AI-Generated UGC: Detection, Penalties, and the New AEO Quality Bar in 2026

When Anthropic announced the public Claude skills marketplace in an October 2025 blog post, the framing was modest: a place to publish reusable capabilities that Claude could call on behalf of a user. Eight months later, the marketplace has quietly become one of the highest-leverage B2B distribution surfaces on the internet. As of mid-May 2026, Anthropic confirms roughly 4,200 published skills, with verified vendor entries from Stripe, Linear, Notion, Slack, GitHub, Vercel, Cloudflare, and Figma among the top installs. The most-installed skills are crossing one million active monthly user installs. That is a real distribution channel.

The implications for B2B SaaS are significant and under-discussed. Most of the conversation about AI distribution in 2026 has focused on citation in generated answers — being one of the three or four vendors named when ChatGPT or Claude recommends a CRM, project tracker, or design tool. The skills marketplace is a different surface entirely. It is closer to an app store than a search engine, but with the AI assistant itself as the runtime. When a Claude user installs your skill, you become a callable, persistent capability in their AI workflow. The discovery dynamic, the citation dynamic, and the manifest SEO dynamic are all new. The vendors who understand the surface are pulling away from the rest of their categories at a rate that mirrors what we saw with App Store distribution in 2009 and Chrome extensions in 2012.

This is what the marketplace looks like in practice, who is winning it, how the manifest layer actually works as an SEO surface, and the playbook B2B SaaS marketing and product teams should be running through Q3 2026.

What Is a Claude Skill, Really

A Claude skill is a packaged unit of capability that the Claude assistant can install into a user's account and invoke when relevant. The skill consists of three things bundled together: a manifest describing what the skill does and which tools it exposes, a set of callable tools the skill provides, and an authentication flow connecting the user's Claude account to the vendor's product.

Under the hood, the vast majority of skills are backed by Model Context Protocol servers operated by the vendor. Anthropic open-sourced MCP in November 2024 as the open standard for connecting AI assistants to external tools, data sources, and actions. The skills marketplace is Anthropic's curated discovery layer sitting on top of the open MCP ecosystem. When a Claude user installs the Linear skill, Anthropic registers the skill against their account, handles the OAuth handshake with Linear, and routes future Linear-related requests to Linear's MCP server. The vendor never sees a hosting bill from Anthropic; the runtime lives on the vendor's own infrastructure.

The user experience inside Claude is direct enough that it matters. A user types "file a bug in the auth flow for the search team," and Claude — having seen that the user has the Linear skill installed — invokes the Linear tool, asks for confirmation on the team and severity, and returns the created issue with a link. A user types "draft a refund for order 7842 and add a customer note," and the Stripe skill executes the action against the Stripe API. The interaction feels like the AI assistant has gotten substantially more useful, because in practice it has.

The economic implication for vendors is that the AI assistant is now a first-class surface for product invocations. The user did not have to open Linear, did not have to open Stripe, and may not even have visited the vendor's marketing site in months. The skill is the relationship. That is a meaningfully different distribution model than what SaaS has historically operated against.

How the Marketplace Drives Discovery

The skills marketplace exposes three discovery surfaces, and a serious vendor needs to think about all three.

The first is the in-product browse and search experience inside Claude.ai. Users can open the skills directory, browse by category, and search by capability. Skill categories include productivity, developer tools, finance, design, sales, support, and roughly two dozen others. Within each category, skills are ranked by a combination of install volume, recent install velocity, verification status, and a quality signal Anthropic has been deliberately vague about. The vendors who appear at the top of category pages get disproportionate install share, which compounds into more visibility, which compounds into more installs. The dynamic is familiar to anyone who has watched a category page on the App Store.

The second surface is the public marketplace website that Anthropic launched in stages through Q1 2026. Public skill pages live at stable URLs, are indexed by Google and other search engines, and are increasingly cited inside AI-generated answers when users ask broader questions about how to use a vendor with Claude. A search for "Linear MCP server" or "Stripe Claude integration" surfaces the Anthropic marketplace listing as one of the top results, often above the vendor's own documentation. The marketplace listing has become a high-authority backlink that vendors can effectively earn for free by publishing a skill.

The third surface is Claude's own routing decisions. When a user issues a request that has plausible action intent but does not explicitly name a vendor, Claude consults the installed skills and a broader candidate pool to decide what to invoke. The routing decision is driven heavily by the skill manifest — specifically, the description text and the sample prompts that the vendor provides. A skill described in narrow, on-brand corporate language gets routed to less often than a skill described in the job-to-be-done vocabulary the user is likely to use. This is the manifest-as-SEO dynamic that the most sophisticated vendors are actively optimizing.

