AI Search KPIs: 7 Measurable Definitions

A useful KPI names its numerator, denominator, engine set, prompt set, and time window. This framework separates technical access, answer-level visibility, referrals, and business outcomes so one signal cannot masquerade as another.

The short answer

Measure seven layers: mention rate, citation rate, recommendation rate and framing, cited-page coverage, verified crawler coverage, AI Readiness, and AI-referred conversion rate. Keep the monitored questions and engines fixed. Report each engine separately before calculating a rollup. Use your own baseline for targets because no public study establishes a universal good citation rate or sentiment score for every market.

This separation matters because access is not indexing, indexing is not citation, citation is not recommendation, and a recommendation is not a conversion. A single composite score can summarize the system, but the underlying rates explain what changed.

Use one measurement contract

Before collecting results, freeze the contract below. A changed prompt set can create apparent growth even when the brand's performance did not change.

Field	Record it as	Why it matters
Prompt panel	Versioned list of buyer questions	Prevents easier questions from inflating the trend
Engine set	Named engines and product surfaces	Different engines retrieve and cite different sources
Market context	Language, country, device, and account state	Locality and personalization can change an answer
Successful answer	Non-error response with usable answer text	Provider errors should not count as brand misses
Observation window	Start, end, cadence, and repeat count	Makes percentages reproducible

Foglift's frozen 75-question research panel found 77.0% to 86.7% mean same-question citation-set turnover across five engine lanes from Q2 to Q3 2026. That result supports repeated measurement on a stable panel. It does not prove that every brand changes at that rate. Read the full citation-drift methodology.

The seven AI search KPIs

1. Brand mention rate

Formula: successful monitored answers that name the brand divided by all successful monitored answers, multiplied by 100.

Report: the total and the split by engine, prompt group, and buyer intent. Add the denominator beside every percentage.

Target rule: compare with your previous fixed-panel window and with the named competitors on that same panel. Do not import a generic industry percentage.

2. Domain citation rate

Formula: successful answers containing at least one citation to your normalized domain divided by all successful answers, multiplied by 100.

Keep separate: count an answer once for citation rate, then report citation occurrences and distinct cited URLs as supporting measures. Otherwise one answer with five links can look like five wins.

Decision: a rising mention rate with a flat citation rate calls for stronger source evidence. A rising citation rate with a flat mention rate can mean your pages support answers without making the brand shortlist.

Worked method: Foglift's AI search tool citation benchmark publishes its usable-answer denominator and separately defines answer-level tool share of voice and citation URL occurrences. That category panel illustrates denominator discipline; its percentages are not universal targets.

3. Recommendation rate and framing

Formula: answers that explicitly recommend or shortlist the brand divided by answers that mention it. Classify the remaining mentions as neutral, caveated, or warning.

Quality control: keep the answer excerpt that justifies the label and audit a sample of automated classifications. A brand list and a direct endorsement are different outcomes.

Target rule: improve the mix from your own baseline. A universal 70% positive target is not defensible without a shared labeling rubric, prompt mix, and category sample.

4. Cited-page coverage and concentration

Coverage formula: eligible pages cited at least once divided by all eligible pages in the monitored content set.

Concentration formula: citations earned by the top cited pages divided by all citations to your domain. Always publish whether “top” means one, five, or ten pages.

Decision: concentration is diagnostic, not automatically bad. Protect pages that already earn citations, then compare uncited pages with the winning sources for the same question.

5. Verified AI crawler coverage

Formula: eligible URLs with at least one verified visit from a named search or user-retrieval crawler divided by eligible URLs, reported by agent and time window.

Classify purpose first: OAI-SearchBot supports ChatGPT search while GPTBot is a separate training control. Anthropic distinguishes Claude-SearchBot, Claude-User, and ClaudeBot. Perplexity distinguishes PerplexityBot from Perplexity-User.

Boundary: a verified request proves access to a URL. It does not prove indexing, citation, recommendation, or training. Use provider documentation for OpenAI, Anthropic, and Perplexity.

6. AI Readiness Score

What it measures: page-level technical and authority inputs that help systems access, understand, and reuse content. It remains separate from answer-level AI Visibility.

How to use it: record the score before a page change, fix the highest-severity findings, deploy, and rescan. Then watch the fixed prompt panel for answer-level movement instead of assuming the readiness change caused a citation.

Evidence: in 854 Foglift Technical Audits, the latest eligible result for each of 194 domains had a median AI Readiness Score of 60 and median SEO of 90. The study describes that sample and scoring window. It does not set a universal pass mark. See AI Readiness across 194 websites.

7. AI-referred conversion rate

Formula: key events from observed AI-referred sessions divided by observed AI-referred sessions, multiplied by 100. Report session count and key-event count beside the rate.

Attribution: GA4's AI Assistant default channel covers sources such as ChatGPT, Gemini, DeepSeek, Copilot, and Grok. It excludes Google AI Overviews and AI Mode. Google reports its generative Search features through Search Console, including a dedicated Generative AI performance report.

Boundary: referrals without preserved source data can appear as direct traffic. Treat analytics as observed click-through contribution. Use assisted pipeline or survey evidence separately and label the attribution model.

