program has already affected over 900,000 sellers. Amazon’s stated goal is to build the most comprehensive product knowledge base in the world. The commercial motivation is explicit: internal documents estimate Project Starfish will contribute $7.5 billion in additional gross merchandise sales. Amazon sees this as a catalog quality initiative. Sellers are experiencing it as an unauthorized rewrite of their brand content, with limited recourse and inconsistent Seller Support responses.
Extractive Summary
Project Starfish is Amazon’s internal AI initiative to make Amazon the best source of product information for all products worldwide. Real sellers are experiencing listing removals, image swaps, variation splits, and category reclassifications triggered by AI-generated content they never approved. Brand Catalog Lock is the most important protective measure available to brand-registered sellers, and most sellers have not activated it. Flat files give sellers stronger catalog authority and serve as documentation of intended content in the event of a dispute. Amazon sometimes sends advance notification emails about planned AI changes, and sellers who miss the two-week response window see changes go live automatically. The sellers hit hardest are those with incomplete listings, because Starfish was built to fill gaps.
Abstractive Summary
Project Starfish marks a structural shift in the relationship between Amazon and the sellers who use its marketplace. For years, Brand Registry gave sellers the expectation of control over their own listing content. That assumption no longer holds. Amazon is now an active editor of the catalog, not a passive host. The sellers most at risk are not those who have done something wrong. They are those who have done something incomplete. Amazon’s AI interprets empty fields and inconsistent data as problems to solve. Sellers who treat their Amazon listing as a standalone asset, disconnected from their website and catalog data, will find that disconnection used against them. The practical response is not to fight Amazon’s AI but to give it nothing to correct.
What Is Project Starfish and Why Should Sellers Care?
Project Starfish is Amazon’s internal AI program to collect, synthesize, and rewrite product listing content across its entire catalog, without requiring seller approval. Reported by Business Insider in July 2025, the initiative has already touched over 900,000 seller accounts. Amazon’s stated objective is to become the best source of product information for all products worldwide. The commercial objective is an estimated $7.5 billion in additional gross merchandise sales.
The program operates in four stages. First, Amazon’s AI scrapes brand websites, third-party retailer pages, manufacturer databases, product reviews, and Q&A sections. Second, it synthesizes the collected data, prioritizing Brand Registry content when conflicts arise. Third, it uses large language models to rewrite titles, bullets, descriptions, and generate images. Fourth, it deploys those changes to live listings, sometimes with a notification email and sometimes without.
The adoption numbers Amazon cites internally show why they are committed to this approach. Ninety percent of sellers accept AI-suggested changes. Listing quality scores increased 40% where the program has been applied. Rufus, Amazon’s AI shopping assistant, served 250 million shoppers in 2024, up 140% year over year. Shoppers who interact with Rufus are 60% more likely to complete a purchase. Amazon is optimizing for conversion, and it believes better listing content drives conversion.
The tension this creates is direct. Amazon’s AI makes decisions about listing content based on what it calculates will convert best. Sellers may disagree with those decisions. Under Project Starfish, disagreement is not enough to prevent a change from going live.
What Has Project Starfish Actually Done to Seller Listings?
Project Starfish has produced documented harm across four categories of listing damage: content violations, catalog lockouts, image replacement, and structural changes to variation families.
How Has AI-Generated Content Triggered Policy Violations?
Amazon’s AI has generated listing content that violated Amazon’s own listing policies. One supplement seller used Amazon’s Generate Listing Content tool when creating a new listing. The AI produced titles, bullets, and descriptions that included prohibited health claims. Three weeks later, Amazon’s compliance bots flagged the AI-generated content. The listing was removed within 48 hours. The seller had already manufactured the product and shipped inventory to FBA.
When the seller attempted to relist with corrected content, Amazon’s catalog retained the erroneous AI data. Standard overwrite attempts failed. Seller Support sent scripted responses. The bad data persisted even after the original listing was deleted. Amazon’s AI generated content that broke Amazon’s rules, and Amazon held the seller responsible for the violation.