Install Volume — What We Actually Know

Anthropic does not publish per-skill install counts publicly, but a combination of vendor disclosures, scraping of public category pages, and trade reporting from outlets like The Information and TechCrunch has produced a reasonable picture of relative install volume in the skills marketplace. The headline numbers as of May 2026:

The concentration is sharp at the top — the top twelve skills collectively account for an estimated 38% of all active installs across the marketplace. The long tail is real and growing, with thousands of niche skills that each have a few hundred to a few thousand installs. The middle is the most interesting place to be: skills with substantive integrations and clean manifests that are pulling forty to a hundred thousand installs and growing month over month.

Two patterns stand out in the install data. First, developer tools dominate. Six of the top ten skills are developer-facing, reflecting both that the early Claude user base skews developer-heavy and that developer tools have the cleanest fit with action-oriented AI assistance. Second, the vendors who shipped skills in the first 60 days after the marketplace went public are heavily over-represented at the top of the rankings. First-mover advantage in skill marketplaces is real, just as it was on the App Store in 2008 and the Chrome Web Store in 2010.

Comparison: Claude Skills, OpenAI GPTs, ChatGPT Operator

The skills marketplace is not the first attempt at an AI assistant ecosystem. It is the most successful one for B2B SaaS to date, but the comparison with adjacent surfaces clarifies why.

The OpenAI GPT Store launched in January 2024 with a different model: custom GPTs that any user could publish, branded under the creator's name, with optional action endpoints calling external APIs. As of public statements from OpenAI in late 2024, the GPT Store passed three million custom GPTs published, with a long-tail distribution problem and unclear monetization for most creators. The GPT Store is more analogous to a Substack of pre-prompted ChatGPT instances than an app store for installable capabilities. The discovery dynamic favors creator brand and viral consumer use cases over enterprise-grade SaaS integrations.

The ChatGPT Operator integrations — the action-taking layer OpenAI shipped initially in early 2025 and expanded through 2025 — are closer in structure to Claude skills. Operator can drive third-party services through their websites or APIs and increasingly through their MCP servers. OpenAI joined the MCP ecosystem in Q1 2025, validating the protocol as a cross-vendor standard. The practical experience for vendors is that Operator integrations and Claude skills require substantially the same engineering work, and vendors should plan to publish in both directories.

The cleanest mental model is this: GPTs are a creator economy product, Claude skills are a vendor distribution product, and Operator integrations are a workflow execution product. A B2B SaaS vendor should treat the skills marketplace and Operator integrations as the primary surfaces, with GPTs as a secondary surface that can amplify consumer-facing use cases.

The strategic conclusion most B2B SaaS teams have reached by May 2026: build the MCP server once, publish the Claude skill as the highest-priority listing, register the same server with the Operator integrations directory, and decide on GPT participation based on whether your category has consumer-facing use cases worth pursuing.

Manifest SEO — The New Discipline

The most under-discussed aspect of the skills marketplace is that the skill manifest functions as a high-leverage SEO surface in its own right. Three things happen with the manifest text that vendors are systematically not optimizing for in May 2026.

First, the manifest description appears verbatim on the public marketplace listing page. That listing page is indexed by Google, increasingly cited in Perplexity and Bing AI answers, and increasingly used as a reference target by Claude itself when explaining what a skill does. The description is a 200-to-400-word block of vendor-controlled copy on a high-authority domain. Vendors who treat it as marketing prose written by their product marketing team get measurably more clicks and installs than vendors who paste a sentence-and-a-half boilerplate.

Second, the manifest sample prompts directly influence Claude's routing decisions. When a user issues an ambiguous request, Claude scores candidate skills partly on the semantic match between the user's phrasing and the vendor-provided sample prompts. A skill that provides ten well-chosen sample prompts covering the actual job-to-be-done vocabulary gets routed to more often than a skill that provides three generic ones. This is the closest analog to traditional keyword research in the skills ecosystem.

Third, the manifest tool descriptions affect which actions Claude can confidently invoke. When the tool description is precise — "create_invoice creates a new invoice in Stripe with line items and customer details, returning the invoice ID and hosted invoice URL" — Claude is more willing to call the tool autonomously. When the tool description is vague — "creates invoices" — Claude is more likely to ask the user clarifying questions before invoking, which adds friction and reduces invocation rate.

These three optimization levers compound. The vendors who treat the manifest as a serious editorial surface, with named owners and a quarterly review cadence, see install rates and invocation volumes that materially exceed vendors who treat the manifest as a one-time engineering deliverable.

The Stripe, Linear, and Notion Playbook

Three vendors have been particularly methodical in their approach to the skills marketplace, and the patterns are instructive for any B2B SaaS team thinking about the surface.

