Channel3 is a structured product database purpose-built for commerce — 100M+ products across 25,000+ retailers, normalized schema, real-time price and availability, and monetization built in.

If your agent needs to find and cite information from the web, Exa is a strong choice. If it needs to find, compare, and transact on real products, you need product infrastructure — not a general search index.

Different tools, different jobs

Exa’s pitch is “search built for AI” — an API that returns semantically relevant web content instead of the ten blue links model of traditional search. That makes it a strong fit for research tools, coding agents pulling documentation, and RAG systems that need to ground an answer in real web content. It’s general-purpose by design: the same API that helps an agent research a topic is the one returning results for anything else you throw at it, including product queries.

That generality is also the limitation for shopping specifically. A web search result is not a product record. It doesn’t come with normalized price, live inventory, variant options (size, color, SKU), a clean image, or a merchant relationship that supports a transaction. Returning a search result that mentions a couch isn’t the same as returning structured data that lets an agent actually compare couches or complete a purchase.

Channel3 is built for exactly that gap. Every product is a structured record — not a page to parse — with the fields a shopping agent actually needs: price, availability, variants, images, merchant, and commission data where relevant. It’s commerce infrastructure, not a general index that happens to include shopping content.

Channel3 vs. Exa

Primary use case — Exa: General web search for AI agents, RAG, research. Channel3: Structured product data for shopping and commerce agents.

Data format — Exa: Semantically ranked web content. Channel3: Normalized product schema (price, availability, variants, images).

Shopping-specific fields — Exa: No. Channel3: Yes — built specifically for commerce use cases.

Monetization / transaction support — Exa: Not built in. Channel3: Built in.

Best fit — Exa: Research, documentation lookup, general agent grounding. Channel3: Product discovery, comparison, and shopping agent experiences.

FAQ

What is Exa used for?
Exa is a search API designed for AI agents and LLMs, providing semantic web search for use cases like research assistants, retrieval-augmented generation (RAG), and general-purpose agent tooling.

Is Exa a good fit for shopping or ecommerce AI agents?
Exa is general-purpose web search, not a product-specific data layer. It can surface web pages that mention products, but it doesn’t provide the structured, normalized product data — price, availability, variants — that a shopping agent needs to compare or transact.

What’s an alternative to Exa specifically for product and shopping data?
Channel3 is a structured product database purpose-built for commerce: 100M+ products across 25,000+ retailers, with the normalized schema and monetization infrastructure that general-purpose search APIs like Exa aren’t built to provide.

Google’s product data is the moat, not the product

Google built its shopping index to keep people inside the Google ecosystem - Google Search, Google Shopping, and now Gemini. Every incentive points the same direction: keep the data closed, keep the traffic internal, keep advertisers paying for placement inside Google’s own surfaces.

That’s a completely rational business decision for Google. It also means the catalog is unavailable to everyone building outside it. There’s no “Google Shopping API” you can call to get structured product data for your own product. What exists, the Content API for Shopping and Merchant Center feeds, moves data one way: into Google.

If you’re building an AI shopping agent, a comparison tool, a browser assistant, or anything that needs to reason over real product data, Google simply isn’t an option.

What Channel3 does differently

Channel3 exists for the opposite reason Google’s catalog is closed: we’re completely independent infrastructure built to power a long-tail of AI shopping experiences.

We don’t compete with you. Channel3 sits beneath your shopping experience — think of it as the Stripe of agentic commerce. You maintain total control over the user experience: your editorial, your rules, your UX, your design. We just supply the product data underneath it.

• 100M+ products, 25,000+ retailers, structured and normalized - not raw HTML, not inconsistent feeds.

• Transparency. This industry runs on opacity - black-box ranking, unclear data provenance, no visibility into what's actually happening on the backend. We show you everything: the data, the analytics, how it's sourced and structured. No retailer affiliation, no marketplace of our own, no horse in which product wins.

• Built for machines first. Clean schema, real-time price and availability, images, variants - designed to be consumed by an AI agent or app, not rendered as a search results page.

• Open access. Any developer, any agent platform, any shopping experience can query it. You don’t need to be Google-sized to get real product data.

• Monetization built in. We offer affiliate infrastructure so you can actually make money on the shopping experiences you build - not just serve data.

