Share of Citation: Measuring 'AI Mentioned Me'

A DTC water-filter brand selling into the US and Germany is three slides into its monthly review when the growth lead asks for a content budget. The slide says AI mentions tripled since January. The saved answers say otherwise: one twelve-source roundup ranks the brand’s blog ninth without naming it in the body; another lists the brand only in a source list under a rival’s recommendation. The mentions are real. The position is not.

“AI mentioned me” is not a number but a fraction — your position in an answer over every source that answer cited — and counting mentions alone turns noise into a budget line.

Three receipts, not three grades of one metric

A mention, a citation, and a recommendation have separate tests, denominators, and conclusions. One line averages three rulers.

Layer	What it looks like	How to score it	Count
Mention	Brand string in body or source list	Exact string, no link or description	Mentions
Citation	Clickable URL to a specific page	Domain plus page a buyer can open	Citation slots
Recommendation	Brand in candidate set, comparison, or conclusion	Brand bound to a situation, plus an exclusion	Recommendation slots

Scoring the source list as a citation is the most common error. Those URLs prove the model read you, not that it used you. If no body sentence matches your page, it is a mention.

Recommendation needs two things at once: a concrete constraint — “fits a 16-inch cabinet gap, no drain line” — and an exclusion — “but the tank is small for a household of five.” Praise alone helps no buyer decide.

The denominator moves without you

Ask one prompt twice, get six sources then twelve, and your share halves while you lost nothing. The model changed the answer’s shape, not its opinion of you. Shrink to three sources and the share rises with no extra buyer.

Three rules follow. Store the total source count with every sample — numerator and denominator, never the ratio. Report absolute counts beside the share; each misleads alone. Compute per engine, never merged: an AI Overview, a chat assistant, and a marketplace search differ in length and source structure, and a blended total averages three rulers.

Then weight the prompts. “Which one should I buy for a rental kitchen” and “how do I descale it” are not worth the same. Group by brand, comparison, and scenario; weight from the revenue mix. Unweighted, the share flattens under low-intent questions: stable, information-free.

Small samples lie in both directions

A share is a proportion; on a small sample, it is not a trend.

Suppose 30 prompts across three engines, asked three times each: 270 answers, 90 per engine. Near a 20% citation rate, the 95% interval on 90 samples is about plus or minus 8 points by a binomial approximation. So 25% to 31% is noise and 30% to 52% a change; without four times the sample, a decimal move does not belong in a monthly report.

Discipline matters as much as size: freeze prompts and IDs; one region, one language; repeat a day apart; stay logged out; save raw answers and cited URLs. Fix four to six rivals in the same price band; swap one and the line restarts.

Example: mentions climb, share does not

Example — brand and figures invented; only the arithmetic is meant to be reused.

Month	Answers	Mentioned	Cited	Recommended	Avg. sources	Share
Jan	90	16	9	2	7.1	9/639 = 1.4%
Feb	90	21	11	2	9.6	11/864 = 1.3%
Mar	90	27	14	3	12.4	14/1116 = 1.3%

Mentions rose from 16 to 27 while share slipped a tenth of a point: exposure worked, evidence did not. Citations grew from 9 to 14, but slots grew faster; the new slots went to review sites and forums.

The response is not twenty more articles. Make cited pages easier to lift: a standalone conclusion, a spec table with units, a certification number anyone can check. See turning specs into evidence and which categories need the most evidence.

A five-step monthly routine

Step	Action	Owner	Passes when
1 Fix the track	Freeze four to six rivals; group prompts	Growth	Every prompt has a group
2 Score the layer	Three checkboxes plus total sources per answer	Ops	Every citation maps to a body sentence
3 Compute	Per engine, weighted, absolute beside share	Growth	No blended total share
4 Read the gap	Each gap maps to one fix	Content, technical, ops	Gap names page, field, or term
5 Change one layer	Re-test same prompts	Owner	Prompt IDs unchanged

Read gaps as assignments, not diagnoses. High mentions with low citations mean no liftable sentence. High citations with low recommendations mean third-party confirmation is missing, which is why the four trust prerequisites for AI recommendation come first. High recommendations with weak conversion usually trace to price, delivery, or returns, not content.

Where this breaks

The category is too small. A handful of real questions a month makes the denominator unstable. Keep a log, not a trend.

The assistant does not reproduce. Marketplace answers vary by account, region, and session. Note whether you appeared, not a share.

AI does not know the brand yet. Share is zero and arithmetic will not fix it. Confirm the entity exists, and keep recognized separate from correctly described.

Answers keep lengthening. When every engine adds sources, the industry’s denominator inflates; a flat share can be progress. Read absolute citations.

Share becomes a KPI for writers. Content’s deliverable is a paragraph that survives extraction. Managed on share, writers chase short answers and lose the evidence that earned the citation.

One error hides in the numerator: the brand is named but the model or capacity is wrong. Score naming and accuracy separately.

This line belongs to measurement and growth; its output is a decision about which layer to fix.

This month: turn 30 customer-service questions into numbered prompts, pick one engine, one market, one language, ask each three times, and log four numbers per answer — mentioned, cited, recommended, total sources. Four weeks will separate brand from denominator.