Measuring LLM Citations: The Real GEO Metric (And How to Track It)

Let me save you a year of confusion. The metric you tracked for SEO does not work for AI engines. There is no position one in ChatGPT. There is no blue link to climb toward in Perplexity. The thing you need to measure is whether the model names your brand and links your page when a real buyer asks a real question. I call that citation share, and it is the only GEO number I trust.

Here is the shift. In classic search, the question was "where do I rank?" In AI search, the question is "do I get cited, and how often, against my competitors?" Those are not the same question. A page can rank on page two of Google and still get pulled into a ChatGPT answer because the model liked one clean paragraph you wrote. The reverse happens too. You own position one and the model never mentions you.

Citation share is simple to define. Take a set of buyer questions. Run each one through the AI engines. Count how often your brand or your URL shows up in the answer. Divide by the total. That percentage is your share. Track it against the brands that keep showing up next to you, and you have a scoreboard that actually maps to whether AI is sending you business.

This connects directly to the bigger picture I lay out in what is GEO. GEO is the practice of becoming the source AI engines quote. Measurement is how you know if the practice is working. Skip measurement and you are just guessing, redesigning pages on vibes, and hoping.

Before you can count anything, you have to decide what counts. This sounds pedantic. It is the whole game. If your definition of a citation is sloppy, your numbers are noise, and you will make bad decisions off them for months. I use three distinct categories, and I score them separately.

The three tiers I track

• Brand mention: the model names you in prose but does not link you. Still valuable, because the name lands in the buyer's head.
• Linked citation: the model names you and attaches a source link or footnote to your domain. This is the one that can drive a click.
• Sole or primary source: the model leans on you as the main answer and treats competitors as secondary or omits them. This is dominance.

Why split them? Because they call for different actions. If you are getting mentioned but never linked, your problem is source attribution and structure, not awareness. If you are not even getting mentioned, your problem is deeper. You are not in the model's consideration set at all. Lumping these together hides the diagnosis.

There is also the question of accuracy. The model can cite you and describe you wrong. I log a fourth flag for that: cited but inaccurate. A confident wrong description is a fire to put out, and it is invisible if you only count yes or no. This ties back to the structural fixes in how to get cited in ChatGPT, where clean, copy-ready claims reduce the odds the model garbles you.

The single biggest mistake I see in GEO measurement is testing the wrong queries. People test the prompts that flatter them, like their own brand name, and then celebrate. A buyer who already knows your brand name is not the buyer GEO wins. You need the questions people ask before they have decided anything. That is where citation share converts to revenue.

Start from the funnel, not from keywords. AI engines get used heavily for research and comparison, so your query set should be thick with problem-stage and comparison-stage prompts. These are conversational, longer, and messier than the keywords you targeted in keyword research for classic SEO. Write them the way a person actually talks to a chatbot.

Five buckets to fill

• Problem-aware: "how do I stop X from happening" with no product in mind yet.
• Solution-aware: "what kind of tool solves X" where they want a category, not a brand.
• Comparison: "best X for Y type of buyer" and "X vs Y" head-to-heads.
• Brand-adjacent: "is [competitor] good for X" and "alternatives to [competitor]."
• Decision: "is X worth it for a small team" and pricing or trust questions.

Lock the set once it is good. The whole point of measurement is comparison over time, and you cannot compare if you change the questions every month. Keep a stable core you rerun, and a small rotating set for experiments. When you do add a query, mark its start date so you do not mistake a new question for a trend change.

Here is the dirty secret nobody selling a GEO dashboard wants to say out loud. Run the same prompt twice and you can get two different answers, with different brands cited. The models are non-deterministic by design. If you test each query once, your data is a coin flip dressed up as a metric. This chapter is the difference between real measurement and theater.

Treat each query like a sample, not a fact. One run is an anecdote. The honest unit of measurement is the citation rate across multiple runs of the same prompt. If you appear in three of five runs, your presence for that query is sixty percent, not a yes and not a no. That fractional view is the truth, and it is far more useful than a single snapshot.

How I control the wobble

1. 1Run each core query several times, not once. More runs for high-value prompts, fewer for the long tail.
2. 2Fix the conditions you can: same model version, same region setting, fresh chat with no memory carried over, search mode on or off held constant.
3. 3Record the citation rate as a fraction, not a binary, for every query.
4. 4Aggregate to the engine level so you compare ChatGPT to Perplexity fairly, since their variance differs.
5. 5Re-run the full set on a fixed cadence so each period is measured the same way.

