GEO-109·Issue 10·Measurement & Growth·109 / 132
Four Layers of AI Visibility: Mention, Citation, Recommendation, Conversion
Mentions, citations, recommendations, and conversions are four separate verdicts, not one score. A metric dictionary, guardrails, and a fix order for gaps.
Issue focusMeasurable: AI Visibility, Attribution, and Experiments
At an October planning meeting, a US DTC standing-desk brand puts one number on screen: how often AI answers name the brand, up three months running. Budget talk starts. The founder asks how many desks that sold; the slide has nothing. Support fills the silence — a shopper bought a frame from a screenshot saying it fits a 71-inch top while September’s bundle ships 78; the return cost more than the order earned.
AI visibility is four verdicts, not one score — seen, used as evidence, shortlisted, worth the money — and a gain at one says nothing about the next.
Four layers, four ways to fail
One figure treats four questions as one; each layer is a claim to falsify, not a funnel stage.
| Layer | Test | Failure |
|---|---|---|
| Mention | Brand string in the body or a list; link or not | Named, wrong size or price |
| Citation | A body sentence mapping to your page | Listed at the end, unused |
| Recommendation | Brand bound to a situation, plus an exclusion | “Also good” — the buyer still cannot decide |
| Conversion | Entry identifiable, joined to orders and refunds | Orders arrive, returns eat margin |
Mention is easiest to over-read: the screenshot is a mention win and a citation failure.
Citation needs a matching sentence: a source list proves the model read you, a body sentence pointing to your page proves it used you. Recommendation needs a constraint and an exclusion — “fits a five-foot-eleven user, under $500, but overkill for one monitor” — not “reliable quality”.
Conversion alone carries money and is easiest to mislabel. AI referrals often land in Direct or organic; only an order-side identifier separates them, and the split depends on which four AI search entry points delivered the session, as AI traffic attribution explains.
Why one score lies
A total assumes the layers compensate. They transmit; they do not substitute.
The weights have no source: why thirty percent for mentions, forty for conversion? Weight comes from order mix and margin, both moving quarterly. A blended total also averages three rulers, so one engine’s rise cancels another’s fall into “flat”.
Transmission is not automatic. A citation becomes a recommendation only with a reason and outside confirmation; a recommendation becomes an order only when price, shipping, and returns hold. A weak return policy turns recommendations into refunds.
Denominators move too: six sources become twelve and your share halves with nothing lost; report counts beside ratios — the arithmetic is in share of citation.
One metric, one guardrail, per layer
Each layer gets one denominator and one guardrail: naming answers over valid samples, guarded by wrong-spec answers; matching-sentence answers, guarded by source position; answers with a scenario and exclusion, guarded by candidate rank; orders from an identifiable AI entry over requests from it, guarded by refund rate and margin.
Freeze prompts, engine, region, language, rivals for a quarter; change one and recompute the baseline. Report per engine, never merged. Mentions and citations read weekly, recommendations monthly because an edit waits for a re-crawl, conversion monthly with a refund window. A weekly number cannot explain a monthly move.
Five steps, layer first
| Step | Action | Owner | Done when |
|---|---|---|---|
| 1 Split the slide | Four columns; mark the empty ones | Growth | Each has a number, or says “missing” |
| 2 Pick one layer | One quarterly target; red lines elsewhere | Growth, content | Only one column carries a target |
| 3 Freeze definitions | Numerator, denominator, sampling, exclusions | Data, growth | Two people get the same figure |
| 4 Assign the fix | Each gap maps to a page, field, or policy term | Content, engineering, ops | Every gap has an owner and re-test date |
| 5 Re-test | Same prompts; read only the changed layer | Ops | The other three held |
Step 2 is the one teams skip. Four targets mean everyone works the softest layer — mention, which any review post or misquoted spec can lift. That is self-congratulation: four metrics scored at once.
Step 4 fixes are specific. Low mentions with a clean entity means access, not content. High mentions with low citations means the answer is not liftable; the four slots of an answer unit address that. High citations with low recommendations means outside confirmation is thin; weak conversion sits with price, delivery, or returns.
Example: mentions climb, orders shrink
Example — brand, SKUs, and figures are invented; only the reading method is reusable.
| Month | Mention | Citation | Recommendation | Orders per 1,000 AI sessions | Refund rate |
|---|---|---|---|---|---|
| Jul | 22% | 11% | 5% | 18 | 9% |
| Aug | 30% | 12% | 4% | 15 | 13% |
| Sep | 38% | 12% | 3% | 11 | 17% |
Read the first column alone and this is a case for budget. Read all five: recommendations fell from 5 to 3 percent while refunds rose from 9 to 17, and “wrong size, will not fit” was half the return reasons. The answers still describe a 71-inch top; September’s bundle is 78, and the model cited the July page.
The fix is not ten more articles but one source of truth for variant dimensions, with page and structured data from the same record, then a re-test of layer three. Whether it sold more desks is separate: before-and-after shows correlation; a holdout shows cause.
Where this breaks
Too few questions, or an unresolved entity. A handful of questions a month makes ratios wobble; a brand AI cannot identify reads zero on all four.
In-platform assistants do not reproduce. Answers shift by account and session; record whether you appeared, not a rate.
Low ticket cannot measure conversion. Decisions complete in one session and AI referrals merge with organic; read recommendation and branded search.
Four layers cannot become four department KPIs. Split that way, content chases mentions and nobody owns whether the answer is right. The layers diagnose; appraisal hangs on the last one and its guardrail.
This month, split the figure you report upward into four columns and recompute 20 answers on one engine and one market. Mark which layer broke first: one, fix access; two, fix the sentence; three, add evidence; four, check price and returns. One layer deserves the quarter’s budget; the other three are its guardrails. The rest of this line lives in measurement and growth.