How to measure AI citations without analytics

Your analytics measures visits. A citation is a mention inside a generated answer, and the whole design of that answer is to resolve the question without a visit. When it works best for the reader, it is most invisible to you.

Filtering referral traffic for the assistant domains is still worth doing. It captures the subset who followed a link, and those visitors tend to arrive with unusually clear intent because their question has already been answered and the click is a verification step.

But that subset is a floor, not a total, and the gap between the two is unknown and varies by query type. Reporting referral sessions as though they were citations is the most common measurement error in this space, and it consistently understates the effect.

Decide this before you start counting, because the definition determines the number and a definition that shifts mid-quarter makes the trend meaningless.

A workable definition has three cases. A linked citation, where your URL appears in the sources. A named mention, where the answer says your brand name without linking. And an unattributed lift, where the answer clearly uses your framing without naming you at all.

Count the first two. The third is real and genuinely valuable, but it cannot be detected reliably, and a metric you cannot compute consistently is worse than one you leave out and mention in the notes.

Start from real buyer language rather than keyword tools. The phrasing people use with an assistant is longer and more conversational than what they type into a search box, and it usually carries their constraints: the sector, the team size, the region.

Cover three intents. Discovery questions, where the buyer is asking what exists. Comparison questions, where they are asking which option fits their situation. And problem questions, where they describe a symptom without knowing the category name.

Keep branded queries in a separate list. An engine naming you when asked about you proves almost nothing, and mixing those results into the same figure will flatter it until nobody trusts it.

Then freeze the set. This is the discipline that makes the whole thing work, and the one most teams break first. Every prompt you add or reword changes the baseline, so improvements and regressions become indistinguishable from edits. When you want new coverage, open a second list and track it separately.

That last column is the one teams leave out and later wish they had. Your own citation rate tells you how you are doing. The competitor set tells you who is winning the slots you are not, and which of them to go and read.

Record the raw answer text too. It costs nothing to store and it is the only way to answer the question that always follows a bad quarter, which is what the answers actually said about the category.

Because generated answers vary between runs, repeat each prompt several times per cycle. Report the share of runs in which you were named. A single check is an anecdote; a rate across repeated runs is stable enough to trend even though no individual run is.

This is the method behind our own tracker: a fixed per-project prompt set replayed against each engine on a schedule, with every response parsed for brand and domain mentions to produce a rate over time (Osoro Solutions, 2026).

Citation rate is the share of runs in which you were named, computed per engine and never averaged across engines that behave differently. A blended figure hides the case that matters, which is being strong in one engine and absent from another.

Share of voice is your citations as a proportion of all brand mentions across the set. It is the number that survives contact with a leadership team, because it answers the comparative question directly.

Report both as a trend. The absolute level depends on how hard your prompt set is, which makes it close to meaningless in isolation and genuinely useful over time.

Say this out loud in the reporting, because it is the honest limit of the method and someone will find it eventually.

When a buyer asks an assistant to shortlist options and your brand is named, that shapes a decision that may not surface for months and will arrive attributed to something else entirely, or to nothing. No sampling method reaches it.

The method also cannot distinguish a citation shown to thousands of people from one shown to nobody, because the engines do not publish query volumes. Treat every prompt as equally weighted and be aware that they are not.

None of this makes the measurement worthless. It makes it directional, which is enough to decide what to work on next, and considerably better than the alternative of asserting that the work is going well.

Weekly sampling gives enough data points that a real change separates from run-to-run variance within a month or so. Daily adds cost and noise without adding signal, given how slowly the underlying signals move.

Report monthly. Technical and content fixes surface on the engines' crawl cadence, so a change you shipped on Tuesday will not show up by Friday and reading the chart that way will have you undoing work that was fine (Google Search Central, 2025).

Set the baseline before you change anything. The most common regret here is starting the measurement after the remediation, which leaves you with a number and no way to say whether it improved.