What a citation rate cannot tell you about revenue

What a citation rate measures, exactly

A citation rate is the number of prompt runs in which you were named divided by the number of prompt runs executed, on a stated set, per engine, over a stated window. That is all it is. It describes presence in an answer.

It does not describe preference. An answer that names four companies has named you without recommending you, and the sentence you appear in matters more than the fact of appearing: being listed as the budget option is a different commercial event from being named as the specialist, and both count identically in the rate. It does not describe position either, because there is no position to occupy. And it certainly does not describe demand, since the rate is computed over questions somebody chose to ask on your behalf.

A high rate over a small number of opportunities is a small sample rather than a strong result, which is why the rate travels with its run count. That much is arithmetic. The harder problem is what happens between the answer and the money.

Why the click is missing

The mechanism that makes AI answers valuable is the same mechanism that makes them hard to attribute. The buyer asks a question, reads a synthesized answer, and forms a shortlist without visiting any of the sources. If they later act, they type a company name, and the visit that reaches your analytics is a branded search or a direct arrival with no trace of where the name came from.

The step before that is invisible too, and for a different reason. The crawlers that fetched your pages to feed those answers do not execute the tag your analytics run on, so they leave no row in a client-side analytics product at all. Their traces are in server or CDN logs, which is a different system that most marketing teams never read.

Google's Search Console does now report impressions from AI experiences, and Google has said an impression there means a link to your site was shown. That is genuinely useful, and it is narrower than it sounds: it counts a shown link. An answer that named you in prose without linking you is the common case in some engines, and it produces no impression to count. So even the best first-party instrument for this sees a subset of the influence you are paying for.

Why old click-through curves cannot carry a forecast

The standard revenue forecast multiplies search volume by a click-through rate by a conversion rate by a value per conversion. Three of those terms are still fine. The click-through term is the one that broke, because the published curves that supply it were measured on a results page that did not have a synthesized answer sitting above it.

We do not publish a replacement curve, because we have not measured one, and the figures in circulation are vendor estimates rather than survey data. The curve that describes your category is the one inside your own Search Console, segmented by whether an answer appears for the query. That is a first-party measurement, it is free, and it is specific to you, which makes it strictly better than any published average regardless of how well sourced the average is.

The consequence for anyone selling this work, us included, is that a forecast built on a borrowed curve overstates revenue in exactly the direction that wins the deal. That is worth saying plainly rather than discovering in a quarterly review.

What to instrument instead

Four things, read together and never blended into one number. Branded query volume and direct arrivals over the same window as the capture, so a rise in people looking for you by name can be set against a rise in how often you are named. Assisted paths rather than last click, because last click in this channel is a branded search taking credit for the answer that caused it. Server-side or CDN logs, for crawler activity and for the referrals that some assistants do pass. And a first-touch question on your enquiry form, which is cheap, imperfect, and worth reading as a sample rather than as attribution.

None of those four is a clean causal link, and presenting them as one would be the same failure this essay is about. What they do is bound the question. If citation share rises across a stated window and branded demand does not move at all across the same window, that is a finding worth acting on. If both move, you have a correlation you can defend as a correlation.

The discipline that makes any of it work is deciding what would count as evidence before the work starts. A metric chosen after the result is a story.

The case where the metric moves against you

One scenario is worth naming in advance, because a client will otherwise read a win as a regression. It is possible for mentions of a brand to rise while citations of its site fall, on the theory that once a model has absorbed a fact it stops attributing a source for it. If that is real, citation rate is an inverted metric for a brand that has succeeded at the slow layer.

We have not tested it and we are not asserting it. It is a hypothesis with a plausible mechanism, and the reason to raise it here is that the moment to explain a metric's possible inversion is before it happens, not in the meeting where the chart goes down.

The line worth holding

We will not convert a citation rate into a revenue figure, and we would encourage you to ask any provider who does for the conversion factor and how they derived it. In this category the answer is usually an assumption stack with a percentage on the end.

What we will do instead is say which of the two visibility problems we diagnosed, what it costs to fix, what would count as it having worked, and over what window we would expect to read it. That is a smaller promise than a revenue projection. It has the advantage of being checkable.