Being mentioned and being recommended are not the same measurement

Never named is a retrieval problem. The assistant either cannot read the brand, does not hold it as a distinct thing, or has nothing to reach for when the question comes up. Named but not chosen is the opposite failure: the brand is in the answer and the sale goes elsewhere, which is demand the brand already paid to create being handed to a rival.

Named wrongly is a repair job. The answer carries something untrue about the brand and the buyer never sees the page that would correct it. Chosen with no way to buy is the last of the four and the least noticed: the answer recommends the brand and gives the buyer nowhere to go.

Collapsing those four into one figure is where mention counting comes from. Three of them have presence in common, and so does an answer that recommends the brand, so a number built on presence alone reports a recommendation and a rejection as the same result.

Five figures are in common use and they answer five different questions.

• Presence. Predicts that the assistant can reach the brand at all. Says nothing about where in the answer it landed or how the answer described it.
• Mention. Predicts that the brand is in the running text. Says nothing about whether the same answer named somebody else as the choice.
• Citation. Predicts which evidence the assistant read. A page can be cited in an answer that recommends a rival, so a citation is proof the brand was consulted rather than proof it was chosen.
• Sole recommendation. Predicts the choice. It is the one figure of the five that reads the buying decision itself rather than an input to it.
• Top three. Predicts contention. Reaching the first three names is reaching the set a buyer actually considers, which is a different question from being the one name they are given.

They are not a ladder a brand climbs in order. They are five separate questions, and a month of work can move one of them and leave the other four exactly where they were.

A question set is not all one thing. Some questions ask what to buy. Others ask how a category works, what a term means, or which specification matters. Both kinds belong in a set, because the second kind is where a buyer assembles a shortlist before they ever ask who to buy from.

The two cannot share a denominator. If a recommendation rate divides by every answer, adding one research question lowers the figure with nothing having changed in the market. Add several more and it moves again. A number a question set can move on its own is a number about the question set.

So one idea gets one denominator and every screen uses it. A share of recommendation here is the share of answers to buying questions where the brand is the recommendation, and nothing else may be labelled with that word. Before that was settled, one idea was printed as three different figures on three surfaces of the same measurement, every one of them arithmetically correct.

They stay in the set and they are reported. Being the source an assistant reaches for on a category question is real ground, and it is the ground a shortlist gets built on. What changes is where the figure lands: presence on informational questions is published beside the score rather than inside it.

The exclusion is enforced at the point the score is computed rather than assumed further up, because an informational question arriving at the recommendation calculation is the exact failure the rule exists to prevent. A brand that has deliberately never published educational content has not failed at anything, and a score that says otherwise sends a team to work on the wrong thing.

The rule was written for recommendation first and was not applied to visibility for a while, so a brand correctly absent from a how-does-it-work question was counted as a visibility failure. That is the same error one step earlier in the chain, and the visibility figure was corrected the same way.

Three things have to be true at once. The denominator is every answer, so the easy half of the question set carries most of the weight. A mention counts the same whether the brand led the answer, sat near the bottom of it, or was named in order to be dismissed. And a research question is easier to be named in than a buying question, because a buying question is the one every rival in the category is competing on.

Hypothetical, and the brands are invented. Ardent is named in most of the answers about running shoes for flat feet, because assistants reach for it whenever the subject comes up. Kestrel is the recommendation on nearly every question that asks what to buy. On a mention count Ardent looks well covered, and it is losing the category.

Citations behave the same way. A page can be the source an answer was built from while that same answer names a rival as the choice, so a rising citation count and a flat recommendation rate are perfectly consistent. Being cited while a rival is the recommendation is not a contradiction. It is two measurements doing their jobs.

Disclosure is a property of the assistant and of the request, not of the brand. Some assistants publish the pages they read. Some publish them only when the question sent them to search. Some publish nothing at all, and an answer from one of those is not evidence that no source was read. It is evidence that nobody said.

A citation figure therefore carries a denominator of its own, separate from every other figure beside it. Delphi records which assistants disclosed sources on a run, so a reading can state how wide it is rather than dividing by the full set and implying that all of them answered the question.

• What counts as a hit. A mention, a first place, a top three place and a cited page are four different numbers wearing one word.
• What it divides by. Every answer collected, or only the answers where there was something to win.
• Which questions are in it. A rate over a set nobody has seen is a rate over a set that can be chosen to flatter it.
• Which assistants it covers. A figure only some assistants can produce is a figure about those assistants, whatever the heading says.
• How far it has to move. A number with no margin of error beside it cannot tell a result from noise, and a monthly report of movement inside the noise is a report of nothing.

A figure that survives those five is worth acting on. A figure that cannot answer the second one is a mention count, whatever it is called on the page it appears on.