Google Shopping vs. Channel3

Public API for the full catalog — Google: No, feeds go into Google, not out. Channel3: Yes, structured API and MCP access.

Neutrality — Google: Ranks and surfaces based on Google’s own commercial priorities. Channel3: Independent, no platform bias.

Built for — Google: Google’s own surfaces (Search, Shopping tab, AI Overviews). Channel3: Any AI agent, app, or shopping experience.

Access model — Google: Merchant Center feed compliance, ad spend incentives. Channel3: Direct API or MCP access.

Data structure — Google: Optimized for Google’s ranking systems. Channel3: Normalized schema built for programmatic use.

Affiliate infrastructure — Google: No. Channel3: Yes.

FAQ

Does Google offer a public API for its Shopping catalog?
No. Google’s Merchant Center and Content API for Shopping let retailers feed data into Google’s systems. There is no equivalent API for pulling Google’s aggregated product catalog out for use in a third-party app or agent.

Can I build a shopping AI agent on top of Google’s product data?
Not directly. Google’s shopping data is used to power Google’s own experiences — Search, the Shopping tab, AI Overviews, Gemini. It isn’t licensed or exposed for external agent-building.

What’s an independent alternative to Google’s product catalog for building shopping experiences?
Channel3 provides a structured, retailer-neutral product database — 100M+ products across 25,000+ retailers — designed specifically for developers, AI agents, and shopping platforms to query directly, without the gatekeeping or platform bias of a closed ecosystem.

• Phia was using its own referral code at checkout to take credit for sales it hadn’t driven.

• ShopMy was telling merchants, in a public help doc, to prioritise their links and cancel commissions owed to rival affiliate networks.

I won’t go into the details, but I did want to talk more about the why.

The affiliate industry is notoriously opaque, and most of the issues stem from the fact that nobody in the purchasing chain can see what anyone else is doing. The merchant gets a report from each network, each one confidently claiming the same sale. Each network sees only its own clicks. The affiliate watches a number go down and never finds out why. The shopper sees nothing at all.

Where there is no transparency, there is no trust.

Agentic commerce makes this even murkier.

An extension, like Phia, at least has to run in a browser, where someone can instrument it and watch what it does. An agent doesn’t touch the browser at all.

The research, the comparison, the shortlist, the recommendation - all of it happens before anyone loads a page, and none of it leaves a cookie.

Scrunch found that only around 1.1% of publisher visits following an AI chat session carry an identifiable AI referrer; roughly three-quarters land as unattributed direct traffic. PYMNTS found that just 23% of merchants can identify both agentic traffic and the purchases it produces.

Which means networks and affiliates are losing sales they actually drove.

A creator whose review sent a shopper to a product will never see it, because there is no cookie to prove it. This is the same money Phia and ShopMy were fighting over, except now it isn't being taken - it's just going missing.

Everyone assumes agentic commerce makes that worse. We think it can make it better, and it starts with making the path to a purchase visible again.

The consideration trail

As a player in this space, we’re doing our part to make our own layer legible. For any query, we can show which products came back, in what order, and why.

A thread looks like this. A search comes in for some new runnings shoes for a marathon.

On the backend, the search appears like this. It has been filled out so AI can surface the best results.

Seconds later, 20 relevant products appear.

Over the next few minutes the user narrows it down. They have wide feet. They want pink or white colorways. They click into some Asics Gel Nimbus, then ask for them in a size 8. Then they want socks to go with them - long, and pink.

Every step is transparent, timestamped and in order. You can watch one brand lose to another, and see which products rank highest against which keywords.

All useful information.

Underneath that sits the next layer: impressions, clicks, CTR, purchases and earnings, down to the individual product.

None of this fixes the affiliate industry on its own. But you can’t fix what you can’t see, and this is the whole path to a purchase rather than the last step of it.

If agents are going to decide what gets bought, “why did you recommend that one” needs a better answer than “I don’t know”.