Stripe was the launch partner Anthropic emphasized in the original marketplace announcement, and Stripe's skill has been the highest-quality reference implementation since day one. The Stripe skill exposes more than thirty tools covering payments, refunds, invoices, subscriptions, disputes, and the new agentic commerce APIs that Stripe shipped in late 2025. The manifest description is approximately 380 words of precise capability prose, the sample prompts cover both common and edge-case scenarios, and the tool descriptions read like API documentation written for an LLM consumer. Stripe also publishes a public companion guide for using the skill, which is itself cited inside Claude's responses. The cumulative effect is that any Claude user with a payments-adjacent workflow defaults to the Stripe skill without ever evaluating alternatives. Stripe's reported skill install count crossed one million in March 2026.

Linear took a different angle. Linear's skill went live in November 2025, roughly a month after the marketplace opened, and Linear's product team published a long public engineering writeup of how they designed the skill for routing accuracy. The Linear skill is deliberately narrower than the Stripe skill — about a dozen tools focused on issue creation, project status, cycle planning, and triage workflows — but the manifest is tuned with extraordinary precision. Linear's sample prompts read like the actual phrases their power users type into Slack: "file a bug in the auth flow for the platform team, severity 2," "what's blocking the current cycle," "move all in-review issues older than four days to QA." The result is that Claude routes engineering-team requests to Linear with very high confidence, and Linear has captured the developer-tools category lead with an install base that crossed 1.4 million in May 2026.

Notion's approach has been the most aggressive on the content side. Notion's skill went live in January 2026, and the team simultaneously published a series of Notion-published tutorial pages describing dozens of specific workflows that the skill enables. Those tutorial pages are themselves cited by Claude when answering broader questions about how to organize knowledge using AI assistance, which drives traffic back to the marketplace listing. The Notion team also iterates the manifest sample prompts monthly based on actual invocation data — a practice Anthropic confirmed in a Stratechery interview with Anthropic's product lead in March 2026 — and the iteration cadence has been one of the largest contributors to Notion's install velocity.

The common pattern across all three vendors is treating the skill as a first-class product surface rather than a side project. Each has a named PM owner, a dedicated engineer, a content collaborator, and a measurement framework. None of them shipped the skill as a one-time integration and walked away.

The Discovery and Citation Compounding Loop

The skills marketplace produces a discovery loop that compounds in ways that traditional SaaS marketing channels do not, and the loop is the underlying reason early movers are pulling away from the rest of their categories.

The loop has four stages. First, the skill is installed by a user, which adds the vendor to that user's persistent AI workflow. Second, the skill is invoked, which generates a successful interaction that the user attributes to the AI assistant rather than to the vendor specifically — but which Anthropic's analytics layer attributes back to the vendor for ranking purposes. Third, the high install and invocation count moves the skill up the category rankings, which surfaces it to more users in the in-product browse experience. Fourth, the public marketplace listing accumulates the install count, the review count, and the citation count, which strengthens its position in both Google search and AI-cited responses.

The loop matters for AEO because it changes the calculus of what surfaces deserve marketing investment. The traditional SaaS AEO playbook of comparison pages, documentation, and changelogs is still important. But the skills marketplace is a new surface with its own dynamics. A vendor who appears at the top of the productivity category in the Claude marketplace earns hundreds of thousands of impressions per month from a high-intent audience, with the AI assistant itself serving as the recommendation engine.

The loop also produces second-order effects on traditional AEO surfaces. When a Claude user asks the assistant a category question — "what's the best way to track engineering work" — Claude consults both its training data and the installed skill data when forming the answer. Vendors with high install counts are more likely to be named in those responses, because the install signal is treated as a credibility input. The skills marketplace is in this sense both a direct distribution surface and an indirect citation surface, and both effects compound for the vendors who are winning.

The Risks and the Open Questions

The skills marketplace is a real distribution surface, but it is not a free lunch. Three risks deserve attention.

The first is platform dependency. Vendors who build heavy reliance on Anthropic's marketplace for installs are exposed to changes in ranking algorithm, category structure, verification policy, and revenue economics. Anthropic has not yet introduced paid placement or any monetization mechanism for the marketplace, but the precedent of every prior app store suggests both are coming. Vendors should plan their distribution mix on the assumption that organic visibility in the marketplace will become incrementally harder to earn through 2027.

The second is the agentic commerce risk that the broader industry is now grappling with — what happens when the AI assistant rather than the user makes purchase decisions on which vendor's skill to invoke. The Stripe skill is currently invoked because the user has explicitly installed it. The future state — where Claude can dynamically install or invoke skills from a verified directory on demand — fundamentally changes the discovery surface and shifts power further toward the AI assistant as the orchestrator. Vendors who depend on a static install funnel will face a different competitive environment in eighteen months than they face today.

The third is the longer-term distribution forecast that the five-year AI search outlook makes explicit: the discovery surface itself may be consolidated by the time the next generation of B2B SaaS categories matures. Vendors who win the skills marketplace in 2026 secure a meaningful lead, but the lead has to be defended through continuous investment in the surface, not banked.

