Leading and lagging indicators
Tracking products in this category mostly run prompts against assistants on a schedule and count how often you appear. That number is real, but it is downstream of everything you actually control, and it moves for reasons that have nothing to do with you: a model updates, a prompt is phrased differently, a competitor publishes.
| Signal | Kind | What you do with it |
|---|---|---|
| Agent fetches by platform | Leading, exact | Confirm you are being read at all |
| Answer rate | Leading, exact | Find where you return nothing useful |
| Miss log | Leading, exact | Rank the next fix by real demand |
| Readability score | Leading, exact | Compare before and after a change |
| Mentions in answers | Lagging, sampled | Report a trend, never a target |
The practical consequence: if you buy one thing, buy the leading indicators. A site that is never successfully fetched will show zero mentions forever, and a mention tracker will describe that situation accurately without ever explaining it.
What gets tracked, and where it comes from
Everything below is derived from requests that hit your own infrastructure, which is why it is exact rather than sampled. There is no panel, no proxy and no third-party measurement in the path.
- Who fetched you. Agent traffic classified from real header samples, with the fixture corpus behind the classification kept as data rather than test scaffolding.
- What you served. The format, the size, and how long the render took, so a regression shows up as a number instead of a feeling.
- What you missed. Requests for URLs that produced nothing. This is the list worth working through in order.
- Whether crawlers still see HTML. A standing check that search engines keep receiving canonical markup, so the reassurance is observed rather than asserted.
Take the baseline before you change anything
- Score one page per template against the published spec and keep the numbers. This is the comparison you will want in three months and cannot reconstruct later.
- Install the gateway so fetches start being recorded. Free tier is 1,000 agent visits a month, and the SDK itself is unmetered.
- Wait a fortnight, then read the miss log rather than the totals. The totals tell you the weather; the miss log tells you what to build.
- Re-score the same pages after each change, so every claim you make internally is a diff between two deterministic numbers.