How Are Brand Owners Being Locked Out of Their Own Listings?
Brand-registered sellers, the actual manufacturers and brand owners, are being prevented from editing their own listing content after AI rewrites. One seller reported that Amazon changed their title to content that violated Amazon’s own style guide. When the seller tried to correct it, Amazon’s system rejected the edit, citing the current title as non-compliant. The non-compliant title was the one Amazon’s AI had written.
Seller Support’s only suggested resolution was to delete the listing and relist under the same SKU with the corrected title. For a title change. On a listing the seller created and owned.
What Kinds of Changes Are Being Made Without Notification?
Image swaps, variation splits, and category reclassifications are occurring without seller awareness. One seller found that Amazon had replaced their listing images with alternatives, with no path to restore the originals as the primary image. The replacement images were not higher quality. The seller could not identify where Amazon had sourced them.
Another seller woke up to find 13 of 16 child ASINs had been broken into separate standalone listings, each with its own review count. No action had triggered the split. No notification had been sent. A sports supplement was reclassified into the baby products category. A second product was flagged as adult content, eliminating organic visibility overnight.
The consistent pattern across these cases is that Amazon’s AI treats its own output as authoritative. Once a change is made, the system is resistant to reversal. Seller Support does not have override access to automated catalog decisions. The sellers who recover fastest are those who can demonstrate, through documentation, that their original content was correct and complete.
Why Are Some Listings More Vulnerable Than Others?
Listings with incomplete data are the primary target of Project Starfish, because the program was designed to fill gaps, not to replace complete content.
Amazon’s AI works through a specific decision process. When it encounters conflicting information from different sources, it prioritizes Brand Registry content. When it encounters empty or inconsistent fields, it does not defer to the seller. It fills the gap with whatever data it scraped from external sources and generates content it calculates will perform better.
Empty backend fields are the most direct vulnerability. Every unoccupied attribute field is, from Starfish’s perspective, a problem waiting to be solved. Incomplete listings also create a secondary risk: when Amazon’s AI scrapes a brand’s own website and finds information that conflicts with the partial Amazon listing, it may treat the website data as supplementary evidence for a rewrite, even when the seller would prefer the existing Amazon content.
Sellers with fully populated listings, consistent data across every field, and matching content between their Amazon catalog and their brand website give Amazon’s AI the least possible justification to intervene. The algorithm is looking for gaps to fill. A complete listing has none.
There is also a compounding effect that sellers in regulated categories face at a higher rate. Health and wellness products, supplements, and beauty products carry compliance requirements that vary by country, platform, and regulatory body. Amazon’s AI does not have visibility into those requirements. When it synthesizes content from a brand website, a competitor page, and a manufacturer database, it produces content that may accurately describe the product but use language that triggers Amazon’s own compliance systems. The seller faces a violation for content they never wrote.
Category misclassification creates a separate vulnerability. Amazon’s AI infers category based on the data it collects. A product with attributes that overlap two categories may be moved based on what the AI determines is the better fit for conversion. That reclassification can change search visibility, advertising eligibility, and fee structure simultaneously, without the seller being notified in advance.
What Is Brand Catalog Lock and How Do You Activate It?
Brand Catalog Lock is a feature Amazon introduced as part of its 2025 Brand Registry updates that prevents unauthorized users, including in many cases Amazon’s own AI, from modifying titles, images, bullets, and descriptions on brand-registered listings.
Activating Brand Catalog Lock requires three things: a registered or pending trademark, active enrollment in Amazon Brand Registry, and correctly assigned contributor roles for the brand. The feature is requested through Brand Registry’s content protection tools.
Once active, Brand Catalog Lock restricts edit permissions to approved brand roles only. Changes from outside those roles, whether from third-party sellers, catalog automation, or AI systems, are blocked. This does not make a listing immune to every form of AI interference, but it eliminates the most common vector through which Project Starfish makes unauthorized changes.
