Skip to main content

AI performance metrics: the seven numbers that describe the channel.

AI discovery is its own channel, and it needs its own report. These are the seven metrics that describe it: training coverage, candidate coverage, referral coverage, growth score, engine mix, crawled but not cited, and AI-attributed revenue. Each one is read off first-party server logs, and each one is a panel in the WISLR AI Channel Analytics dashboard.

A funnel showing the four stages of the AI channel, training crawls narrowing to citations, referrals and revenue, with AIO tools covering only share of voice above it and GA4 covering revenue weakly below
The short version
  1. The channel has four stages, not one number. A model reads your pages, fetches some of them mid-conversation, sends a few real people to you, and some of those people buy. Training, candidates, referrals, revenue. Every useful AI metric answers a question about one of those four stages, and a metric from one stage tells you almost nothing about the next.
  2. Coverage is the unit, not volume. A raw crawl count rewards whoever has the most pages. Coverage asks how much of what matters got read, got fetched, and got clicked. That is what makes training, candidate and referral numbers comparable to each other and to last month.
  3. Crawled but not cited is the most actionable number on the page. It isolates pages a model has already read and still does not use. The content is reachable and the model has it, so what is missing is on the page itself, which makes it the shortest path from a report to a decision.
  4. None of this comes from browser analytics. Training crawls and mid-conversation fetches never run JavaScript, so GA4 cannot see them at all. Even for real people arriving from AI, browser-based tools undercount by 2.5x to 5x. These metrics are read off server logs or not at all.

Why the AI Channel Needs Its Own Report

Paid, organic, social and email each get their own report, with their own metrics, because each behaves differently. AI discovery is the newest of them and the least understood, and it is still mostly reported as a footnote inside a channel it does not belong to.

The reason it needs its own report is not novelty. It is that the mechanism is different. There is no ranking position, no impression count, no results page. A model reads your pages on a schedule, decides mid-conversation whether to reach for one of them, and occasionally sends a person to you with a link. Those are three separate events, produced by three separate behaviors, and a number from one of them predicts very little about the next.

So the useful question is not "how visible are we in AI." It is which stage of the channel is failing, because the fix for each one is different.

The Four Stages Every Metric Sits In

Before the metrics, the shape. Everything below answers a question about one of four stages, and they run in order. A page cannot be a citation candidate if it was never read, and it cannot send you a referral if it is never cited.

Stage 01 Training Has a model read this page at all? Bots like GPTBot and ClaudeBot reading on a schedule, feeding what the model knows by default.
Stage 02 Candidates Does it reach for the page mid-conversation? Live fetches by agents like ChatGPT-User while answering someone's actual question.
Stage 03 Referrals Did a real person click through? Humans arriving from an AI answer, landing on a specific page.
Stage 04 Revenue Did any of it turn into money? Orders and leads matched back to the session that produced them.
Stages 01 and 02 are server-to-server requests that never execute JavaScript. A browser-based analytics tool is not undercounting them, it cannot observe them at all.

The Seven Metrics

Each of these is a panel in the WISLR AI Channel Analytics dashboard, read off first-party server logs. They are listed in funnel order, and the diagnostic value is in reading them against each other rather than one at a time.

01

Training coverage

Training
What it measures
The share of your commercially important pages that AI training crawlers have actually read, not the raw number of crawl hits.
How WISLR reports it
Learned pages and training coverage, with daily crawl volume, your top training days, and a breakdown by engine.
What good looks like
Coverage concentrated on category, product and comparison pages. Broad coverage of a blog while the commercial catalog sits unread is a common and expensive pattern.
Where it breaks
Blocks in robots.txt, content that only exists after JavaScript runs, and slow responses that cause crawlers to give up.
02

Candidate coverage

Candidates
What it measures
The share of pages fetched during live conversations. These are the pages a model considered good enough to pull while answering a real question, which makes this the closest thing to a vote of confidence the channel produces.
How WISLR reports it
Candidates and candidate coverage, with the individual candidate pages listed by title so you can see which content is being reached for.
What good looks like
Your money pages appearing as candidates, not just the homepage and one popular guide.
Where it breaks
Pages with no self-contained claim to quote. A model fetching your page still has to find a sentence it can stand behind.
03

Referral coverage

Referrals
What it measures
The share of pages that earned an actual click out of an AI answer, and where those people landed.
How WISLR reports it
Referrals and referral coverage, with landing pages ranked, and referral volume tracked against training and candidate activity on the same timeline.
What good looks like
Referrals concentrated on pages that convert, rather than on informational content that answers the question so completely nobody needs to click.
Where it breaks
Mobile in-app browsers strip the referrer, so the visit lands in direct or other. This is the single largest source of undercounting in browser-based tools.
04