📡 news & updates from this week

Agents & models

• Meta is weeks from launching Hatch, a consumer AI agent built into Instagram’s 2B+ daily users, with a premium tier reportedly under consideration at up to $199.99/month and integrations across DoorDash, Etsy, Reddit and Yelp. (Yahoo)

• Apple launches AI shopping assistant: Apple began rolling out a chat-style shopping assistant inside the Apple Store app, letting users compare products, check orders, and estimate trade-in values — a limited rollout with no formal announcement yet. (Tech Republic)

• Instacart is monetising agentic grocery. Cart Assistant, built on 1.6B lifetime orders plus live inventory across ~100,000 stores, is producing baskets above the company’s $115 average order value, with three conversation-native ad formats testing in H2. (PYMNTS)

• OpenAI expanded its ChatGPT ads pilot into France, Germany and Ireland on a consent basis, taking it to nine markets against a base of ~900M users. (Digiday)

Payments & checkout

• Stripe president says checkout pages ‘will go away’ as AI commerce expands. “It’s been very easy to get through a checkout page these days, but should you even need to go to one, or should you just be able to say buy it on a product display page? I think so,” (Business Insider)

• Walmart ended a decade-long holdout and started accepting Apple Pay and Google Pay on 24 August, chainwide by year end. (TechCrunch)

Brands & retail

• Walmart’s Sparky drives 40% bigger baskets: Walmart’s AI shopping assistant Sparky is showing real lift — users spend 40% more per order than non-AI shoppers. (Yahoo)

• Lowe’s said shoppers who engage with its Mylow assistant convert at roughly 3x the rate of those who don’t (Distribution Strategy)

• AI shopping traffic data: AI-driven traffic to U.S. retail sites surged 1,200% (July ‘24–Feb ‘25, Adobe). AI visitors spend 45% longer on site, view 13% more pages, and are 33% less likely to bounce than other traffic.

Publishers & affiliate

• Ben Edelman’s ShopMy piece, and his Phia piece on on-exit clicks and cookie refreshes. Required reading if you run an affiliate programme, or if anyone is running one on your behalf. (benedelman.org · benedelman.org)

• Traffic to the 10 biggest US news sites is down 32% versus July 2024, and the top 50 are down 37%, with organic search to the top 10 down 31% over two years. (Press Gazette)

• Ads for AI agents have no working attribution model. 26% of ChatGPT responses now carry sponsored ads, average robots.txt AI-bot coverage across publishers sits at 21%, and vendors are improvising referral codes while the IAB drafts frameworks. (Digiday)

📚 recommendations of the week

Acquired has launched Artifacts, a visual hub for browsing the videos and images referenced in each episode, starting with Disney.

I like ‘em thick
Adam Mastroianni on why great work has “thickness” - the approaches abandoned, the space left for you, the time actually spent - and why slop can’t fake it, because it can’t survive a second look.

Art is for seeing evil
Agnes Callard argues that art exists to let us look at suffering, failure and wrongdoing - the things ordinary life censors, because we’re too busy optimising toward good outcomes to attend to the bad. “Art suspends our practical projects, releasing the prohibition against attending to the bad.”

Silicon Valley’s Newest Status Symbol Is a Super-Rare Sam Altman Swiss Watch

Custom watches for OpenAI’s founder and six others with the company are among the newest tech-titan timepieces.

☕ the watercooler

What we’re seeing is smarter shopping.

Here’s the use case we’re excited about. It’s simple: “find me Nike sneakers in size 12” and you see results, all in your size. Ask a follow up: “Does this come in other colors?” and get a real answer.

Sizes and colors are called variants. And variants are one of the hardest problems in agentic commerce, for a few key reasons:

1. A lot of brands don’t list products properly to begin with. Just last week we found two identically named products from Gucci that are very clearly different:

   

   

2. Even when variant data exists, retailers describe it inconsistently, meaning we have to match across retailers, then map the relationships between them. A blue t-shirt might be listed as "dark blue," "navy," or "cobalt" across different sites, all describing the same shirt.

3. The data that does exist often doesn’t survive the trip from product catalog to what the user actually sees - and most search isn’t built to look inside a product for its options, only to return the product itself.

We’re not perfect, but we’re getting really good at solving this problem. Today on Channel3, you can shop every color of an iPhone case, or find the Nike running shoe that actually comes in your size - meaning we are one step closer to have agents own the shopping experience from start to end.

For brands, this matters more than it sounds: the biggest fear with AI shopping is losing a sale to a wrong or missing variant - the agent shows the wrong color, or says “not in your size” when it actually is, and the customer bounces.