Most sellers with Brand Registry enrollment have not activated Brand Catalog Lock. It is not enabled by default. It requires a deliberate request. For sellers with more than a handful of SKUs, this should be treated as a priority action, not a future consideration.
There is an important distinction between what Brand Catalog Lock protects and what it does not. It prevents unauthorized edits to the specific fields it covers: titles, images, bullets, and descriptions. It does not prevent Amazon from reclassifying a product’s category based on attribute data, and it does not prevent AI-generated review summaries from affecting conversion. Sellers should treat Brand Catalog Lock as a necessary first layer of protection, not a complete solution.
Sellers who have not yet enrolled in Brand Registry face a more fundamental exposure. Without Brand Registry, there is no access to Brand Catalog Lock and significantly less recourse when AI-driven changes occur. For brand owners who have delayed Brand Registry enrollment, Project Starfish makes that enrollment more urgent. The program prioritizes Brand Registry content as its authoritative source. Sellers without it have no authoritative source in Amazon’s system at all.
What Is the Eight-Step Defensive Playbook Against Project Starfish?
Eight specific actions reduce exposure to Project Starfish and improve a seller’s position when AI-driven changes do occur.
Defense 1: Activate Brand Catalog Lock
This is the highest-priority action. Brand Catalog Lock prevents unauthorized edits to core listing fields. Without it, every listing is exposed to AI rewrites regardless of how complete the underlying data is. Sellers with Brand Registry should request activation through content protection tools before taking any other step.
Defense 2: Fill Every Backend Field
Every empty field in Seller Central is an invitation for Starfish to write something the seller did not approve. Amazon’s AI prioritizes Brand Registry content as the authoritative source, but only when that content exists. Backend attributes, search terms, subject matter, and all product-specific fields should be populated completely and accurately. Incomplete listings get rewritten. Complete listings give the AI nothing to fix.
Defense 3: Use Flat Files for Uploads
Flat files expose more fields than the Seller Central user interface and give sellers stronger catalog authority. They also create a documented record of intended content. If Amazon’s AI overwrites a listing and the seller needs to dispute the change, a flat file upload provides evidence of the original intended state. Content entered through the UI and later overwritten has no equivalent record. Sellers managing high-value SKUs should upload and maintain content through flat files as a standard practice.
Defense 4: Monitor Your Top Listings Daily
Monitoring catches AI-driven changes before they compound. Sellers should track title changes, bullet modifications, image swaps, category reassignments, and variation splits across at least the top 20% of their SKUs by revenue. Tools that support this include SentryKit for variation monitoring, Bindwise, and SellerPulse. A daily manual check of top sellers works for smaller catalogs. The goal is to identify changes within 24 hours, before the AI system treats its output as an established state that is harder to reverse.
Defense 5: Audit Your Brand Website for Consistency
Project Starfish scrapes brand websites. Outdated product pages, inconsistent specifications, or old descriptions on a brand’s own site become source material for AI rewrites on Amazon. Every product page on the brand website should match the corresponding Amazon listing exactly: same names, same specs, same feature descriptions. Outdated content should be removed. When Amazon’s AI finds a conflict between a website and an Amazon listing, it uses the scrape data as justification for a change. A clean, consistent website removes that justification.
Defense 6: Respond to Amazon Notification Emails Immediately
Amazon sends advance notification emails about planned AI changes to listing content. The response window is typically two weeks. If the seller reviews and rejects the proposed change within that window, the change does not go live. If the seller does not respond, the change is applied automatically. These emails arrive formatted like routine compliance communications and are easy to miss. Setting up email filters to flag anything from Amazon related to listing updates and assigning daily review responsibility are necessary steps for any seller with more than a few active SKUs.
