Here is the awkward question every marketing leader investing in AI search eventually gets asked: “how do we know it's working?” The old answers, rankings went up, sessions went up, don't survive contact with 2026. Rankings don't capture whether ChatGPT names you. Sessions are falling across the board because AI answers the question without the click. If you measure AEO with an SEO scoreboard, you will conclude it is failing exactly when it is working.
This is the measurement system we run for clients, end to end: what to track, how to track it without fooling yourself, and how to connect it to the number the board actually cares about.
8% vs 15%
click rate with vs without an AI summary (Pew)
0%
of questions where 4 engines agreed on a top source (CiteTrack)
+38%
more likely to buy: visitors arriving from AI tools (Adobe)
First, accept what changed about the funnel
Pew Research found users click a result on just 8% of visits when an AI summary appears, versus 15% without one. The influence is real, buyers act on the answer, but the click that used to prove it evaporates. Meanwhile the AI-referred visitors who do click convert unusually well: Adobe measured them 38% more likely to buy. So the funnel now has a layer above your website, and most measurement stacks are blind to it.
The core metric: citation share, measured as a distribution
The AEO equivalent of a ranking is citation share: across the buying prompts that decide your category, what percentage of AI answers cite or mention you, per engine, over time. Three design decisions make it trustworthy instead of theater.
- Revenue-weighted prompts, not vanity prompts. Build the prompt set from 12–18 months of closed-won data and real sales conversations, the questions that map to deals, then add 20–30 conversational variations of each, because models answer different phrasings differently.
- Sampling, not counting. Ask “best X” ten times and the answer drifts. One run is noise; a screenshot is marketing. Run each prompt several times per engine and report the distribution, closer to polling than rank-checking.
- Every engine, separately. In our 101-question benchmark, the four major engines agreed on the top source 0% of the time, and on 70% of questions no two even shared one. A single-engine view is a quarter of the picture at best.
Grade each sample on three levels: cited (you are a source the answer links or attributes), mentioned (named but not sourced), or absent. Add sentiment and position where you can. The trend of that distribution is your AI-search ranking report.
The supporting metrics, layer by layer
| Layer | Metric | How to get it | What it tells you |
|---|---|---|---|
| Answers | Citation share & share of voice vs competitors | Sampled prompt runs across ChatGPT, Perplexity, Gemini, Claude | Whether you are winning the answer layer |
| Sources | Source coverage: presence on the domains engines cite for your prompts | Record the citations behind every sampled answer | Which third-party gaps to close next |
| Crawl | AI crawler hits (GPTBot, PerplexityBot, ClaudeBot, Google-Extended) | Server logs or an edge/CDN report, verified by reverse-DNS | Whether engines can even read you |
| Referrals | AI-assistant sessions & conversion rate | GA4's AI channel grouping plus custom referrer rules | The visible slice of AI-driven demand |
| Demand | Branded search & share of search | Search Console + free trends data | The proxy that rises when invisible influence rises |
| Revenue | Self-reported attribution + blended efficiency | “How did you hear about us?” + pipeline ÷ total spend | Whether any of it becomes money |
Two notes on the referral layer, because it is where teams most often fool themselves. GA4 added an AI-assistant channel grouping in 2025, but it only catches visits that pass a referrer, in-app clicks and copy-pasted links still land in Direct, so treat GA4's AI number as a floor, not the truth. And expect the crawl-to-referral ratio to look absurd: engines read hundreds of pages for every visitor they send. That is not failure, that is the medium. The influence happens inside the answer, which is why the citation layer, not the session layer, is the headline. We covered the underlying mechanics in the dark funnel.
If you measure AI search by sessions, you will kill the program right as it starts working. Measure the answer layer, and let sessions be a bonus.
Leading vs lagging: what to report to the board
AI search stretches the gap between cause and effect, so separate your indicators. Leading: citation share trend, source coverage, branded-search lift, AI crawler activity. Lagging: AI-referral conversions, self-reported “heard about you from ChatGPT,” pipeline and closed-won. With a 90-day sales cycle, the leading indicators are the only forecast of next quarter you will get. Put both tiers on one page, headline the blended-efficiency number no platform can inflate, and add a one-paragraph “what we can and can't see” note. Executives forgive imperfect attribution; they don't forgive discovering the gaps themselves.
The honest caveats
Citation share is a sampled estimate, engines change weekly, and no tool sees inside a private ChatGPT conversation. Self-reported attribution is directional and biased toward what people remember. None of that is a reason to skip measurement, it is the reason to triangulate: answers + proxies + reported, with an incrementality test when signals disagree. Treat every number as evidence, not verdict, and you will still be far ahead of competitors reporting a rankings screenshot.
One more honesty note: this is exactly why we built our own tracking rather than reselling a dashboard, sampling across engines, validating the cited sources (roughly a third are dead or hallucinated), and joining it to revenue data. The difference between a tool and results is covered in Enginekick vs AI-visibility tools, and the full method is The Citation Engine.
Set it up this month
- Week 1: build the revenue-weighted prompt set (20–50 prompts + variations) from closed-won data.
- Week 2: baseline citation share: run every prompt across all four engines, several samples each, and record cited / mentioned / absent plus every source cited.
- Week 3: wire the plumbing: GA4 AI channel grouping, AI-crawler log tracking, a required “how did you hear about us?” field.
- Week 4: ship the first one-page report: citation share trend + source gaps + branded search + blended efficiency, with the transparency note.
Or start with a free read: our AI visibility checker scores your readiness in two minutes, and on a strategy call we run your Answer Map, your prompts, every engine, live.
Frequently asked
What is citation share?
Citation share is the percentage of AI answers, across the buying prompts that matter in your category, that cite or mention your brand, measured per engine and over time. Because AI answers vary run to run, it is computed from repeated samples as a distribution, like polling, rather than from a single response.
What is the AEO equivalent of keyword rankings?
Citation share, tracked across ChatGPT, Perplexity, Gemini and Claude on a revenue-weighted prompt set. Where SEO asks "what position do we hold for this keyword," AEO asks "in what share of sampled answers to this buying question are we cited, mentioned or absent," and reports the trend.
How do I track ChatGPT traffic in GA4?
Use GA4's AI-assistant channel grouping plus custom rules for referrers like chatgpt.com and perplexity.ai. But treat it as a floor: in-app clicks and copy-pasted links often pass no referrer and land in Direct, and most AI influence never produces a click at all. Pair it with citation-share tracking and a self-reported attribution field.
How do I prove AEO drives revenue?
Triangulate three signals: citation-share growth on revenue-weighted prompts (leading), self-reported attribution naming AI tools on your forms (reported), and blended pipeline-versus-spend efficiency (lagging). When the three move together, the case is strong; when they disagree, run an incrementality test rather than an argument.
How often should I measure AI visibility?
Weekly sampling with a monthly reported trend is the practical cadence. Engines update constantly, so daily numbers are noisy; quarterly is too slow to catch a competitor taking your answers. What matters is consistent methodology, same prompts, same engines, same sampling depth, so the trend is real.
Can I measure AEO with my normal SEO tools?
Only partially. Rank trackers and backlink indexes do not sample AI answers. Some suites now bolt on AI-visibility modules, and dedicated trackers measure it directly, but the layer that matters, sampled citation share on your revenue prompts plus validation of the sources behind each answer, usually requires purpose-built tracking. That is why we built our own.