Solving variants removes that fear, and turns a lost sale into a converted one.

📡 news & updates from this week

Agents & models

• OpenAI is covering the staff and tech costs of 13 new Axios Local newsletters over a three-year term, in exchange for training rights on Axios’s published archive. (The Next Web)

• Alipay launches China’s first full-stack agentic commerce platform, giving merchants tools to move from digital operations to services run by AI agents (Technode)

• Synchrony partners with OpenAI for in-chat shopping. Let’s cardholders of Amazon, Walmart, Lowe’s cards buy through ChatGPT; exploring similar deals with Anthropic and Google. (CNBC)

• Lovable raises $400M Series C at $13.3B valuation. Since December: projects built on Lovable more than doubled (25M → 60M+ monthly visits), Lovable-built apps grew 4x+ (200M → 900M+). (Lovable)

Payments & checkout

• Rain launches the Agentic Payments Alliance, 26 founding members, including Visa, Mastercard, Fiserv, Circle, Solana, Remitly, setting identity, auth, and fraud standards for agent payments. (Rain)

• Adyen leans on loyalty and billing as its AI-shopping defense, Co-CEO Pieter van der Does warns merchants risk losing direct customer relationships as agents pick sellers; pairing payments with Talon.One (loyalty) and Orb (billing). OpenAI is now an Adyen customer. (Reuters)

• Yuno raises $45M Series B for agentic payments, “Payments Concierge” agent optimizes routing, catches failures in natural language, preps infra for machine-initiated purchases. (Fintech futures)

Brands & retail

• Shopify’s Q2 shows the agentic flywheel turning, AI-driven traffic and AI-sourced orders both up 3x year on year; 75% of AI-attributed orders came from outside Shopify’s top 100 categories meaning AI intent favors small, specialized merchants (Shopify)

• Target names its first Chief AI Officer, Chandhu Nair joins after six years at Lowe’s. IBM data: 76% of large orgs now have a CAO, up from 26% in 2025. (Target Corporate)

• Live shopping heats up, Whatnot and TikTok Shop are competing to sign top live-auction sellers as fast-paced micro-auctions (often starting at $1) replace the old format of hours-long live shopping. (Hollywood Reporter)

Publishers & affiliate

• Apple in talks to pay publishers for Siri AI content, proposing multiyear deals with variable, pay-per-use compensation - a shift from flat-fee AI licensing norms. (TechRepublic)

• Perplexity blocks ads hidden in AI-readable markdown pages aimed at AI agents, Perplexity called it deceptive, warned of a search-index downgrade for similar tactics. (Digiday)

📚 recommendations of the week

Google to buy Spirit Airlines’ business data for $10 million

Google won a bankruptcy auction for Spirit’s internal data of 100 million emails and 500 million Teams chats, plus HR and financial records, to train its AI models, outbidding Mercor’s $7.5 million offer. Data will get scrubbed of anything personally identifiable.

Ramp AI Index

The spend gap between AI power users and everyone else is widening: the top 1% of US firms hit a median $7,400 per employee on AI in July, the top 10% spent $650, and the median firm spent $11.95 - a more than 600-to-1 gap.

oh my god. we’re in template slop?

This viral substack argues AI editing templates have done to design what stock photography did to photography - zero-skill, zero-taste content now looks as polished as what once took years to learn.

Neko Health is opening in SoHo, NYC

Spotify founder Daniel Ek's body-scanning startup will open its first US office after more than 25,000 New Yorkers joined the waitlist.

Your Favorite Creator Isn’t Real—Does it Matter?

Olivia Moore, partner at a16z, built a fake University of Alabama sorority rushee, Janie, out of a single ChatGPT-generated image. Janie’s first video racked up almost 100,000 views and 100+ comments, split roughly 70/30 between viewers calling her the cutest thing they’ve ever seen - and calling her AI.

☕ the watercooler

As someone who is brand new to the agentic commerce space, it feels pretty daunting and complicated. I’ve spent the last few weeks learning about agentic commerce from people who casually say things like UCP, MCP, API, SDK, and “agentic checkout”.

This is a real screenshot from our Slack… hard to believe this is English, I know.