Defense 7: Never Publish AI-Generated Listing Content Without Full Review
Amazon’s Generate Listing Content tool produces content quickly, but it does not know a seller’s compliance requirements, category-specific restrictions, or brand positioning. Sellers in regulated categories including supplements, health products, and beauty should write all listing content offline before creating a new listing, then upload via flat file. For sellers who do use Amazon’s generation tool, every line requires manual review before publishing. The supplement case documented above shows the consequences of publishing AI-generated content that contains claims Amazon’s compliance systems will later flag.
Defense 8: Know Your Escalation Paths Before You Need Them
Standard Seller Central editing fails in some Project Starfish cases, and sellers who do not know the escalation paths lose time. Specific paths that have worked for affected sellers include submitting identical content changes across multiple SKU offers simultaneously, using the Investigate Other Product Issues flow with My Issue Is Not Listed selected (this forces human review), contacting advertising representatives to escalate internally, opening a Brand Registry support case requesting Brand Catalog Lock specifically, and updating the brand website to match the desired Amazon content, then referencing the website in the support case as evidence of brand intent.
What Is the Core Principle Behind All Eight Defenses?
The core principle is that Project Starfish fills gaps, so the defense is to eliminate gaps. Amazon’s AI looks at a listing and identifies what is missing, inconsistent, or potentially improvable. A listing with complete data, consistent attributes, matching content across Seller Central and the brand website, and an active Brand Catalog Lock gives the system nothing to act on.
This principle holds across all eight defenses. Monitoring detects gaps that appear after the fact. Flat files document what the intended state was. Website audits close the gap between brand content and Amazon content. Email monitoring closes the gap in response time. Escalation paths close the gap when standard tools fail.
Sellers who approach this defensively, treating listing completeness as a catalog hygiene standard rather than a one-time setup task, have the strongest position against AI-driven rewrites. The sellers most damaged by Project Starfish are those who set up their listings once and assumed that was enough. Under the current program, it is not.
What Does Project Starfish Signal About Amazon’s Direction?
Project Starfish signals that Amazon intends to become an active editor of its catalog, not a passive host for seller-generated content. The internal estimate of $7.5 billion in additional gross merchandise sales is not a projection Amazon will abandon. The 90% acceptance rate among sellers who receive AI-suggested changes tells Amazon its approach is working.
The sellers who will be least affected by this shift are those who understand what Amazon’s AI is trying to optimize and align their listings with that objective. Amazon is optimizing for conversion. Listings that convert well, that answer the questions Rufus is likely to surface, that have complete attribute data and consistent content across all sources, are listings Amazon’s AI has less reason to touch.
The sellers who will be most affected are those who treat their Amazon listing as a static asset. The catalog is now dynamic. Amazon’s AI is updating it continuously based on data from sources the seller may not even be aware of. The practical response is not resistance to that dynamic. It is staying close enough to the listing to catch changes early and maintain enough data completeness that the AI defaults to leaving the content alone.
Amazon is building the largest product knowledge base in the world. Sellers can position their listings as reliable, complete contributions to that knowledge base, or they can position them as gaps waiting to be filled. The sellers who treat listing quality as a continuous practice, not a launch task, are the ones who will keep control of how their products are represented.
The sellers who responded most effectively to prior Amazon algorithm changes, the A9 shifts, the introduction of Rufus, the rollout of AI review summaries, were those who adjusted their operational habits rather than waiting for the impact to be undeniable. Project Starfish is in the same category. It is not a temporary experiment. The $7.5 billion commercial case guarantees Amazon will expand it. The question for every seller is whether their catalog is ready to withstand continuous AI review, or whether it is relying on a version of control that no longer exists.
One data point that warrants attention: 250 million shoppers used Rufus in 2024, and that number is rising. Rufus draws on listing content to answer shopper questions in real time. Listing content that Amazon’s AI has rewritten is the content Rufus cites. Sellers who do not control what is on their listing do not control what Rufus tells shoppers about their products. That consequence extends well beyond the listing page itself.
