GEO-060·Issue 05·Measurement & Growth·60 / 132
From Citation to Checkout: The Conversion Path of Trust Evidence
Citations rose while orders stayed flat. Read branded search, direct visits, and a self-reported source together, count once per order, then swap one asset.
Issue focusEvidence of Trust: Reviews, UGC, and Third-Party Corpora
First Monday in May, the growth standup at a brand selling portable power stations into the US and Germany. A marketer shares a screenshot: six citations across three assistants for “best portable power station for RV camping,” double last month’s count. Orders are flat, branded search barely moved, and AI-attributed sessions round to nothing. The founder asks what the slide cannot: what did that evidence buy?
A citation proves a model was willing to use your sentence, not that a buyer will pay for it — judge the asset by joining it to branded search, direct visits, and a self-reported source at checkout, counted once per order.
A citation happens upstream of the decision
The answer box compresses the research phase. A buyer asks three follow-ups — capacity for a 30-amp hookup, charge time from a panel, noise at night — and leaves with three to five candidates. The click that analytics would have logged never happens: they type the brand into Google, enter the domain, or search the model number on a marketplace. The session dissolves into the answer; the decision does not. This is the four AI search entry points seen from the reporting side.
Session data undercounts that work; a buyer who remembers your model from a comparison table gets filed under Direct. Citation counts overcount it; a citation can sit in a paragraph about a shortcoming, or describe a variant you do not ship there.
Assistants also cite you as the control — “Brand A claims 3,000 cycles, Brand B 2,000” — which helps the buyer, not you. Citations that precede orders carry a specific number with its test condition and one difference a buyer can repeat out loud. Without both, it is a mention, the failure the answer unit framework exists to prevent.
Three numbers, or you will misread all three
| Number | Question it answers | What contaminates it |
|---|---|---|
| Branded search | How many remembered you and came looking | Campaigns, promos, seasonality |
| Direct visits | How many skipped every entry point | Bookmarks, email, pasted URLs |
| Self-reported source | What the buyer says they did | Sample bias, low response |
Branded search is the closest proxy for “the assistant made you memorable,” but split it by market and term type; paid brand campaigns fake the lift. Direct is the most flattered number: a pasted URL, an email click, a bookmark all land there, and calling Direct an AI referral is the common self-deception — see how to separate AI referrals from Direct.
Self-reported source is crude and useful: one question at pre-checkout, “How did you find us today?”, with “AI assistant” written out and “can’t remember” kept. Low response rate, real direction.
Windows, deduplication, and a definition you stop editing
Attribution breaks first on the window. Assistant research happens days to weeks before the order, so a 7-day window misses part of it and a 90-day window credits unrelated touches. Fix two windows rather than hunt for the correct one — short for direct effect, long sized to the purchase cycle — publish both, and write the choice down so it survives an analyst leaving.
Then deduplicate. One person can ask in ChatGPT on Tuesday, search the brand on Thursday, and order from a phone on Sunday; by session that is three AI-influenced visits and one buyer. Count by order: one primary source, everything else an assist. Cross-device stays unsolvable: state the error band, not precision.
Freeze the definition last. Switching between “share of sessions from AI domains” and “citation count” produces charts that contradict each other. Whether the evidence was worth citing is a separate question — the four trust prerequisites cover it.
Four steps from citation to order
| Step | Do this | Owner | Passes when |
|---|---|---|---|
| 1 Tag citations | List pages cited in 30 days; label each linked or unlinked, positive or used as a drawback | Ops, content | One URL and one quoted sentence per row |
| 2 Add one question | “AI assistant” in the pre-checkout survey; keep “can’t remember” | Front end, support | Response rate visible, nothing pre-selected |
| 3 Fix the definition | 7- and 28-day windows sized to the category, per-order dedup, documented | Analytics | Two people get the same number |
| 4 Run one change | One market, one category, one asset changed for two weeks | Growth, content | Two of three numbers move together, or a declared miss |
Step 4 is the only falsifiable step and the one teams skip; a flat result is harder to present than a screenshot. One asset changed against a held comparison beats ten more pages: only here does no movement mean something.
Example: one evidence swap, two weeks
Example — brand, product, and figures are invented; no real company is described.
A pet water fountain brand ran this on its German store, replacing a pump-lifespan paragraph with a table of test conditions: water hardness, run hours, failure criterion. The same sprint added the source question to pre-checkout.
Two weeks in, the invented read: branded search 40 to 44 queries a day, “AI assistant” survey share 3% to 4%, orders 120 to 126. Same direction, small movements — enough to conclude “the chain has a pulse and earns four more weeks,” not “GEO drove growth.”
Branded search up with orders flat means the buyer remembered you and was not persuaded: fix the offer or the page. Survey up with branded search flat means the option is leading people: fix the question.
Where this breaks
Revenue sits inside a marketplace. The buyer asks and buys on Amazon, where the site chain sees nothing; measure branded search there and repeat purchase.
Low-ticket impulse categories. Nobody runs three rounds of research on a $19 item; the chain costs more than it returns.
A live reputation problem. Mid-crisis, an assistant citing you is citing the risk; publish the remediation record before measuring conversion.
Attribution as proof. It orders priorities; it does not reimburse. An ROI quoted to two decimals collapses at the first review.
This month, list every page an assistant cited in the last 30 days, tag each linked or unlinked and positive or used as a drawback, pick one unlinked page that keeps getting cited, and add a sentence at the top pointing to the product page. Recheck branded search and the survey answer in two weeks. The rest of the chain lives on the measurement hub.