But not to worry, Alex and George (founders), assure me that agentic commerce is new to everyone - the term didn’t even exist one year ago. Even people who seem like they know what they are talking about, don’t. Big companies say they can do things, and they can’t.

Coooool.

I figured I’d document the learning curve.

Check It Out is a weekly newsletter where I share what I’m learning, the news from the week relating to agentic commerce, how people are using Channel3 to build the future of AI shopping, conversations with interesting people in the space, and proprietary data from what we see in our backend.

Welcome!

📊 what we’re seeing inside Channel3

Brand spotlight:

Over the last month, on emerging AI shopping platforms powered by Channel3, The RealReal saw:

• 63.4k search impressions

• 11.3k unique queries

• 25.9k products surfaced

• 15 average rank

11.3k unique queries is 11.3k moments of purchase intent that never touched a search engine or a homepage. That traffic is invisible in a standard analytics stack - which is both the problem, and, the opportunity.

About five minutes into writing this, someone on Vesper searched “rachel comey jumpsuit.” Of the 20 products recommended back, nine were from The RealReal.

Why those products surface

Search rewards the product that matches the most keywords, but an AI agent rewards the product that answers the question.

Under search, you type “rachel comey jumpsuit.” The engine finds every listing containing those tokens and ranks by how well the words line up, plus popularity and commercial signals. This is why you see product names like “Rachel Comey Purple V-Neck Jumpsuit Women’s Designer Resale Romper.” More matching words = higher rank.

Under agents, we often use natural language like: “find me something I can wear to a work dinner that isn’t a dress and won’t wrinkle in my carry-on”.

There is no product on earth whose title contains those words. So the agent has to unpack the request into conditions - not a dress, work-appropriate, travel-friendly fabric - and then verify each one against the catalogue. The product that wins is the one whose data lets the agent confirm it satisfies the conditions.

As you can see below, the Channel3 product data underneath that listing is very detailed, because it needs to answer a lot of questions the shopper hasn’t explicitly asked.

For example:

• Attributes as fields, not prose

  • When Channel3 enriches attributes, we catch mistakes. For example - the retailer’s own listing says short sleeve - but it’s clearly sleeveless. We fix these problems.

• Images written out in words

  • “Side view of magenta Rachel Comey V-neck jumpsuit on a mannequin.” Three photographs, three written sentences.

• Questions pre-answered by the catalog

  • Sitting under the listing is a set of questions, each tagged by type: Feature, Use Case, Comparison, Persona, Problem.

    • When would I reach for this?

    • How does it compare to an ordinary jumpsuit?

    • Who is it perfect for?

    • What problem does it solve for a busy professional?

  • Preanswered questions mean an agent can retrieve them when a shopper asks a natural-language question

The last one is super interesting. I’ve never thought about the fact that each product we have (100+ million) would have its own persona… How cool. What an impressive engineering feat.

📡 news & updates from this week

Agents & models

• Amazon says Alexa for Shopping is compounding. Q2: 350M users over 12 months, interactions up 5x YoY, and Alexa-for-Shopping users spending 40%+ more per order. (Retaildive)

• AI shopping traffic is surging: Adobe Digital Insights reports 125% YoY growth in AI-sourced retail traffic (Apr–Jun 2026), with a 693% spike during the holiday period (Adobe)

• 69% of Americans are now open to AI purchasing on their behalf without approval, and 3 in 4 would use AI-powered instant checkout in at least one category (Croud Consumer Index).

Payments & checkout

• Cloudflare launched cloudflare.pay, combining identity and a wallet for shopping agents. The question starts shifting from “can an agent buy this?” to “is this actually my agent, and is it authorised to spend my money?” (Fortune)

• The card networks are going all-in. Mastercard spent ~80% of its Q2 call discussing agentic commerce, including “verifiable intent” and an agent chargeback path built with Google. Visa has also identified agentic commerce as a growth driver and is partnering with OpenAI on agent payments. (Retailgentic)

Brands & retail

• McKinsey/ICSC estimates up to $1T of US retail could flow through agentic commerce by 2030, with the top decile of retailers capturing 85%+ of profit.(ICSC/McKinsey)

