Skip to main content
This page defines the terms and metrics you’ll encounter throughout Lightsage. Use it as a reference when you need to understand what a score measures or how a feature works.

​
Optimization categories

GEO is the practice of optimizing your content to appear in AI-generated answers from answer engines — systems like ChatGPT, Perplexity, and Google AI Overviews that respond to questions in prose.For consumer brands, GEO is the primary channel: if ChatGPT mentions your product when someone asks “what’s the best X?”, you win.For API companies and developer tools, GEO covers only part of the picture. Answer engines drive awareness, but they don’t drive adoption — that happens in coding agents. See AEO below.
AEO extends GEO to cover action engines — AI systems that don’t just answer questions but take actions. For developer tools, this means coding agents like Claude Code, OpenAI Codex, and Cursor that recommend APIs, write integration code, execute it, and debug failures.AEO tracks two dimensions that GEO doesn’t:
  1. Visibility — Does the agent recommend your API for relevant implementation prompts?
  2. Usability — When the agent writes code with your API, does it work?
Lightsage is built specifically for AEO. It’s the only platform that tracks coding agents alongside answer engines, and the only one that measures API usability — not just brand mentions.

​
AI platform types

Coding agents are AI systems that take actions: they receive an implementation prompt, choose tools and libraries, write code, execute it, and iterate on failures. When a developer says “add payments to my app,” a coding agent like Claude Code or Cursor recommends an API, generates working integration code, and potentially runs it to verify.Lightsage tracks 10 coding agents: Claude Code, OpenAI Codex, Cursor, GitHub Copilot, Gemini CLI, OpenClaw, OpenCode, Hermes, Pi, Kilo.Because coding agents execute real tool calls against your API, evaluating them costs more: 5 credits per run.
Answer engines respond to questions with prose. When someone asks “what’s the best authentication API?” or “compare Clerk vs Auth0,” an answer engine returns a written response. Success is being mentioned favorably.Lightsage tracks 7 answer engines: ChatGPT, Google AI Overviews, Google AI Mode, Gemini Search, Perplexity, Claude, Microsoft Copilot.Answer engine prompts cost 1 credit per run.
Models are the underlying large language models that power agents and answer engines. Tracking models separately lets you see how your API fares across different model providers before those models are integrated into specific agents.Lightsage tracks 4 models: DeepSeek, Kimi, Z.ai, Grok.Model-level evaluations cost 1 credit per run.

​
Visibility metrics

Your Visibility Score is a 0–100 number that reflects how consistently AI platforms recommend your API across the prompts you’re tracking. It weights recommendations by platform reach and prompt intent — an implementation prompt from a coding agent carries more weight than a brand query in an answer engine.A rising Visibility Score means more agents are recommending you for relevant prompts. A falling score means a competitor is gaining ground or your docs have degraded in AI discovery.Track your Visibility Score trend over time in the Visibility Analytics dashboard.
Share of voice (SOV) measures what percentage of relevant recommendations go to you versus your competitors. If five APIs compete for a “best payments API” prompt and you appear in 30% of responses, your SOV is 30%.SOV is a competitive metric — your score is relative, not absolute. You can improve your SOV by gaining recommendations, but you can also lose it if a competitor improves faster than you do.Lightsage tracks your SOV per platform and per prompt type so you can see exactly where you’re winning and where you’re being displaced.