This is also why I distrust any single-run "GEO audit" screenshot. It captures one roll of the dice. The discipline of repeated sampling is unglamorous, and it is exactly what separates a number you can act on from a number that lies to you. The same rigor I apply to measuring Core Web Vitals, where field data beats a single lab run, applies here.

You need to be precise about what got cited. Sometimes a model names your brand from memory with no link. Sometimes it links one specific page it retrieved live. Sometimes it links a page about you that you do not even own, like a third-party review. If you only watch your own URLs, you will miss most of how AI talks about you, and you will optimize the wrong things.

Split attribution into two streams from day one. Stream one is the entity: is the brand named, regardless of any link? Stream two is the URL: which specific pages get linked, yours or anyone else's? These tell different stories. Strong entity presence with weak URL presence means the model knows you but does not trust a specific page enough to send a click. That is a fixable structural problem.

Third-party citations count too

Often the linked source is not your site. It is a roundup, a review, a forum thread, or a directory that the model trusts more than your homepage. That is still a citation that shapes the buyer, and it is a roadmap. If models keep citing a particular review site when they describe your category, getting represented well on that site can move your share faster than any change to your own pages. This is the off-site half of E-E-A-T doing its quiet work.

A citation count on one day tells you almost nothing. Citation share that climbs over six runs while a competitor's falls tells you everything. GEO measurement is a longitudinal sport. You are looking for direction and gap, not a hero number to screenshot. Here is how I structure the trend so it actually informs decisions.

Build a share-of-voice view, not a vanity count. For each period, total the citations across your whole query set, then express your slice as a percentage of all brand citations that appeared. Do the same for your top competitors. Now you have a chart where everyone competes for a fixed pie, and your wins and losses are visible relative to the field, not in a vacuum.

Slice the trend so it is useful

• By engine: your share in ChatGPT may climb while Perplexity stalls. Average them and you lose the lesson.
• By query bucket: dominating comparison prompts but missing problem-stage prompts is a specific, fixable gap.
• By competitor: track the two or three brands that keep appearing next to you, not the whole universe.
• By citation tier: rising mentions but flat links means a structural job, not an awareness job.

You do not need a four-figure subscription to start measuring GEO. You need a clear method and the patience to run it. I built my first citation tracker in a spreadsheet, and for many businesses that is genuinely enough. Here is the exact process, start to finish, that you can run this week.

Set up the sheet

1. 1List your locked query set, one prompt per row, tagged by funnel bucket.
2. 2Add columns per engine for each run: named (yes/no), linked (yes/no), primary source (yes/no), accuracy flag, and the linked domain.
3. 3Add a calculated column for citation rate per query per engine, the fraction of runs you appeared in.
4. 4Add a competitor presence column so you can compute share, not just your own count.
5. 5Date and version-stamp every batch so each period is comparable.

Run the test

1. 1Open each engine in a clean, logged-out or fresh session with personalization minimized.
2. 2Paste each prompt, run it the agreed number of times, and record every field honestly, including the runs where you do not appear.
3. 3Capture the linked sources verbatim, including the third-party domains, not just whether you showed up.
4. 4Hold conditions constant across the whole batch: same day, same model versions, same search setting.
5. 5Roll the per-query rates up into per-engine share and your overall share of voice.

I trust a humble spreadsheet I understand over a slick dashboard I cannot interrogate. If you cannot explain how a number was produced, do not bet a strategy on it.Shmul

Data that does not change a decision is decoration. The whole reason to measure citation share is to know exactly where to point your effort next. Once you have a few periods of clean data, the patterns tell you what to fix in plain terms. Here is how I read the sheet and turn it into a work list.

Start with your biggest gaps, weighted by query value. A comparison prompt where buyers decide is worth more than a vague problem-stage one. Find the high-value queries where your citation rate is low and your competitors are high. That intersection is your priority list. You are not trying to win every prompt. You are trying to win the prompts that move money.

Map the pattern to the fix

• Not mentioned at all on a query: you lack a page that clearly addresses that question. Build it, structured for extraction, as I cover in content writing.
• Mentioned but never linked: the entity is known but no page earns trust. Tighten structure, add clear claims and schema, per schema markup.
• Cited but inaccurate: the model has stale or muddled facts. Publish an unambiguous, current source page and reinforce it.
• A third-party domain keeps winning: go earn accurate representation on that source.
• Strong on one engine, weak on another: study what the winning engine rewards and adapt, since Perplexity and ChatGPT do not weight the same things, as I explain in ranking in Perplexity.