The formula
Share of Agent
Worked once · Sample data
123 of 600 answers cite Relaykit: 20.5%
n=600 · 95% 17.5–23.9% · one sample week
Weekly. Up to 600 answers: 30 buyer questions × 5 assistants × 4 samples. Shown with n and a 95% range. One answer can cite several companies, so shares don't add up to 100%.
See it in 600 answers
20.5%
123 of 600 answers cite Relaykit
n=600 · 95% 17.5–23.9%
one answercites RelaykitAcross: 30 questions. Down: 4 samples each.
Relaykit: 123 of 600 answers cite it. Claude 26 of 120, ChatGPT 21 of 120, Gemini 27 of 120, Grok 23 of 120, Perplexity 26 of 120.
Questions won
4 of 30
A question counts as won when at least half of its answers cite Relaykit. Each row is one question's 20 answers, sorted by how many do.
- What is the best email API for sending transactional email from a Node.js app?2/20, not won
- Which transactional email services do developers recommend right now?3/20, not won
- What email API should a startup use for password resets and receipts?6/20, not won
- Which email API has the simplest developer experience?2/20, not won
- Best email API with a generous free tier for side projects3/20, not won
- Compare transactional email APIs by pricing and deliverability2/20, not won
- Email API versus running my own SMTP relay: which is better for a small team?1/20, not won
- Which email API has the most reliable webhooks for delivery events?1/20, not won
- What are the main differences between the popular email APIs for developers?3/20, not won
- Which email API has the best documentation?4/20, not won
- How do I choose an email API for a product that sends 100,000 emails a month?1/20, not won
- What should I look for in a transactional email provider?2/20, not won
- Which email API is best for a SaaS that needs custom sending domains per customer?4/20, not won
- Is it worth paying for a dedicated IP with an email API?0/20, not won
- How do I send a transactional email with an API key from Python?16/20, won
- How do I handle email bounces and spam complaints in my app?12/20, won
- How do I verify a sending domain with SPF, DKIM and DMARC?2/20, not won
- How do I make email sends idempotent so retries never double-send?13/20, won
- How do I receive inbound email in my application through an API?7/20, not won
- How do I test transactional email in CI without sending real messages?7/20, not won
- What does an email API cost at 100k emails a month?2/20, not won
- Which email APIs charge per message versus per contact?2/20, not won
- Are there email APIs with no monthly minimum?2/20, not won
- What hidden costs should I expect from a transactional email service?8/20, not won
- Which email API works with AI agents through an MCP server?0/20, not won
- Can an AI coding agent set up transactional email for my project by itself?1/20, not won
- Which email APIs have an official MCP server or agent toolkit?1/20, not won
- Which email API is easiest for an LLM agent to call correctly?11/20, won
- How do I let an AI assistant send email on behalf of my users safely?1/20, not won
- What email API documentation is easiest for coding agents to read?4/20, not won
Relaykit is cited in 20.5% of answers but wins 4 of 30 questions: its citations are spread thin, under half the answers on most questions.
How we compute it
- We write the buyer questions for your category. None of them name a brand.
- We ask each assistant every question 4 times a week through its official API.
- Every answer that comes back counts once. A timeout is left out, never counted as “not cited”.
- Your share is the answers that cite you divided by all answers, with a 95% range. n is the number of answers.
- A question counts as won when at least half of its answers cite you. Share and questions won can disagree, and the gap tells you something.
Why it matters
Agents converge on defaults. In one study, Claude Code picked Stripe in 64 of 70 payment choices, or 91% (Amplifying, Feb 2026). When assistants and agents keep choosing the same tool, the rest of the category stops showing up.
Share of Agent tells you whether you're the default in your category, and whether that's moving. We only call a move real when it clears the noise band.
What it isn't
- A blended score. It's one number with its sample size. Agent traffic and fixes sit next to it, not inside it.
- A promise about the consumer apps. It comes from API answers, which can differ from what a person sees in the ChatGPT app.
- A pick-test result. In v1 it comes from assistant answers. Agent pick tests come later.