启动 HN：Channel3（YC S25）——一个涵盖互联网上所有产品的数据库
嗨，HN——我们是乔治和亚历克斯，正在构建 Channel3（<a href="https://trychannel3.com">https://trychannel3.com</a>），这是一个涵盖互联网上每个产品的数据库，可以通过文本/图像进行搜索，并内置了联盟营销功能。这里有一个演示：<a href="https://www.youtube.com/watch?v=Mx8FyP7KvJg" rel="nofollow">https://www.youtube.com/watch?v=Mx8FyP7KvJg</a>。
寻找优质产品数据出乎意料地困难。如果你希望你的软件能够推荐产品并深度链接到商家，你会很快发现你所需的数据——干净的标题、标准化的属性、去重的列表、当前价格和库存、变体选项、图片以及品牌信息——不仅杂乱无章，而且分散在众多零售商那里，且常常被复杂的机器人检测系统所屏蔽。
在构建一个能够推荐相关用品的AI教师时，我们遇到了这个问题。我们向Exa请求产品，但得到的却是文章，而不是结构化数据。Tavily和Bing（自2025年8月13日起已停用）也是如此。我们还被联盟计划拒绝，建议我们先积累上千个TikTok粉丝。Channel3就是我们希望拥有的API。
产品详情页面（PDPs）通常会展示主要商品以及推荐产品。我们使用计算机视觉技术来识别主要产品并找到其属性，如标题和价格。我们将相同的逻辑应用于该域名上的其他PDPs。
产品通常在多个零售商处销售，且没有保证它们的标签一致。我们通过使用大型语言模型（LLMs）和多模态嵌入将网络上的产品整合为一个规范化的集合，以真正理解每个产品。
为了将所有内容规范化为一个尽量简洁且可扩展的模式，我们必须持有明确的观点。一个50美元的10寸平底锅和一个60美元的12寸平底锅是同一产品吗？可能不是，但T恤的S/M/L变体是。我们的目标是，任何你特定搜索的产品都被视为独立的产品。
我们处理大量数据。我们很快就用完了Cloudflare Vectorize索引的空间，转而使用全新的AWS S3 Vectors平台，并与OpenSearch同步，以实现更快的响应时间和更动态的过滤。我们经常遇到速率限制，因此我们将工作分散到多个云服务提供商和AI模型上。
你可以搜索诸如“户外烧烤，低于1000美元”或“防汗无线跑步耳机”或“来自Paige的女性牛仔裤，样式类似于[<a href="https://www.gap.com/webcontent/0020/666/799/cn20666799.jpg" rel="nofollow">https://www.gap.com/webcontent/0020/666/799/cn20666799.jpg</a>]”这样的内容。产品以JSON格式返回，包含标题、品牌、图片、价格、规格等信息。
开发者可以从他们推动的销售中获得佣金（平均5%）。Channel3会抽取一部分。我们希望你从Channel3获得的收入远超你在其上的支出。你赢了，我们也赢。
我们提供API、SDK（Typescript和Python）和MCP。我们提供1000次免费搜索，之后每1000次搜索收费7美元。你可以在我们的仪表板上查看每个品牌的预期佣金。
到目前为止，产品仅限于美国（抱歉！我们会扩展）。我们已经上线了数百万个产品和数百名开发者。
要开始使用，请在<a href="https://trychannel3.com">https://trychannel3.com</a>上注册一个免费账户，然后选择你想要销售的品牌（从50000多个品牌中选择或请求你自己的品牌），生成API密钥，开始销售和赚钱。
我们非常希望得到这个社区的反馈。如果你之前构建过产品搜索，我们在模式中遗漏了什么？如果你尝试在应用中加入商业功能，遇到了什么阻碍？如果你尝试自己构建这个，你学到了什么？有没有你希望存在的端点（例如，价格提醒、补货Webhook、产品数据源）？对于任何建议，我们都非常欢迎。
我们会在这个讨论线程中全天回答问题，分享更多技术细节，并倾听任何能让这个工具对你更有用的建议。请随意评论！
查看原文
Hi HN — we’re George and Alex, building Channel3 (<a href="https://trychannel3.com/">https://trychannel3.com/</a>), a database of every product on the internet, searchable via text/image, and with built-in affiliate monetization. Here’s a demo: <a href="https://www.youtube.com/watch?v=Mx8FyP7KvJg" rel="nofollow">https://www.youtube.com/watch?v=Mx8FyP7KvJg</a>.<p>It’s surprisingly hard to find good product data. If you want your software to recommend products and deep-link to merchants, you’ll quickly discover that the data you need—clean titles, normalized attributes, deduped listings, current prices and inventory, variant options, images, and brand info—is not just messy; it’s also spread across a long, long tail of retailers, and often lives behind advanced bot-detection systems.<p>We ran into this problem while building an AI teacher that could recommend relevant supplies. We asked Exa for products, but got back articles, not structured data. Same for Tavily and Bing (deprecated as of 8/13/25). And we got rejected from affiliate programs, who suggested we come back with 1000s of TikTok followers. Channel3 is the API we wished we had.<p>Product detail pages (PDPs) usually present the main item alongside recommendations. We use computer vision to isolate the main product and find its attributes, like title and price. We apply the same logic to the rest of the PDPs on the domain.<p>Products are often sold across multiple retailers, with no guarantee they’ll be labeled consistently. We collapse products across the web into a canonicalized set by using LLMs and multimodal embeddings to actually understand each product.<p>To normalize everything into a schema that tries to be both minimal and extensible, we have to be opinionated. Are a $50 10” and $60 12” skillet the same product? Probably not, but the S/M/L variants of a T-shirt are. Our goal is that any product you’d search for specifically is treated as its own product.<p>We process a massive amount of data. We quickly ran out of room on our Cloudflare Vectorize indices and moved to the brand-new AWS S3 Vectors platform, syncing to OpenSearch for faster response times and more dynamic filtering. We hit rate limits constantly, so we spread our work over multiple cloud providers and AI models.<p>You can search things like “outdoor grill, less than $1000”, or “sweat-resistant, wireless running earbuds”, or
"women's jeans from Paige that look like [<a href="https://www.gap.com/webcontent/0020/666/799/cn20666799.jpg" rel="nofollow">https://www.gap.com/webcontent/0020/666/799/cn20666799.jpg</a>]”. Products come back as JSON with title, brand, images, price, specs, etc.<p>Developers earn commission on sales they drive (averaging 5%). Channel3 takes a cut. We want you to earn way more money from Channel3 than you spend on it. We win when you win.<p>We provide an API, SDK (Typescript and Python), and MCP. We offer 1000 free searches, and charge $7/1000 searches after that. You can view expected commissions per-brand on our dashboard.<p>So far, products are US-only (sorry! we will expand). We’re live with millions of products and hundreds of developers.<p>To get started, make a free account at <a href="https://trychannel3.com">https://trychannel3.com</a>, then select which brands you’d like to sell (choose from 50k+ or request your own), generate an API key, and start selling and earning.<p>We’d really appreciate feedback from this community. If you’ve built product search before, what did we miss in the schema? If you’ve tried to add commerce to an app, what blocked you? If you tried to build this yourself, what did you learn? Are there endpoints you wish existed (e.g. price alerts, back-in-stock webhooks, product feed)? For any suggestions, we’re all ears.<p>We’ll be in the thread all day to answer questions, share more technical detail, and hear whatever would make this most useful to you. Comment away!