Rank is not citation
Almost every measurement habit you carried over from SEO points at the wrong thing. Rank tells you where a page sits on a results list. Traffic tells you how many people clicked. Neither tells you whether an AI engine said your name when a buyer asked who is best, and that is the only outcome Answer Authority is built to produce. So the scoreboard has to change. The metric that matters is citation share: how often each engine names you, per query, over time. If you keep grading yourself on rank while the decision has moved into the answer, you will feel busy and stay invisible.
This is not a fringe claim. The foundational GEO research, a 2023 study from Princeton, IIT Delhi, and Georgia Tech, first showed that content and citation choices measurably change how generative engines quote a source. Once you accept that the answer is engineerable, the follow-on question is obvious: how do you know it is working? You measure the citation, not the ranking.
Baseline before you touch anything
The first move in any honest program is to capture the current AI answers for your category before a single change ships. Ask each engine the questions your buyers ask, and record verbatim what comes back: who gets named, in what order, with which sources cited. That record is your starting line. If you cannot show where you stood on day one, you cannot later claim you moved anything, and you should not pretend to. A baseline is also humbling in a useful way. It usually shows you are absent from answers you assumed you owned, which is exactly the gap the rest of the method closes.
Build a prompt library
You cannot compare month to month unless you ask the same questions the same way every time. So you build a prompt library: a fixed set of real buyer questions, organized by persona and by platform, run consistently on a schedule. Vague, one-off prompting produces noise you cannot trend. A disciplined library produces a signal you can.
A working library has three kinds of prompts. Core prompts are the high-intent questions that decide deals, run every cycle without change. Experimental prompts test new phrasings, new personas, and adjacent categories you might expand into. Monitoring prompts watch competitor movement and category shifts so you see a threat before it costs you a citation. Keep the core set stable, because stability is what makes the comparison honest.
Track by engine, not by blended average
ChatGPT, Perplexity, and Gemini or Google's AI Overviews do not source the same way, so a single blended number hides more than it reveals. One engine may name you consistently while another never does, and a blended average paints that as mediocre-everywhere when the truth is strong-here, absent-there, which points at completely different work. Report per engine. Measure ChatGPT, Perplexity, and Gemini and AI Overviews separately, track each on its own line, and read them as three different rooms you are trying to get invited into.
Competitive share of voice
A citation count on its own is thin. What makes it meaningful is context: who else shows up in the same answers, and how you stack against them. So you measure share of voice against the named competitors who appear alongside you, and you track that relationship over time. Progress that is relative is honest progress. Going from named in two of ten answers to five of ten while a competitor slips from eight to five is a real, defensible win, and it is the kind of movement a buyer and a boardroom both understand.
Leading indicators: watch the crawlers
Citations are a lagging signal. They move after the engines re-read the web and re-weight their sources, which takes time. The leading signal lives in your server logs. AI-bot crawl activity tracks closely with how often you get quoted, so a rising crawl frequency from the AI user agents is an early sign that citations are about to move. When you see GPTBot, ClaudeBot, PerplexityBot, and the rest hitting your new pages more often, you are watching the mechanism warm up before the scoreboard changes. That is why log monitoring belongs in the measurement stack, not just the technical one.
Cadence and reporting
Measurement without a rhythm decays into a spreadsheet nobody trusts. Run the prompt library on a monthly cadence, and report it plainly. A good report shows the questions asked, the engines queried, the sources the engines cited, and the movement since last month, per engine and against competitors. No jargon wall, no vanity chart. The test of an honest report is that a reader can see exactly what was measured and could, in principle, reproduce it. Show the questions and show the sources, or the number is just an assertion.
Original data as lever and as metric
There is one input that is both a way to earn citations and a thing you can track: first-party research. When you publish original data, you give the engines something quotable that nobody else can offer, and you give yourself a clean number to watch as that asset gets picked up. The lever is real and measurable. Original data marked up in Dataset schema has been linked to as much as a 43% AI-visibility uplift (SearchX, 2026), which is exactly why it earns its own line in the measurement plan. Publish the study, then track how often engines quote it and how your citation share moves after it lands.
An honest note on precision
Measurement here is noisy, and pretending otherwise costs you credibility. Engines return different answers to the same prompt on different days, and they change how they choose sources without notice. So report ranges and trends, not false precision. A number like "named in four to six of ten answers, trending up over three months" is more honest, and more useful, than a fabricated "62.4% citation rate." The goal is a true picture of direction, not a decimal that implies a control you do not have.
Where to go next
Measurement only means something once there is work to measure, so if you have not read it yet, start with Chapter 5 on getting cited, which produces the movement this chapter tracks. Then go to Chapter 7 on the video layer, the part of the method most competitors cannot copy. And if you want the whole system framed on one page, with measurement built in and run for you, read the GEO agency guide.