A scorecard with formulas and decisions

KPI	Required denominator	Primary decision	Target
Mention rate	Successful answers	Which prompt or engine gap to address	Own baseline and same-panel competitors
Citation rate	Successful answers	Where stronger source evidence is needed	Own baseline by engine and intent
Recommendation rate	Brand-mentioned answers	Positioning or reputation work	Improve labeled mix from baseline
Cited-page coverage	Eligible content set	Protect winners or strengthen gaps	Content-set baseline
Crawler coverage	Eligible URLs	Robots, CDN, WAF, or rendering fix	Expected allowed agents and paths
AI Readiness	Audited pages	Which page-level issue to fix	Pre-change score and issue closure
Conversion rate	Observed AI-referred sessions	Landing-page and funnel investment	Own channel baseline with sample shown

Choose cadence from the decision

Decision	Useful cadence	Comparison rule
Active content or positioning experiment	Weekly fixed panel	Compare multiple post-change observations with the pre-change window
Operational health	Daily or weekly alerts	Alert on defined deltas and provider failures separately
Executive reporting	Monthly rollup	Show denominator, prior period, and intervention log
Source-market change	Quarterly frozen-panel study	Keep questions and aggregation method unchanged

The KDD 2024 Generative Engine Optimization paper measured experimental content changes with its own visibility metrics and benchmark. It did not establish universal brand-monitoring targets. Use research metrics inside their published method, then define business KPIs around your own decision and denominator.

Monthly reporting template

Contract: prompt-panel version, engines, market context, cadence, successful-answer count.
Outcomes: mention rate, citation rate, recommendation mix, each split by engine and buyer intent.
Sources: cited-page coverage, top cited URLs, and source concentration with the cutoff named.
Access and readiness: verified crawler coverage, Technical Audit changes, and unresolved issues.
Business impact: observed AI-referred sessions, key events, revenue, and the attribution boundary.
Interventions: dated page changes, earned mentions, releases, and the next measurement date.

How Foglift measures the full loop

Foglift keeps the layers separate. Unlimited single-page Technical Audits are available at $0 and report SEO, AI Readiness, performance, security, and accessibility. Active Free workspaces can run weekly Perplexity monitoring. Launch starts at $49 per month and adds daily monitoring across ChatGPT, Claude, Perplexity, Gemini, and Google AI Overview, plus REST API, CLI, and MCP access.

AI Crawler Analytics and referral tracking add access and traffic evidence. Tracked prompts, citations, sentiment, competitors, recommendations, and content briefs connect the answer layer to an improvement queue. That combination lets a team diagnose a weak page, observe whether the brand appears, ship a change, and measure the next fixed-panel window without collapsing every signal into one number.

Frequently asked questions

What are the most useful AI search KPIs?

Start with mention rate, citation rate, recommendation rate and framing, cited-page coverage, verified crawler coverage, AI Readiness, and AI-referred conversion rate. Keep the prompt set, engine set, geography, and cadence fixed so changes are comparable.

What is the difference between mention rate and citation rate?

Mention rate is the percentage of successful monitored answers that name your brand. Citation rate is the percentage that link to your domain. An answer can name a brand without citing it, or cite a brand page without making a recommendation, so report the metrics separately.

What is a good AI visibility benchmark?

There is no defensible universal percentage for every category. Build a baseline from the same buyer questions across the same engines, then compare your own trend, competitors on that panel, and changes after dated interventions. Publish the denominator with every percentage.

How often should AI search KPIs be measured?

Match cadence to the decision. Use a fixed weekly panel when a team is actively shipping changes and a monthly rollup for planning. Keep a longer quarterly comparison because cited sources can move substantially over time. Avoid treating a single answer as a trend.

Can AI search conversions be measured in GA4?

Yes, with boundaries. GA4 can group some referrals into its AI Assistant channel and report session source or medium, key events, and revenue. Google AI Overviews and AI Mode are excluded from that GA4 channel and have dedicated Search Console reporting. Missing referrer data means analytics should be treated as an observed lower bound, not complete influence attribution.

Sources and further reading

Google Search Central, AI features and your website. Defines Google AI-feature eligibility, Search Console reporting, and the difference between access and serving.
Google Search Central, Search Generative AI performance reports. Documents the dedicated Search Console views for generative Search features.
Google Analytics, default channel groups. Defines the AI Assistant channel and its exclusions.
Google Analytics, Traffic acquisition report. Defines session source, medium, key events, and revenue dimensions used for referral analysis.
OpenAI, Publishers and Developers FAQ. Separates OAI-SearchBot from GPTBot and documents ChatGPT referral tracking.
Anthropic, web crawler controls. Defines ClaudeBot, Claude-User, and Claude-SearchBot roles.
Perplexity crawler documentation. Defines PerplexityBot, Perplexity-User, and published verification data.
Aggarwal et al., Generative Engine Optimization, KDD 2024. Reports an experimental evaluation framework for generative-engine visibility; its metrics are bounded to that study.
Foglift Research, AI Engine Citation Drift 2026. Frozen 75-question Q2 to Q3 comparison across five engine lanes.
Foglift Research, AI Readiness Across 194 Websites. Privacy-safe aggregation of 854 Technical Audits using the current score contract.

Start with a measured baseline

Audit one page for free, then monitor the buyer questions that should surface your brand.

Run a free Technical Audit Track AI Visibility

Fundamentals: Learn about GEO (Generative Engine Optimization) and AEO (Answer Engine Optimization) (the two frameworks for optimizing your content for AI search engines).