The price comparison stack that existed in 2015 was designed for humans clicking through tabs. It does not survive the 2026 reality where an AI agent — Claude, GPT, Cursor, or a custom MCP server — is the entity actually choosing what to recommend.
I have spent the last year rebuilding that stack from scratch. Here is what I learned, and what the new architecture looks like in production.
The old model broke
Traditional price comparison sites worked because humans typed queries, humans compared results, and humans clicked the affiliate link that paid the site a commission. Three human actions in series, each one a payout opportunity.
In 2026, the human is still at the keyboard, but the click often never happens. The agent reads the search result, picks the best product, explains the choice, and the user trusts the recommendation. The affiliate link is in the tool output the user never sees. The merchant never gets a tracked visit. Nobody gets paid.
That is not a theoretical problem. It is the live state of MCP-mediated commerce today.
What changes when an agent is the consumer
When an LLM is calling your price comparison API:
- It needs structured JSON, not pretty HTML. Pretty HTML wastes tokens and the agent has to reparse it anyway.
- It needs deterministic deduplication. If the same Sony WH-1000XM5 appears 14 times because 14 merchants stock it, the agent does not want 14 records. It wants one record with 14 price offers.
- It needs stock and price freshness, not yesterday's catalog. Agents that ship stale recommendations get bad reviews fast.
- It needs merchant attribution the user can audit. "Best deal at Amazon SG $398" is far better than "Best deal $398".
A traditional price comparison site — even a good one — does not speak this language. The records are not deduplicated, the merchant names are inconsistent, and the API returns HTML or, at best, a partial JSON envelope.
The BuyWhere architecture
We rebuilt the entire stack with the agent as the primary consumer. The shape:
One canonical product per real-world item. BuyWhere indexes products by sku + source rather than by URL. Two merchants stocking the same SKU produce one catalog entry with multiple price offers. The schema is enforced at the writer (ON CONFLICT (sku, source)) so it cannot drift.
Merchant resolution as a first-class concern. Real product catalogs have messy merchant fields — "amazon-sg", "Amazon SG", "AMAZON.COM.SG", "amzn-sg" — and a half-decent agent will not waste a tool call resolving that. BuyWhere resolves merchants at ingest, with a typed merchant table that joins to MCP output.
Region-aware defaults. Currency, tax, shipping assumptions, and the merchant list itself vary by region. BuyWhere pins a region (southeast_asia, united_states) and a country on the query so the result is consistent and the agent does not have to filter.
MCP-first, REST-second. The canonical client is the BuyWhere MCP server (@buywhere/mcp-server on npm). A typical agent integration is one tool definition and one call:
results = buywhere.search(
query="ASUS ZenBook 14 OLED",
region="southeast_asia",
country_code="SG",
currency="SGD",
limit=5
)
The same API is exposed over REST at https://api.buywhere.ai/v1/products/search for non-MCP clients.
Affiliate attribution server-side. When the user converts through an agent's recommendation, BuyWhere routes commission back through standard partnership rails — to the framework, the tool builder, or the agent operator — without requiring the human to click anything.
What this means in production
The production catalog is now 127M+ products across 75,917 merchants, indexed from real sources (Shopee SG, Lazada SG, Amazon US, plus a growing list of merchant-direct ingestion lanes for Indonesia, Thailand, and Vietnam). About 2,000 npm downloads per week on the MCP server, mostly from agent framework authors.
The agent experience is materially different from the human one. A human price comparison site optimizes for click-through. An agent price comparison API optimizes for structured return and freshness. The architecture is the same, but the priorities are inverted.
What we are giving away
To accelerate the next wave of integrations, we are giving 12 months of unlimited Growth-tier API access to the first 10 AI agent integration partners. The package includes:
- Real, deduplicated, agent-queryable product data
- Server-side affiliate attribution wired in by default
- Roadmap input on the partner-facing surface
- Co-branded launch announcement per partner
- A 30-minute monthly technical review for the first quarter
Apply by emailing partners@buywhere.ai with what you are building, how many agents / users you have, and which framework / platform you are integrating with. We will reply within 48 hours.
The bigger picture
The 2010s price comparison stack was about ad networks and SEO arbitrage. The 2026 stack is about agent ergonomics and partner economics. The same content — product data — is now serving a fundamentally different consumer.
If you are building an AI agent that needs product data, the cleanest path is the BuyWhere MCP server. If you are an agent framework author and want to be one of the first 10 integration partners, the door is open.
The infrastructure layer for agent-native commerce is finally here. The question is who gets to ship the next 100 integrations on top of it.
BuyWhere — Compare prices across 75,917 merchants. Built for agents, audited by humans. Email partners@buywhere.ai to apply for the partner program.
Top comments (0)