10 AI visibility KPIs every marketing team should track

The most useful AI visibility KPIs are eligible mention rate, citation rate, qualified recommendation rate, share of voice, sentiment, factual accuracy, referral sessions, engaged visits, assisted conversions and influenced pipeline. Do not compress them into one unexplained score. The first six describe what assistants say; the last four describe what identifiable visitors do. Track every rate with its numerator, denominator, prompt set, provider mix, market and collection date. That separation lets a marketing team see whether it has a discovery problem, an evidence problem, a reputation problem or a downstream conversion problem.

Measure presence before popularity

Eligible mention rate is the share of answers in which the brand appears among prompts where it could genuinely satisfy the stated requirements. A mention means the name appears; it does not imply endorsement. Citation rate is the share of answers that visibly link to or name a source from your domain. Keep source citations separate from mentions because an assistant can mention a brand without linking to it, or cite a page while recommending another option. Define eligibility before running prompts so irrelevant questions do not depress the denominator.

• Mention rate: answers naming the brand divided by eligible answers.
• Citation rate: answers citing an owned URL divided by answers with visible source treatment.
• Qualified recommendation rate: answers recommending the brand while matching the buyer's stated constraints.
• Share of voice: your eligible appearances divided by all eligible brand appearances in a fixed competitor set.
• Sentiment: favorable, neutral, mixed or unfavorable framing, reviewed in context.
• Accuracy: material claims judged correct, incomplete, outdated, unsupported or wrong against approved evidence.

Define recommendations, sentiment and accuracy narrowly

A recommendation requires more than a mention: the answer must position the business as a suitable choice, ideally with a reason tied to the prompt. Conditional recommendations count separately when the stated condition matters. Sentiment describes framing, not truth. A warmly written answer can contain a material error, while a critical answer can be accurate. Accuracy therefore needs its own claim-level review against current product, policy, pricing and company records. Report material-error rate beside sentiment instead of letting favorable language hide customer risk.

Connect answer metrics to behavior carefully

Referral sessions are visits whose referrer or campaign evidence indicates an AI assistant. Engaged visits add a quality test such as meaningful time, depth or a key event. Assisted conversions are conversions where an identifiable AI referral occurred earlier in the measured journey but was not necessarily the final touch. Influenced pipeline is qualified opportunity value associated with a documented AI touch or buyer disclosure under an agreed attribution rule. None of these proves the assistant alone caused the outcome. Preserve direct, organic and unknown classifications rather than forcing every "dark" visit into AI.

1. Freeze a commercially relevant core prompt panel by audience, market and buying stage.
2. Capture complete answers across ChatGPT, Claude, Gemini and Perplexity under matched conditions.
3. Apply written labels for mentions, citations, recommendations, sentiment and accuracy.
4. Configure GA4 events, referrer groupings and CRM fields before interpreting revenue.
5. Report sample counts, changes and uncertainty instead of an isolated percentage.
6. Open the underlying answer or journey whenever a KPI crosses a decision threshold.

Build a KPI scorecard that creates action

Give each KPI an owner, cadence and response. Content can own citation gaps, product marketing can own inaccurate positioning, communications can review harmful framing, analytics can maintain referral rules, and revenue operations can audit pipeline classifications. Start with the AI visibility scan and use the AI readiness checker to inspect the public evidence behind weak answers. ModelSaid Monitor and higher plans add recurring trend monitoring; their Google connection can connect either GSC or GA4, so check current pricing and plan limits before designing the dashboard. A disciplined scorecard shows separate evidence layers and tells an owner what to inspect next; it never turns correlation into guaranteed ranking or revenue.

Review the KPI system, not just the latest values

Review the scorecard monthly and after material launches, but do not reset targets merely because one provider changed. Keep the original observations, mark methodology breaks and explain whether movement was broad or concentrated in a few prompts. A useful quarterly review asks which inaccurate claims were resolved, which cited pages need investment, which recommendations fit the intended buyer, and whether identifiable demand signals support continued work. Retire KPIs that never change a decision. Recalibrate human labels on a shared sample so a new reviewer does not create an artificial sentiment or accuracy trend. Also audit the competitor universe and prompt eligibility, versioning any necessary change. The result should be a small, durable measurement system rather than a growing inventory of numbers.