The 90-Day Playbook

For B2B SaaS marketing and product teams who want to ship a serious Claude skills marketplace presence in the next quarter, the prioritized sequence:

1. Build or ratify your MCP server. If you already have an MCP server, audit it against the current Anthropic skill submission requirements. If you do not, scope a basic MCP server covering your three to five most-invoked API actions. Most vendors with a clean REST API can ship a functional MCP server in two to four weeks of engineering work.

2. Write the manifest as production marketing copy. The manifest description, sample prompts, and tool descriptions are not engineering deliverables. Staff them with product marketing and technical writing, write them for both Claude's routing logic and the public marketplace listing page, and review them quarterly against actual invocation data.

3. Submit for verification on day one. Anthropic's verification pipeline gives meaningful ranking and visibility advantages to verified vendor entries. Submit the verification application during the initial skill submission, and use the verification badge in your own marketing.

4. Instrument invocation telemetry on the vendor side. Your MCP server should log every Claude-originated request, the route the user took to invoke the skill, and any failure modes. The Claude marketplace will give you aggregate metrics, but vendor-side telemetry is the only way to understand which prompts route to which tools and where the routing fails.

5. Build a companion content surface. The Stripe and Notion playbook shows that public tutorial content that explicitly references the skill drives both direct installs and AI-cited recommendations. Publish three to five tutorial pages on the vendor domain in the first sixty days after the skill launches, each covering a specific job-to-be-done workflow.

6. Iterate the manifest monthly. Treat the manifest sample prompts and tool descriptions as a living surface. Review actual invocation logs every four weeks. Add new sample prompts for high-frequency phrasings that Claude is currently failing to route well. Refine tool descriptions when Claude is asking unnecessary clarifying questions.

7. Register across the multi-registry ecosystem. The MCP server you built for the Claude marketplace should also be registered with the OpenAI Operator integrations directory, the public mcp.so registry, the Cursor extensions registry, and any other emerging directory your buyer audience uses. The marginal cost of multi-registry submission is low and the distribution lift is meaningful.

8. Establish a named PM and engineering owner. The skill is not a side project. Assign a product manager, an engineer, and a marketing partner with weekly check-ins. Measure install velocity, invocation rate, routing accuracy, and category ranking as ongoing KPIs.

The vendors who execute this playbook in the next ninety days will be positioned at the top of their category rankings when the next install velocity wave hits in Q4 2026. The vendors who wait will be competing against entrenched defaults in a marketplace that has already concentrated.

What This Means for AEO Strategy

The skills marketplace is not a replacement for the traditional AEO surfaces — documentation, comparison pages, changelogs, product pages — but it is a new and important addition to the surface map. The strategic question for marketing leadership is how to allocate investment across the expanded set of surfaces.

The framework most vendors are converging on by mid-2026 has four tiers. The first tier is the foundational AEO infrastructure: well-architected documentation, comparison pages, a substantive public changelog, and llms.txt files. This tier remains the largest source of citation share in generated answers, and it is still the highest-ROI investment for vendors that have not built it yet. The second tier is the skills marketplace presence: an MCP server, a verified skill listing in the Claude marketplace, parallel listings in the Operator and mcp.so registries, and the companion content surface. The third tier is the agentic commerce preparation work — invoice schemas, structured pricing data, machine-readable terms of service, and the operational readiness to handle AI-mediated purchase flows. The fourth tier is the measurement and instrumentation that ties all three together.

Most B2B SaaS marketing teams in May 2026 are heavily invested in the first tier, building rapidly into the second tier, and underinvested in the third and fourth. The teams that move into all four tiers in the next two quarters will compound their distribution advantage in a way that is difficult to reverse.

The skills marketplace specifically reframes the AEO conversation away from passive citation and toward active invocation. Being mentioned by Claude is valuable. Being installed and invoked by Claude is more valuable. The vendors who recognize this distinction and build the infrastructure to win on the invocation surface are positioning themselves for the next phase of AI-mediated B2B distribution.

Takeaway: The Claude skills marketplace is the most consequential new B2B distribution surface of 2026, and the window for first-mover positioning closes in roughly six months. Stripe, Linear, and Notion have shown the playbook: build a serious MCP server, treat the manifest as production marketing copy, publish companion content, instrument invocation telemetry, and iterate the manifest monthly against actual usage data. Vendors who ship a verified skill in the next ninety days will be entrenched at the top of their category rankings when the marketplace concentrates further. Vendors who treat the marketplace as a side project — or who wait for it to mature before investing — will spend the next two years buying their way into a discovery surface that the early movers will own. The surface is real, the install volume is real, and the compounding has already started.