Doclight documentation
What Doclight measures, how its features fit together, and how to connect it to AI tools through MCP.
Doclight is an AI visibility tracker. It tracks whether ChatGPT, Claude, Perplexity and Gemini mention or cite your brand, and turns what it finds into evidence-backed actions. These docs explain what is measured, how to read the numbers, and how to read (and optionally write) that data from an MCP client such as Claude Code.
What Doclight measures
You register a brand (a project) and the questions your customers might ask an AI assistant. Doclight runs those questions on a schedule or on request and records what comes back. Several independent signals feed the product:
| Signal | What it is | What it is not |
|---|---|---|
| Tracked-question measurements | Each question is sent to provider APIs (ChatGPT, Claude, Perplexity and Gemini are measured; Google AI Overviews is on the roadmap and is not measured). Answers are parsed for the domains they cite or mention. | Not what consumer chat or search apps show you. See Metrics and methodology. |
| Crawler observations | Visits by known AI crawlers, reported to Doclight by a site integration you install. | Not citations, and not proof of blocking when empty. Doclight does not read your server logs itself. |
| Website Crawl | A bounded crawl of your site that records technical facts: broken internal links, sitemap errors, noindex and canonical tags, robots.txt rules that block answer-surface crawlers, and structured-data validity. | Not a ranking prediction. If Doclight's crawler was refused, that is reported as a coverage gap, never as "bots are blocked". |
| Agent Task | A bounded, read-only browser agent that tries one task on your site (for example, finding a product or a fact) and reports an outcome. | Not a customer simulation. A run that could not execute (provider not configured, timeout) is recorded as failed with no outcome and never counts as a website result. |
| Recommended actions | Rules that combine crawl data, validated AI content assessments, measurements and agent runs into prioritised actions, each with its evidence. | Not a forecast. The evidence support score (0–100) says how much observable evidence backs a finding; it is not a probability of success or a predicted uplift. |
| Reports and alerts | A report summary with citation share, citation rate, provider coverage and data sufficiency; calendar-month and quarter snapshots; exports; alerts such as visibility drops, competitor overtakes and first-observed crawlers. | Percentages are never fabricated: when there is not enough data the value is null. |
How the pieces fit together
Measure
Tracked questions produce provider responses. Those responses are the denominator behind citation share and citation rate.
Observe
Crawler observations and Website Crawl results describe what happens on and around your site. They give context for a measurement result but do not change the measurement itself.
Act
Recommended actions are generated from the signals above. Content actions are always labelled as hypotheses produced by an AI assessment whose quotes are verified against the page text; technical actions are deterministic.
Re-check
Re-running a measurement or a crawl updates an action's status (for example verified_resolved or still_present).
Absence in a newer scan is reported as "no longer observed", not as "fixed".
Read the measurement basis
Doclight's numbers come from calling provider APIs with your tracked questions. They describe those API responses. They do not describe what a person sees in a consumer chat or search interface, which can differ.
Where to go next
Getting started
Prerequisites, creating a scoped API credential, and your first successful MCP read.
Metrics and methodology
Denominators, comparisons, data sufficiency, and what each metric does not claim.
MCP concepts
How an MCP client, the local Doclight server and the Doclight API relate.
MCP tool reference
Inputs, outputs and examples for all seven shipped tools.