How to Measure GEO ROI Without Perfect Attribution

An honest GEO ROI compares gross profit from extra new customers with what the work cost, and the hard part is "extra". Count only what the worked part of your business gained over a held-out part, in AI answers and in post-purchase survey answers, and put an error band on both gaps before you call a result.

Why GEO ROI does not show up in GA4

GA4 sees only the AI visits that keep a referrer, and most of what GEO changes happens before any visit: a buyer reads an answer, remembers a name and searches for it later. Those customers arrive as branded search or Direct. AI-referred revenue in GA4 is a floor, so an ROI built on it comes out too low.

Google's own documentation draws the boundaries. GA4's AI Assistant channel covers arrivals from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok, while clicks from AI Overviews and AI Mode are filed under Organic Search, "including Google’s AI Overviews and AI Mode". In Search Console those features sit in the Performance report within the "Web" search type, mixed in with ordinary results.

Our GA4 setup guide opens with a founder whose GA4 showed eleven ChatGPT sessions in ninety days, while the checkout survey on the same store said one in nine customers found them through an AI assistant. On stores we work with, the survey "consistently reports AI discovery at a multiple of what the channel group shows". That guide covers the tracking setup. Here we use the numbers to decide whether GEO paid for itself.

Start with a break-even target, not a dashboard

Write down, before the work starts, how many extra new customers the quarter's GEO spend has to produce to pay for itself. That number turns "is GEO working?" into a test with a pass mark, and it stops a rising visibility chart from passing as a return. Use whatever you actually pay.

Three inputs set it. Use gross margin, not revenue, because a €50 order does not return €50. Use new customers only, because a repeat buyer did not need an AI answer to find you. And decide how far ahead you value a customer: first order only is strict, twelve months is fairer for a brand with real repeat purchase.

On Shopify, the customer reports under Analytics > Reports include New vs returning customers and Customer cohort analysis. Shopify defines a returning customer as one whose order history already includes at least one order, so the first report gives you first orders per quarter and the cohort report gives you how often new customers come back within a year. Take gross margin from your own P&L.

Take a hypothetical online tea brand selling in Germany and the Netherlands at €1.2M a year with a €50 average order. That is 24,000 orders a year, or 6,000 orders a quarter, and say half of them, 3,000, are first orders. At a 60% gross margin, each order leaves €30 of gross profit. If a new customer orders 1.6 times in their first year, they are worth €48 of gross profit over twelve months.

For cost, we use the price we can cite exactly, our own. The Core GEO retainer on our pricing page is €2,200 a month excluding VAT with a 3-month minimum, so the first quarter costs €6,600 before any audit or one-off sprint. We ask for a first quarter because, as our full growth partner page puts it, "GEO does not produce a readable signal in a month".

The tea brand will do the work on half its product lines, 1,500 first orders a quarter, so extra customers can only come from that half. The break-even is the rise in the share of those buyers who credit an AI assistant, over and above the rise in the held-out half:

Written down first, those gaps show that a brand of this size cannot expect one quarter of GEO to pay back on first-quarter customers. Our cost guide shows how to price the spend itself; this table shows what that price has to earn.

The three layers of evidence

Each data source misses part of the effect, so read three together. Answers show whether the engines changed what they say. Visits show clicks that kept a referrer. The post-purchase survey shows customers who credit an AI assistant, including those who never clicked. ROI rests on the survey; the other two tell you whether to believe it.

For layer 2, Search Console has split branded from non-branded queries since November 2025. Google says the split is made by an internal, AI-assisted system, not by a keyword list, and that it is offered only to top-level properties and sites with a "sufficient volume of queries and impressions". Branded clicks rising while non-branded stay flat fits people hearing your name somewhere and searching for it. Paid social and PR produce the same pattern, so read it next to the other two layers.

For layer 3, ask after the order, on the confirmation page, rather than adding a required field to checkout, which costs you friction where it hurts most. Name the option plainly, keep the wording fixed, and export the answers with the order so you can group them by product line. Changing the question changes the share.

Hold out half your prompts and product lines

AI use is growing on its own, so more customers would credit an AI assistant next quarter even without GEO. To separate your work from that drift, split your buyer questions, and the product lines they lead to, into two fixed halves. Work on one half only and judge it on the gap between the halves.

This is the logic Google's CausalImpact package rests on: it estimates an effect by comparing the treated series with control series that it assumes were not affected by the intervention. A held-out half of your own business is a control you can build without a statistics package. Growth in AI use hits both halves, and so does seasonality if both halves sell in the same pattern through the year, so the gap removes them. Do not put gift sets in one half and everyday lines in the other.

The split only holds for work aimed at specific questions and products: new or rewritten pages, product data, comparison content, placements about one range. Brand-wide work such as fixing your entity data, review volume or press coverage lifts both halves, and then the gap understates what the work did. Do the brand-wide fixes before the baseline quarter, or accept that the control will make the work look smaller than it is.

