Incrementality for AI search
Your AI visibility went up. Did your sales?
We run pre-registered lift studies for GEO. Your team or agency rolls out AI-search work in waves across categories, use cases or markets. We measure what it added to sales and gross profit, with an honest interval.
The measurability check is free. Each study has a fixed price, agreed before it starts.
The problem
Visibility is not revenue
Mentions count answers, not buyers
Trackers tell you how often AI answers name your brand. They can't tell you whether those answers changed what people bought.
AI referrals miss most of the effect
Many shoppers read the answer, then search your brand or buy on a marketplace. Others clicked through but would have bought anyway. Referral counts get both wrong.
AI traffic grows on its own
In the one public natural experiment on answer-engine optimization, pages that got no changes grew about 3.5× over the same months. Before-and-after charts credit that growth to your work. Open data
The protocol
How a lift study works
Four steps. The decision rule is written down before the first week of data comes in.
-
Check
free · 2 minutesThe calculator below shows the smallest effect a test on your business could detect, in dollars and in ROI. If a realistic effect wouldn't show up, you know before spending anything.
-
Design and register
1–2 weeksWe split your catalog into units whose AI questions don't overlap: categories, use cases or countries. Similar units are paired, and a coin decides which one goes first. Metric, window and decision rule go into a protocol you sign off.
-
Run in waves
8–12 weeksYour agency or team does the GEO work on wave 1 only. Wave 2 waits its turn and serves as the comparison, so nobody goes without for good. We track AI answers to each unit's questions to confirm the work reached them.
-
Readout
1 weekIncremental orders, revenue and gross profit; ROI with a 90% interval; the probability that GEO paid back; and the decision the protocol committed to. Plus a plain list of what the study can't tell you.
Step 0
Can a test on your business see the effect?
This doesn't estimate your ROI. It shows the smallest lift a test could reliably detect and the ROI it takes to show up. The defaults describe the sample client further down.
Smallest ROI a test confirms, by test length
To lower it
Step 3, on synthetic data
Sample readout
Orders, indexed to the 26 weeks before launch
Raw weekly totals. Wave 1 got GEO from week 0, wave 2 waited. Promotions make single weeks jump; the model accounts for them.
Show data
Effect on wave-1 orders, by week
Before launch the estimate should sit near zero. That is the placebo check.
Show data
Share of tracked AI answers that name the brand
Show data
ROI across all posterior draws
Show data
Same pipeline on invented clients
Show data
What this readout can't tell you
- It measures this GEO work, in these categories, over these 12 weeks. Another tactic or season can behave differently. That is what the next wave tests.
- It reports the total effect of the work. It doesn't split it into "through AI answers" and "through search or the site". The AI-answer chart only shows that the work reached the answers.
- Spillover to wave-2 categories, such as a brand-wide halo, isn't counted. If there is one, the estimate is too low.
- It can't say which individual orders came from AI.
Data and ground rules
What we need from you
Weekly aggregates, no personal data. For B2B SaaS, read signups, demos or pipeline by use case, vertical or language wherever this says orders.
| What | Why | Status |
|---|---|---|
| Weekly orders and revenue by product: 26+ weeks before launch, then through the test | The outcome, and how noisy it is | Required |
| Which unit each product belongs to | Units are what gets randomized | Required |
| The GEO plan: what changes, in which units, from which date | The treatment | Required |
| GEO cost and gross margin | ROI in profit, not revenue | Required |
| Prices, promotions and stock-outs by unit | Separates them from the GEO effect | Recommended |
| Other unit-level marketing changes during the test | Same reason | Recommended |
| An export from your AI-visibility tracker for each unit's questions | Shows the work reached the answers. We can track the questions ourselves instead | Optional |
| AI referrals by landing page (GA4) and non-brand search (Search Console) | Faster, less noisy signals next to sales | Optional |
Random order
A coin, not a hunch, decides which units go first.
The comparison waits
Wave 2 gets no GEO work until the readout window ends.
Changes get logged
Any other unit-level change during the test goes into a shared log.
No moving the goalposts
The protocol isn't edited after launch.
Who it's for
Agencies, brands and SaaS teams
GEO agencies
Show clients what your work added to sales. Each report carries a "verified" mark and a link to the protocol registered before launch, so it holds up in front of their CFO.
Brands
Decide whether to scale, change or stop GEO with a number your finance team accepts. Works with any AI-visibility tracker you already use.
B2B SaaS
Split by use case, vertical or language and measure signups, demos and pipeline instead of orders.
Questions
What people ask first
We already see AI traffic in GA4. Why run a test?
Referrals show who clicked, not who was persuaded. They miss shoppers who read an answer and then searched your brand or bought elsewhere, and they count people who would have bought anyway. A lift study compares units with and without the work over the same weeks, so both errors cancel out.
Do some categories have to go without GEO?
Only for the test window. Wave 2 starts as soon as the readout window ends. Most teams roll out in batches anyway; the study only decides the order with a coin.
What if the result is zero?
The protocol fixes the decision rule in advance, so a weak result can't be explained away, and a strong one can't be argued down. A null with a known upper bound is information: change the tactic before spending more. Agencies usually start with the tactic they are most confident in.
Our sales data is confidential.
We work with weekly aggregates by unit, with no personal data, under NDA. The calculator on this page runs entirely in your browser.
How is this different from SEO split testing?
Page-level split tests measure traffic to pages you can edit. A lift study measures sales across whole categories or markets, so it also covers work that happens off your site, like reviews, roundups and PR.
AI answers change every week. Will the result hold?
Comparison units live through the same model updates, so platform-wide changes cancel out. Whether the effect lasts is a question for the next wave, which doubles as a replication.
Do you measure paid placements in AI answers too?
Yes, as a second arm of the same study: ads on or off crossed with GEO on or off, by unit. It shows whether ads add sales where you are already recommended organically, or only where you are not.
Get started
Start with a free measurability check
Send rough numbers or a weekly sales export. Within five working days you get the calculator's verdict on your own data and a draft protocol.
We are running the first studies with a small group of design partners at a reduced fixed price, in exchange for a case study we can publish, anonymized if you prefer.