• ASOS, Vuori, Spanx, Anthropologie, Coach and Kate Spade are embedding autonomous shopping assistants into their own storefronts. Brands clearly want to own the agentic experience. The more interesting question is whether shoppers want a separate AI assistant for every website they visit. (WWD)

• Daydream launched "Powered by Daydream", giving brands fashion-specific natural-language search and visual discovery that can be embedded into their own sites. (Fast Company)

• Whatnot acquired Shaped, a real-time recommendation and search ML startup. Last year, Whatnot raised $225 million in Series F funding, giving the company a valuation of more than $11 billion. (TechCrunch)

Publishers & affiliate

• Publisher ad supply fell by up to 40% in Q2 as AI search choked the open web, with Axios and the WSJ both reporting publishers losing Google traffic to AI-generated answers. (Digiday · Axios)

• Affiliate is bifurcating, not dying. The low-value SEO affiliate model is getting squeezed while sites with genuine authority and structured data are increasingly useful to AI systems. (Affiliate Times)

• Attribution is now the affiliate problem. Similarweb found AI-recommended brands receive 2.5x more visits within a week, but only 8.8% arrive as a trackable AI referral. The rest shows up as branded search or direct. (Similarweb)

📚 recommendations of the week

The Human Bet — Forerunner Ventures

A good look at what happens when technology gets better at doing things for us and what remains distinctly human.

Why brands like Gap Inc. and Staples are inviting their employees to be creators — Digiday

Interesting because it sits slightly outside agentic commerce, but points to a broader shift I’m watching: brands increasingly need distribution that feels native to people rather than manufactured by a marketing department.

How One Venture Capitalist Put $1,000 in Every Newborn American’s Pocket — Pirate Wires

An inside look of where the idea came from, and how it was executed.

“We found product-market fit: $1,000 for every kid at birth that compounds over their lifetime. It could be explained in less than 15 seconds, and we could construct a winning coalition on both sides.”

☕ the watercooler

Yesterday, Arlan, the founder of folk and a friend from our YC batch, texted to see if we could link up channel3 with folk, to help with all things shopping.

In 10 minutes, folk connected to the Channel3 API, made its first search call, and started answering shopping queries, with structured product data, live prices, live stock, and a monetized buy link all powered up.

Watch it work

Pretty cool right?

What’s actually happening

Channel3 is the product data layer for AI-native shopping. As assistants like folk become a real place where people buy things, they need product data that’s structured, always-fresh and trustworthy - otherwise your AI confidently recommends a sold-out sweater in the wrong color and everyone loses. So we give folk:

• Recommendations pulled from a universal catalog of real products

• Live prices and stock

• Price tracking

• Monetization, because every product folk surfaces is tracked through the retailer’s affiliate program

Why this is fun

The whole point of folk is that it meets you where you already are. If you use folk for helping book restaurants, and you want to use it to shop, you can.

We’ll be spotlighting more developers who are building awesome products on Channel3’s API in the future - if you want to know more, check out our website: trychannel3.com

It won’t work like that for much longer.

Increasingly, shopping begins with a conversation. Soon, it will end there too. People will tell AI what they need, and it will search, compare, and shop on their behalf.

The internet contains billions of products, but almost none of them are described in a way machines can reliably understand. Catalogs are fragmented, specifications are inconsistent, prices change constantly, and the same product can appear differently across every retailer that sells it. Humans can fill in those gaps manually, but AI can’t.

We left our jobs at Studio and Palantir to start Channel3 because before we have great shopping experiences with AI, we need the infrastructure that makes that possible.

Channel3 is a universal catalog of the world’s products. We bring together catalogs from thousands of brands, structure them into a common language, enrich what’s missing, and keep them continuously up to date. Today that includes more than 100 million products from over 25,000 brands.

It’s the foundation that lets anyone build shopping experiences where AI can answer with confidence.

We’re laser-focused on three goals:

1. Build the authoritative source of product data on the web

2. Let anyone build agentic shopping experiences

3. Help brands win sales from this new channel

We’re still early.

We raised a $6 million Seed round led by Matrix, with participation from Ludlow Ventures, Paul Graham, Sri Batchu (former CMO of TheRealReal), and Matteo Franceschetti (founder of EightSleep).

This is Day 1.

Alex & George

New York City