If an agency runs the work, the query set has to stay fixed for the split to mean anything. On our pricing page, Core re-tests 38 buyer queries monthly under identical conditions on 4 engines, and the terms say "The query set does not change month to month." Split those 38 into 19 worked and 19 held out, and each half is 76 prompt-engine pairs. Core runs each query once a month, so count volatile pairs yourself, with three runs on different days in the first week.

A worked example: the tea brand after one quarter

The tea brand ran the split for one quarter. Its answers moved just beyond what chance usually produces; its survey answers did not, yet. All numbers are hypothetical, chosen to show how the reading works, including where one quarter stops short of a conclusion.

Answers. On the worked half, the mention rate went from 18 of 76 pairs to 30, or 23.7% to 39.5%. On the held-out half it went from 17 to 19, or 22.4% to 25.0%. The worked half rose 15.8 points, the held-out half rose 2.6 points: a gap of 13.2 points. Using the noise-band formula from our guide to re-measuring AI visibility, with 24 volatile pairs in each half at one run per reading, chance alone rarely produces a gap larger than about 12.6 points, about one time in twenty. The gap clears that band, narrowly. A gap of 10 points would not have.

Survey. About 40% of first-time buyers answered, 1,200 answers a quarter, or 600 answers in each half. Before the work, 2.5% in both halves credited an AI assistant. After one quarter, the worked half was at 6.0% and the held-out half at 3.5%: a gap of 2.5 points. For four groups of 600 answers, the standard formula for proportions puts the band on that gap at about 3.0 points. The gap is inside it. One quarter of survey answers cannot yet prove that the work brought customers.

Value. As a point estimate, a 2.5-point gap across 1,500 first orders is about 38 extra new customers, worth €1,800 of twelve-month gross profit. That leaves the quarter €4,800 short of its €6,600 cost, and the band on the survey gap includes zero.

Verdict. A first quarter often ends like this. The answers moved where the work was done and barely moved where it was not, which is the part that can be shown in one quarter, and our published timeline puts consistent movement in the third-party layer at three to six months. The brand should run a second quarter, pool both quarters of survey answers so the band narrows, and carry the €4,800 shortfall into the payback sum rather than starting the count again. Had the worked half risen 5 points and the held-out half 3, the gap of 2 points would sit far inside the band. Unless brand-wide work explains the held-out rise, the right call would be to change the work.

Check one more thing before believing a survey gap: that first orders in the worked lines grew faster than in the held-out lines, not just the share crediting AI. If first orders grew at the same rate in both halves while the AI share climbed only in the worked half, the survey is recording customers who would have bought anyway, now giving AI the credit.

What the example does not count: customers who bought through Amazon or a retailer after an AI answer, value beyond twelve months, and buyers who skipped the survey, who may differ from those who answered. Compare respondents' average first order with everyone else's as a rough check.

What the ranking guides get wrong

On 18 October 2026 we scored the ten GEO ROI guides that two US English searches for this question returned, all from agencies or tool vendors, against five checks fixed before reading. Most recommend a "how did you hear about us" field and branded search. None computes ROI on margin, gives a break-even or uses a control.

The pages were Siege Media, WebFX, Foundation, Wild Creek Studio, Addlly, LightSite, Mersel AI, Temso, Single Grain and AuthorityTech. A page passed a check only if its text gave a method, not just the word. Of the seven that pass on self-reported attribution, Foundation does so only in passing, and most place the question on B2B demo or intake forms rather than after an order.

The common formula is revenue-based. WebFX gives "GEO ROI = (AI-attributed traffic × conversion rate × customer lifetime value)" divided by the GEO investment, which multiplies the visits GA4 can see by a revenue figure and ignores both margin and the customers who never clicked. Addlly uses the same formula.

On incrementality, the closest any page came is annotation. Siege Media asks readers to mark every brand-building event on the branded search trend and says "The remaining unexplained lift is your defensible GEO signal." That subtracts the causes you remembered to write down; a held-out half also catches the ones you did not. AuthorityTech skips the question: "the delta in your scores is the return", with no control and no repeat runs.

Only LightSite mentions noise at all, warning that "single-query noise is high", and it gives no range. None puts an error band on a survey share.

Renew, change or stop: a decision table

Decide on two gaps: the answer gap between worked and held-out halves against its band, and the survey gap against its band. Visits and branded search break ties. A single quarter can tell you whether to continue; payback usually takes pooled quarters of survey answers.

Two more rules apply. Do not value AI-referred orders above other orders: our claims audit rates "AI-referred traffic converts better than organic" as unsupported, since AI users "arrive later in their decision process, which the AI did not cause." And hold agency reports to the same test as your own numbers. As Dom puts it on our about page, "If it does not move citations, CPA or ROAS, it did not happen — however good it looked."

What to set up before the next invoice

Before the next billing period, fill in the break-even table with your own margin, first orders and GEO cost. Add a post-purchase survey question with fixed wording and record a quarter of answers before work starts. Split your buyer questions and product lines into worked and held-out halves, and give the split to whoever runs the work.

Then run three answer captures on different days to count your volatile pairs, so you know the size of gap you need before the first report arrives, not after.