AI Summary
AI search has no rankings, and it does not even give the same answer twice. In our public tracker demo, the same prompt swung from 57 to 89 visibility on Copilot between two runs, and from 63 to 93 on Perplexity. Single-query checks are noise. These are the metrics that replace ranking position, and how to collect them.
Why rank tracking breaks in AI search
Three structural reasons. Generative engines are probabilistic, so identical prompts produce different answers run to run. There is no position one, only presence or absence inside a synthesized answer. And each engine runs its own retrieval, so visibility on one says nothing about another. The fix is statistical: measure rates across a stable prompt set and multiple runs, not single observations.
The volatility is not a bug. It reflects how generative engines actually work: sampling temperature in the LLM plus shifting retrieval results means the same query draws from a slightly different pool each time. A brand that appears in 70% of runs has genuinely strong retrieval coverage; a brand that appears in 20% of runs has a foot in the door but no reliable presence. The difference is invisible in a single manual check. This is why a measurement framework built on rates, not spots, is the prerequisite for any meaningful optimization work.
The four metrics that matter
1. Visibility rate
The share of prompt runs where your brand appears, per engine, per month. This is your headline number. Because answers vary, it only stabilizes across repeated runs of a stable prompt set; our demo workspace averages it across 60 runs. Report it per engine: a 60% visibility rate on Perplexity and 15% on Gemini are very different problems requiring different fixes. Visibility rate is the top-line KPI you track monthly and report to leadership. Everything else is diagnostic beneath it. Our GEO KPI framework covers how to set targets by prompt intent and funnel stage.
2. Citation share and source mix
Which of your pages get cited, and which third-party sources power your category’s answers. In our live CRM-category audit, two listicle publishers shaped most recommendations. The source list is your off-site roadmap. In our May 2026 analysis of 153,425 citations across 6 AI platforms, YouTube absorbed 9,868 citations and Reddit absorbed 6,595. If your category mirrors this pattern, a presence strategy for those channels is not optional. Track which of your own pages get cited most: those are your highest-value assets for internal linking and content updates. Pages that get cited on several engines are worth protecting; pages that never appear despite strong organic ranking need structural rework.
3. Competitive share of voice
Your mention rate against named competitors on the same prompts. Visibility without context flatters; share of voice tells you whether you are winning the answers that matter. A 40% visibility rate sounds good until you learn the category leader runs 85% on the same prompts. Run share-of-voice calculations on decision-stage queries specifically: “best [category] tool for [use case]” prompts reveal competitive standing better than awareness queries. Share of voice has replaced search volume as the primary content metric for AI search because it captures relative competitive standing rather than absolute traffic potential. Track it monthly on a fixed competitor set so trends are comparable.
4. Sentiment and position
How you are framed (recommended, mentioned, cautioned) and where in the answer you appear. Our 42,971-citation analysis found engines favor early, declarative placements, so a buried neutral mention and a lead recommendation are different outcomes worth separating. Sentiment labels should be: recommended (explicit endorsement), neutral mention (named without judgment), cautioned (named with caveat or warning), and absent. A brand moving from cautioned to neutral to recommended across six months has achieved something that visibility rate alone does not capture. Position matters because AI answers often have a primary recommendation and then secondary alternatives: being named first carries different conversion weight than being listed third.
The measurement setup: prompt set and cadence
The prompt set is the foundation. Build it once and never change it between periods, or your trend data is meaningless. A working prompt set for B2B SaaS typically has 30-50 queries spread across three intent tiers: awareness (“what is [category]”), consideration (“best [category] for [use case]”), and decision (“compare [brand] vs [competitor]”). The consideration tier drives the most actionable data. Cover all six major engines: AI Mode, Gemini, ChatGPT, Perplexity, Copilot, and Grok. Our prompt research guide covers how to build this set from scratch using buyer language rather than keyword planner output.
Cadence rules
Monthly reporting is the minimum for meaningful trend detection. Weekly runs are better for catching algorithm shifts early. Never compare a single run to a single run: compare the average of 10+ runs in period A to 10+ runs in period B. Keep the prompt set identical for at least a quarter before making additions. When you add prompts, keep the old ones running in parallel for one period to establish a continuity bridge. The citation velocity framework adds a leading indicator: how fast new content gets picked up, which signals retrieval freshness before it shows in monthly visibility rate.
The metrics at a glance
| Metric | What it measures | Reporting cadence | Primary use |
|---|---|---|---|
| Visibility rate | % of runs brand appears, per engine | Monthly | Top-line KPI |
| Citation share | Which pages and third-party sources are cited | Monthly | Content and off-site roadmap |
| Share of voice | Your visibility vs competitors on same prompts | Monthly | Competitive standing |
| Sentiment and position | How and where brand appears in answers | Monthly | Brand health and narrative |
| Citation velocity | Speed of new content pickup | Weekly | Leading indicator |
Build or buy the tracking loop
DIY is real: the three free methods plus our open-source GEO/AEO Tracker (BYOK, six engines, scheduled runs) cover everything above. The tracker runs on GitHub at github.com/danishashko/geo-aeo-tracker with bring-your-own API keys so your data never leaves your stack. Pair it with GA4 referral attribution so visibility connects to sessions. The Looker Studio dashboard guide shows how to wire the two together into a single weekly report.
When scale or competitor depth outgrows DIY, the managed tracking service handles prompt set design, multi-engine runs, and monthly reporting with competitive share-of-voice included. Start either way with the audit checklist and keep the prompt set stable so months are comparable. How to find those prompts is covered in prompt research vs keyword research. For agencies running multiple clients, the competitive intelligence tools roundup covers platforms that support white-label reporting at scale.
From metrics to action
Metrics without action loops are vanity. Each metric should drive a specific work queue. Low visibility rate triggers a content audit and structural rework. Low citation share on your own pages triggers a citation-chain review of your best pages. Low share of voice relative to competitors triggers competitive content gap analysis on the prompts where they outperform you. Negative sentiment triggers a hallucination defense review and a push to get accurate brand descriptions into the sources engines cite. The GEO content audit framework maps all four metrics to specific remediation tasks so the monthly report becomes a work order, not just a dashboard.
Prefer to automate this measurement? These are the GEO tools that track these metrics for you.