Build the Baseline: What Is Your AI Visibility Today?

At a quarterly review, a DTC portable power station brand shows one chart: AI-assistant sessions up three months running. Orders did not follow, and the analyst admits many of those sessions are logged as Direct. Forty minutes later, nobody can answer the simpler question: three months ago, when a buyer in Ohio asked which unit runs a CPAP overnight in a small apartment, was this brand in the answer — and what did it call the brand?

A baseline is not a screenshot of one good day. It is the same frozen questions and fields, repeated monthly, so you can see which layer this month’s work moved.

Four layers that never add up to one score

Four layers, four collection methods, four failure modes — never one score.

Layer	What it is	Record	Common misread
Mention	Name appears	Sentence around it	A mention is a recommendation
Citation	A link a buyer opens	URL and source type	Counting domain, not page
Recommendation	Shortlist entry	Position and reason	One appearance as a stable spot
Transaction	Order traceable to AI	Session, cart, order	AI referrals logged as Direct

The gaps diagnose better than any score. High mention with low citation: the model knows you but has no sentence it can lift. High citation with low recommendation: it read your evidence and still lacked a reason — usually third-party proof, not more self-description. High recommendation with low transaction: the content worked; price, stock, delivery, or returns did not. That is measurement and growth work: a dashboard that cannot name the layer to fix should not take team time.

Thirty prompts, frozen

A prompt set is not a keyword list. Buyers do not type “best portable power station”; they ask whether one can fly, whether a third-party solar panel voids the warranty, and how return shipping works in Germany.

A workable split for 30: six discovery, eight comparison, ten constraint-and-scenario, six direct brand. Weight it to your order mix — brand-heavy revenue needs more of the fourth, acquisition more of the second and third. Source them from support macros, site-search logs, and return reasons; the buyer question chain shows how to expand each into a real follow-up.

Two rules beat the split. Freeze the set: new questions get new numbers, old ones never change, or last month’s comparison breaks. And never test only questions you already win. A baseline measures the gap, not the trophy shelf.

Seven fields, and a sampling habit

A table that records only “mentioned: yes” is dead by month two: it cannot say what the model said. Record seven fields.

Field	Record	Why it is mandatory
Prompt	Verbatim, fixed number	Re-tests compare words
Engine, mode	Product, mode, signed in?	Entry points assemble differently
Region, language	Location, interface language	Answers differ by market
Date	To the day	Indexes move in batches
Description	How you were described	“Chinese seller” is not “EU warranty”
Citation	URL, source type	Owned, marketplace, media, community
Recommendation	Entry, position, reason	Position beats count

Discipline beats fields: three runs per prompt, a day apart, accounts and regions in separate tables. At 30 prompts, a decimal place is false precision — report presence or absence, plus direction.

Prove the page can be read before testing whether it is cited. If product pages collapse without JavaScript, low citation is a crawl problem; start with whether AI can render your store.

One page, five numbers, one sentence

Four weeks, no fifth step.

Week	Work	Owner	Passes when
1	Write and number 30 prompts, freeze	Growth	Each traces to a real question
2	Three engines, three runs, raw text	Ops or intern	Every cell has answer and URLs
3	Five numbers, one sentence	Growth	Report fits on one page
4	Next month’s action, one layer	Lead	Names pages and fields

Five numbers: mention rate, citation rate, recommendation rate, share of citations on your own domain, source-type mix. The first three are counts out of 30; the last two explain movement.

Then say it plainly:

In [market], in [language], for [category], AI describes us as a [ ] company, on the evidence of [ ].

A rate invites an argument about method; this sentence is either right or visibly wrong, and both produce a decision. If you cannot fill it in, you measured presence, not meaning.

Example: a fictional brand’s first baseline

Example — a fictional power station brand, US only, 30 prompts, three engines, three runs. First baseline: 11 mentions, three citations, two of those to a page older than two years; most cited sources are community threads and review sites, the owned domain near zero. Answers call it “a Chinese seller” and file the category under “power bank”.

The count was not the point. It killed a plan for 20 more long articles: with almost no owned citations, the constraint was an inconsistent category description and no liftable conclusion. The month’s work became one category definition, a conclusion sentence on five core pages, and three quotable policy facts — starting with whether the brand entity exists, not with output.

Where this breaks

It fails in five places. Where a category has little question volume, the prompts were never asked and you are measuring noise; validate demand first. In-platform shopping assistants vary by account, region, and session: record whether you appeared and why, never a trend. Model or index updates can lift or depress everything at once, so read trends against engine versions, or you will score a system change as your own success. Without someone to re-run it, ten prompts monthly beat 30 that ran once. And a baseline never explains causation: it says which layer did not move, not why.

For why the entry points diverge, see how AI search splits cross-border traffic.

If you take one action this month, copy 30 questions from support tickets and site-search logs into a numbered sheet, pick one market, language, and engine, ask each three times, and store raw answers with citation URLs. Then write the sentence about what AI says you are; if you cannot, fix that before planning growth.