Growth score

All stages
What it measures
Direction and speed rather than level. A single composite read on whether the channel is accelerating, holding, or decaying.
How WISLR reports it
A growth score with a momentum label, alongside a daily activity view so you can see whether a change is a trend or one unusual day.
What good looks like
Momentum that holds after a content push instead of spiking and returning to baseline within two weeks.
Where it breaks
Reading it daily. Crawl behavior is bursty by nature, and a single day is noise.
05

Engine mix

All stages
What it measures
How your visibility distributes across OpenAI, Anthropic, Gemini and Perplexity, at each stage separately.
How WISLR reports it
A breakdown by LLM applied to training, candidates and referrals, so the mix can differ by stage and you can see it.
What good looks like
Presence across several engines. Concentration in one is a dependency, and engines change their crawling and citation behavior without notice.
Where it breaks
Assuming one engine's behavior generalizes. They differ in what they crawl, how often they fetch, and whether they link out at all.
06

Crawled but not cited

Training to candidates
What it measures
Pages a model has read and never fetched into an answer. The gap between what it has and what it uses.
How WISLR reports it
A standing list of pages crawled but not cited, so the gap is a work queue rather than an inference.
What good looks like
A short and shrinking list on commercial pages. On thin or duplicative content, a long list is the correct outcome.
Where it breaks
Treating a crawl as the win. A crawl means the door was open, nothing more.
07

AI-attributed revenue

Revenue
What it measures
Orders and leads traced back to the AI session that produced them, in currency.
How WISLR reports it
IP-to-order attribution with revenue over time, buyer-level detail, and time-to-purchase, split by engine where the data supports it.
What good looks like
Revenue per AI session that stands comparison with your other channels, and a time-to-purchase curve you can plan retargeting around.
Where it breaks
Session continuity. Someone who arrives from an AI answer and returns three days later to buy is invisible to a naive last-click model, which is why time-to-purchase belongs next to the revenue number.

Reading Them Against Each Other

Individually these are just numbers. The diagnostic value is in the steps between them, because each drop has one likely explanation and a different fix.

Training lowThe model has never read the page. This is an access problem: robots rules, rendering, or response times. Nothing downstream is worth looking at until it clears.
Training high, candidates lowThe model has your content and does not reach for it. This is a content problem, usually the absence of a self-contained claim worth quoting.
Candidates high, referrals lowYou are being used and not credited. Some of this is the engine's linking behavior and outside your control; some is that the page answers the question so completely there is no reason to click.
Referrals high, revenue lowThe traffic arrives and does not convert. This is the ordinary conversion problem every channel has, and it is the one your existing tools are already good at.
Or skip the build

All seven are panels in WISLR.ai. It captures training crawls, mid-conversation fetches and AI referrals at the edge on your own domain, matches sessions back to orders, and reports coverage at every stage broken out by engine. No log pipeline to stand up, and the free plan keeps 30 continuous days of history so the baseline starts today.

Start with WISLR.ai →

How This Maps to Your SEO Report

These metrics sit alongside your SEO report, not on top of it. They come from different sources and answer different questions, and the closest analogue is often not an exact one.

Familiar SEO metricAI channel equivalentWhy it is not the same
Index coverageTraining coverageBeing indexed is a yes or no. Being read by a model is per engine, on each engine's own schedule.
ImpressionsCandidate coverageThere is no impression count to report. The nearest observable event is the model fetching your page to answer a question.
ClicksReferral coverageSame idea, far worse instrumentation. Referrers are stripped often enough that browser tools undercount by 2.5x to 5x.
Average positionGrowth scoreThere is no position. Direction over time is the only stable read available.
Revenue by channelAI-attributed revenueThe same question, answered from server logs and order records instead of a tag that often never fires.

Where to Start

Start at the top of the funnel, because a problem at stage one makes every number under it meaningless. Training coverage and crawled but not cited are enough to make the first round of decisions, and both come from the same log data.

The one thing worth doing immediately is starting the clock. Every metric here is a trend, and a trend needs a baseline. A brand that begins recording now has months of history when a competitor is still deciding what to measure, and history is the part you cannot backfill.

Work with WISLR

See these seven numbers on your own data.

WISLR AI Channel Analytics reports all seven from your own first-party server logs: training and candidate coverage by engine, the crawled-but-not-cited queue, growth score, and revenue matched back to the session. The free plan keeps 30 continuous days of history, which is enough to establish the baseline.