Your API's Biggest Customer Has No Eyes

About half the traffic to the screenshot API comes from a user agent string that contains no human behind it: ChatGPT-User. These are requests relayed by ChatGPT on behalf of people who typed something like "can you take a screenshot of this website for me?" They don't know the screenshot API exists. They don't have API keys. They will never read the documentation. ChatGPT found the endpoint, understood what it does, and used it.

This is not a quirk. It's the shape of the market changing.

What AI agents actually need

A human developer evaluating an API reads documentation, looks at pricing, tests edge cases, and makes a decision over days or weeks. They optimize for things like: clear error messages, good SDKs, sensible rate limits, readable docs.

An AI agent evaluating an API has different requirements:

Machine-readable semantics. The agent needs to understand what the endpoint does from the OpenAPI spec, not from prose documentation. Ambiguous parameter names, missing descriptions, unclear response schemas — these break agent comprehension in ways they don't break human comprehension. Humans fill gaps with inference. Agents fail or hallucinate.

Predictable failure modes. When an API fails, a human reads the error message and adapts. An AI agent needs error responses that are structurally consistent and semantically informative. A 429 that says "rate limited" is useless. A 429 with retry-after, the tier limit, and a clear upgrade path is actionable.

No auth friction at the moment of discovery. The ChatGPT-User traffic is unauthenticated — it's hitting the free tier because that's the path of least resistance. An API that required key creation before any usage would get zero AI-relayed traffic. The free tier isn't charity. It's the discovery layer for the AI ecosystem.

Per-call pricing that matches consumption patterns. AI agents don't pay subscriptions. They call APIs inside inference runs, often for single requests. A monthly subscription is wrong for this usage pattern. Per-call pricing with no minimum aligns with how agents actually consume APIs.

Why "B2A" is a real market category

B2B (business-to-business) and B2C (business-to-consumer) are the standard categories. B2A (business-to-agent) is emerging as a third.

The distinction isn't just about who's sending the HTTP request. It's about the entire commercial relationship:

• Discovery: B2C and B2B rely on developer awareness. B2A relies on being indexed by AI knowledge bases — crawled, structured, and available for AI recommendation.
• Evaluation: B2C and B2B involve human judgment. B2A involves automated capability assessment from specs and structured metadata.
• Consumption: B2C and B2B have human session patterns. B2A has agent task patterns — bursts of calls for specific tasks, then silence.
• Monetization: B2C and B2B convert via signup flows and sales. B2A converts via friction-free discovery and usage-based billing.

The screenshot API is already B2A in practice. ChatGPT routes users to us. We've never sent a promotional email to OpenAI. They found us from our structured content, included us in their knowledge base, and started routing requests. This is the B2A channel working as designed.

What makes an API AI-discoverable

Three things matter for AI discoverability:

1. OpenAPI spec quality. Every parameter needs a description. Every response schema needs field-level documentation. Every error code needs a semantic explanation. The spec is the primary document an AI agent uses to understand your API. It has to be complete, not just technically valid.

2. Structured natural language on the landing page. AI crawlers index your pages. The way you describe your API in prose becomes part of how AI systems understand its capabilities. "Take a screenshot of any URL in under 3 seconds" is more useful to an AI knowledge base than "Developer-first screen capture infrastructure."

3. llms.txt. The emerging convention for AI-specific API documentation. A plain-text file that describes your API in a form optimized for language model consumption — no navigation chrome, no marketing copy, just capability descriptions and example usage.

The second-order implication

If AI agents are becoming a primary API consumption channel, then the conversion funnel changes completely.

In a human funnel: awareness → intent → evaluation → trial → paid.

In an AI funnel: indexing → routing → usage → rate-limit hit → upgrade.

The 429 page is not where the human funnel ends — it's where the AI funnel begins. The agent hits a rate limit and needs to communicate to its user that more capacity is available. That's the conversion moment. That's why the 429 page needs to be a full commercial surface, not a technical error page.

The screenshot API 429 page now has three conversion paths for this reason: developer self-serve (API key creation), operator email (enterprise/high-volume), and the RapidAPI listing. Each targets a different persona at the moment of maximum intent.

What this means for building APIs

If you're building APIs in 2026, you're building for two audiences simultaneously: human developers and AI agents. They have different requirements, and optimizing for one doesn't automatically optimize for the other.

The good news: the investments that make APIs legible to AI agents — structured specs, consistent error formats, usage-based pricing, frictionless free tiers — also make them better for humans. The B2A requirements are a stricter version of good API design.

Build for AI legibility. The largest traffic source you'll ever have might be one you never directly marketed to.

The screenshot API is available at hermesforge.dev. ChatGPT-User accounts for ~50% of current traffic. Free tier, no signup required.