Which AI visibility metrics should a CMO track?

Map AI marketing KPIs to the new customer journey

The buyer no longer starts on a search results page. They ask an assistant. They read a shortlist. They form a view before they ever reach your site. Six metric categories track that journey, one per stage. A report that opens with referral sessions has skipped the first two.

The old funnel let you watch every step. The new one does not. The first two stages happen inside someone else's app, which is why visible AI clicks are about 1.08% of web traffic while the influence behind them is much larger.

The evaluation moved too, and that is the part most teams miss. In the search era a buyer opened your page, a competitor's page and a review site, then compared them across tabs. Your pricing page did the work. Your comparison page did the work. Now the assistant does that work. It reads all three sources, weighs them, and hands back one shortlist with the comparison already made.

So the pages you built to persuade are read by a model before a human sees them, and often instead of a human seeing them. That is why being cited in an answer is now the thing you optimise for. You cannot rank in an answer. You are either in it or you are not, so the first metric is how often you appear.

Share of voice in AI search replaces the old rank report

Position fails as a stable measurement, so any dashboard reporting it is selling noise. SparkToro and Gumshoe.ai ran 2,961 prompts with 600 volunteers and found the odds of the same brand list appearing twice were under 1 in 100. The same list in the same order came in near 1 in 1,000.

AI referral traffic misses most of your visits

Every number here is a floor, so say so on the slide. The real figure is higher and you cannot see it. Similarweb found 55.9% of AI-influenced visits arrive as branded search rather than a trackable referral, so the visit stage structurally under-reports.

Be honest about the AI search conversion rate evidence

Published multiples range from no significant difference to 23×, and the most disciplined study sits at the bottom. Amsive analysed 54 sites over six months: LLM referrals converted at 4.87% against organic at 4.60%, with a paired t-test returning p = 0.794. The B2B subset ran 2.03% against 1.68%, also not significant.

Plan on 4× to 10× for B2B, hold it loosely, and say where each figure came from. The range comes from differences in sample size, vertical and what counts as a conversion: a software trial start is not a retail purchase.

Pipeline attribution is what wins you more marketing budget

Every earlier stage points at this one. Report AI-sourced and AI-influenced pipeline created, win rate, average deal size and cycle length, each split by first-touch source.

The mechanics of joining a source to a closed deal are in revenue attribution.

Revenue attribution metrics count paying customers

Define revenue before you measure it. Here it means a new customer on a paid plan. That is the moment a card is charged for the first time. A trial start does not count. Nor does an MQL, or pipeline created.

Both motions need attribution, and they fail in different ways. In a product-led, self-serve motion the whole chain is observable: AI answer, signup, activation, first charge. It can close in days, so the origin flag rarely has time to get lost.

Sales-led is harder. The same buyer takes months, several people from the account get involved, and the person who signs the order is almost never the person who asked the assistant. Roll touches up to the account level before assigning credit, otherwise the channel that created the demand looks like it contributed nothing.

At current volumes most companies cannot produce a statistically significant per-citation revenue figure. Say that in the first meeting rather than being forced into it later. What you can defend is the AI-sourced customer count and the directional trend.

AI visibility tracking tells you what is coming next

Branded search predicts revenue better than clicks do. Similarweb found AI-recommended brands were 2.5× more likely to get a site visit within seven days, with 55.9% arriving as branded search. Ahrefs' correlation work put branded web mentions against AI visibility at 0.50 to 0.74, while backlinks and ad spend sat below 0.30.

Track citation share weekly, then test correlation with lag against branded search volume in Search Console, direct traffic in GA4, and net-new pipeline. When citation share rises and branded search follows within a sales cycle, you have a defensible story. When it rises and nothing follows, the citations are low quality, and that diagnosis is worth more than a precise number.

Attribution reporting the board will believe

One slide. Six stages. Show movement rather than scores. Tie every metric back to pipeline. State the method behind each figure and show the gap between self-reported and tracked attribution rather than resolving it.

• Citation share and mention rate against named competitors, with confidence intervals.
• Mention sentiment and accuracy, because being named badly is worse than not being named.
• AI referral sessions, labelled as a floor every time.
• Branded search and direct trend, which is where the invisible influence lands.
• AI-sourced and AI-influenced pipeline, with the reconciliation gap named.

What a CMO dashboard for AI search looks like

The clearest public template for this is Kevin Indig's AI visibility ladder, published in Kyle Poyar's Growth Unhinged in July 2026. Indig uses it with Airbnb, Asana and Xero, and it borrows its structure from Andy Grove's paired leading and lagging indicators in High Output Management. The logic is simple: when revenue attribution lags, you want to know whether you are on track long before the revenue arrives.

His ladder runs leading indicators, then quality guardrails, then lagging indicators, mirroring a Retrieved, Cited, Trusted progression. Three setup rules make the dashboard trustworthy:

• Freeze 20 to 50 high-intent prompts across personas, use cases and buying stages, for at least four weeks, so you measure real change rather than prompt drift.
• Log every run: prompt, model, location, answer, cited URLs, brands mentioned and shortlist position. That table is the raw material every rung reads from.
• Run two clocks. Weekly the team checks signal quality. Monthly the CMO checks allocation.

Indig names three traps worth quoting directly: vanity metrics treated as the destination, false precision from decimals on a number that re-rolls monthly, and mixing leading indicators with outcomes without a model connecting them. His framing of last-click AEO measurement is the sharpest in the literature: valuing it by referral clicks is like valuing a Super Bowl ad by QR-code scans.

Every row needs a decision and an instrument

A number that triggers no decision, or that you have no way to move, is the first trap on that list. The scorecard above is only worth presenting if each row, on every rung, comes with two more columns: what you do when it moves, and what has to be instrumented before you can move it. Without both, the dashboard is a weather report.

Read the table as a build list. Any row where the instrumentation column is not yet true is a row you should leave off the scorecard entirely. A dashboard with three populated, actionable rows outperforms one with twelve uninstrumented scores.

What to ask AI visibility tools before you buy one

Most of the category measures presence in answers and stops at the citation. Three questions separate a dashboard from an attribution system.

• Can you show me a closed-won deal and the AI answer that started it? Almost nobody in the category can answer this.
• How many prompts, runs and engines does my tier cover? Most price on tracked prompt volume in tiers of 15, 100 or 500. Given the variance above, a small set run once is meaningless, so entry tiers sell an underpowered sample.
• Do you report ranking position? If yes, they have not read the research on non-determinism.

How ThriveStack citedby differs from other AEO tools at every stage

This metric set is only useful if the last stage is populated, and the last stage is paid customers.

citedby is the AI visibility product inside ThriveStack. It samples answers across roughly ten engines and tracks which of your pages get cited. It then carries that origin into your CRM and billing data, so a citation can be traced to a closed deal. The loop it runs is analyze, fix, attribute.

• Citation share, mention rate and co-mention on a frozen prompt set across roughly ten engines, reported as movement with confidence intervals.
• Sentiment and factual accuracy of each mention, so quality sits beside volume in one view.
• AI referral capture covering the engines GA4's native channel omits, including Perplexity and Claude.
• Branded search and direct correlated against citation share with lag, recovering the influence that lands elsewhere.
• AI-sourced pipeline and closed-won revenue through CRM and billing integration, which is the stage that renews budget.