​
API Performance metrics

The API Performance Score is a composite measure of how well AI coding agents can discover and use each of your API endpoints. It rolls up five sub-metrics:
  • Eval score — Overall pass rate across all test runs for the endpoint
  • Discovery score — Whether agents find and correctly identify the endpoint from your docs
  • Tool call success rate — Whether agent-generated code executes without errors
  • Error recovery rate — Whether agents can diagnose and fix failures when errors occur
  • Time to completion — How long agents take to finish a task using the endpoint
Each endpoint gets its own score, so you can prioritize fixes on the endpoints that matter most.
Tool Call Success Rate measures what percentage of agent-generated code executes successfully against your endpoint without errors. When an agent writes POST /v1/charges code and runs it, does it work?A low Tool Call Success Rate on an endpoint means agents are writing broken code — usually because of unclear authentication patterns, inconsistent parameter naming, or unhelpful error messages. Fix these at the source rather than in your docs.Target: 80%+ is considered healthy. Below 50% means the endpoint is effectively unusable by agents.
Error Recovery Rate measures what percentage of failed tool calls the agent successfully recovers from. When your API returns an error, is the message clear enough for the agent to diagnose the problem and retry with corrected code?Generic errors like “Error 401” or “Invalid request” have low recovery rates because agents can’t determine what to fix. Descriptive errors like “Invalid API key — check that you’re using a live key, not a test key” help agents self-correct.Target: 70%+ is considered healthy. Below 40% means your errors are opaque enough that agents abandon tasks rather than retry.
Completion Rate measures what percentage of agent tasks finish using your API versus being abandoned or switched to a competitor mid-task. An agent that starts with your API but encounters repeated failures may switch recommendations partway through — a loss that’s harder to detect than an outright non-recommendation.Target: 75%+ indicates agents reliably complete tasks with your API. Below 50% suggests critical usability problems at the execution layer.

​
Billing

Credits are the billing unit in Lightsage. Each time a prompt is run against an AI platform, it costs credits:
  • AI search platforms (answer engines and models): 1 credit per run
  • Coding agents: 5 credits per run — coding agent evaluations execute real tool calls, which require more compute
The free plan gives you 250 credits to start. The Pro plan includes 1,500 credits per month at $100/month, with additional credits available as top-ups. Enterprise plans have custom credit allocations.You can estimate your monthly credit usage with the pricing calculator at lightsage.com/pricing.

​
Features

Prompt Tracking is the scheduled system that runs your defined prompts across AI platforms on a recurring basis — daily or every other day. Instead of one-off snapshots, you get a continuous time series of visibility and usability data.You define implementation prompts that reflect real developer use cases, select which platforms to run them on, and set a frequency. Lightsage runs them automatically and surfaces trends, anomalies, and competitive movements in your dashboard.
Devtool Arena is Lightsage’s free, public leaderboard ranking APIs and developer tools by how well AI coding agents can discover and use them. Rankings are based on real benchmark runs across discoverability, tool call success, and error rates.You can use Devtool Arena to:
  • See where your API ranks without a paid account
  • Benchmark against specific competitors
  • Track LLM-specific performance differences
The leaderboard is updated continuously as new evaluations run. View it at lightsage.com/leaderboard.
The Opportunities Engine analyzes your visibility and usability data and surfaces specific, prioritized actions to improve your scores. It identifies competitor gaps (prompts where you’re losing share to a specific competitor), content opportunities (topics where you have no presence), and usability fixes (endpoints with low tool call success).Each opportunity includes an estimated impact and effort level so your team can prioritize effectively.
Lightsage’s Content Generation feature produces AI-optimized content based on your Opportunities Engine findings — comparison pages, quick-start guides, llms.txt files, and endpoint documentation structured to improve agent discoverability. Generated content uses your brand kit for consistent voice and messaging.

​
Technical standards

llms.txt is a machine-readable markdown file placed at the root of your domain (yourdomain.com/llms.txt) that gives AI agents a curated description of your API — what it does, when to use it, and how to call it.When a coding agent encounters a relevant prompt and retrieves your llms.txt, it has immediate structured context that biases it toward recommending your API correctly. This is Layer 3 of the 4-layer decision stack — one of the highest-leverage optimizations for newer or smaller APIs.See the llms.txt guide for a template and implementation instructions.
MCP (Model Context Protocol) is an open standard that lets coding agents connect directly to your API as a tool. When your API has an MCP server, agents can discover it, call your endpoints directly within their workflow, and verify that responses work — turning a passive recommendation into an active integration.MCP presence is one of the strongest signals in Layer 3 of the decision stack. An agent with your MCP server installed will prefer your API for relevant tasks because the connection is already established.See the MCP servers guide for setup instructions.
The @lightsage/agent-tracker middleware detects AI coding agents visiting your website and documentation in real time. Install it in one line in Next.js, or integrate it with Vercel, Cloudflare, Express, Astro, or Netlify.It identifies which agents are visiting, which pages they’re reading, and how often — giving you real traffic data to complement the evaluation data from prompt tracking.See the agent tracking guide for installation instructions.