Proving GEO pays: the test design your CFO can accept

Your CFO is right to ask for proof that GEO pays, and the honest answer has two parts. I could not find that proof anywhere in the public record. And for most brands, a direct revenue test could not settle the matter even if GEO works, because normal sales noise would hide a plausible effect.

So split the question into two links: does your work change what the engines answer, and do those answers change demand? The first can be tested in twenty weeks with holdout topics. The second can only be tested where you have separate markets, and with limited statistical strength, which is better announced before you start.

The sums below use a fictional brand. They show a method, not a result.

Is there published proof that GEO pays?

No, not as of 7 November 2026: I found no published study with a control group that measures GEO's effect on revenue. Private tests may exist, but none can be checked.

The broadest review, by Olivier Martinez (45 studies from November 2023 to July 2026, a single-author preprint), rates confidence in the link from citation scores to clicks, conversions or revenue as "very low". Three pieces of work come close, none of them on revenue:

• The closest test measures traffic. Watanabe and Nakayashiki (June 2026, preprint) study their own employer's site, Glasp. ChatGPT referrals to the whole domain grew 5.7x, but untreated pages on that domain also grew 3.5x. With those pages as the control, the estimated multiplier is 1.82x (95% CI 1.31 to 2.54), and a placebo-in-time test returns p = 0.16. Their own verdict: "suggestive, not conclusive".
• The industry figure cannot be checked. Pinterest reported 20% organic traffic growth in February 2026. Martinez notes that it is not described in enough detail to support a general estimate.
• The most direct search came up empty. Kevin Indig (Growth Memo, 19 October 2026) compared AI visibility with Direct traffic and branded search across four brands and found no reliable link. He adds that an effect could still exist and that he could not isolate it.

Absence of proof does not show absence of effect. It only rules out presenting GEO's return as settled. For the evidence behind each tactic, see the ranking of GEO tactics by strength of evidence.

Why does a correlation between visibility and sales prove so little?

Because the same causes lift both lines. A brand that sells more gets cited, searched and discussed more, and a seasonal peak or a TV campaign raises everything at once.

The precedent to show a CFO comes from paid search. At eBay (Blake, Nosko and Tadelis, Econometrica, 2015), a standard regression put the return on non-brand search ads at over 4,100%, and still over 1,400% with time and geographic controls. The experiment, which stopped those ads for a random 30% of US traffic over 60 days, found -63% (95% CI -124% to -3%). eBay is unusually well known, and the paper does find a positive effect on new and infrequent customers, so the number does not transfer. The mechanism does: a channel gets credit for demand that already existed.

GEO already has its own version. On one site, Indig got r = 0.72 between AI share of voice and Direct traffic at a three-week lag. He then gave misaligned series the same chance to search for a fit. The real dates beat most of them, not all, and several matched or beat the real alignment. Hunting for the right lag is enough to manufacture a correlation worth a slide.

What is an incrementality test, applied to GEO?

An incrementality test is a comparison between what happened with an action and what would have happened without it, estimated from a group that did not receive the action and was chosen by chance.

Advertising can randomize individuals because an ad server decides who sees what. GEO cannot: nobody controls which user receives which answer, and Search Console's generative AI report, announced on 3 June 2026, is limited to impressions. So you have to vary something other than the person.

Why hold out topics rather than prompts?

Because engines do not respect the boundaries of your prompts. Google says AI Overviews and AI Mode may issue several related searches behind one question, so a page written for one prompt can surface on its neighbors. A holdout prompt inside a topic you are working on benefits from that work, and the test understates the effect.

You pay for this in the arithmetic. 120 prompts in 12 topics, across 4 engines and 20 weeks, produce 9,600 answers but only 12 units of analysis. Even read one by one, with a correlation of 0.2 between prompts in a topic (an assumption), 120 prompts are worth about 43 independent observations. How many prompts a reliable measure needs covers the sizing.

Three rules protect the test:

• Randomize, never choose. Treating your weak subjects and keeping your strong ones as the control manufactures a gain. In my simulation (120 prompts, true rates spread around 25% with a 15-point standard deviation, 8 answers per prompt before and after, no real effect), the weakest half at baseline gains 8 points and the strongest half loses 8. That is regression to the mean. Pair topics by starting level, then flip a coin inside each pair.
• Measure the brand name, not the cited pages. Schulte, Bleeker and Kaufmann (April 2026, German-language prompts, Swiss market) find brands steadier than sources from one day to the next: a Jaccard index of 0.45 to 0.59 against 0.34 to 0.42. Steadier does not mean fixed: in the audit by Uberti-Bona Marin and colleagues (117 queries repeated three times, September 2026), ChatGPT shares only 18.1% of exact pages across repeats, and overlap of recommended products reaches only 0.178. Only a rate aggregated over many answers holds up.
• Never read a single prompt. Schulte and colleagues recommend at least 7 runs per prompt per day to monitor a brand. With one answer per prompt, per engine, per week, as in the plan below, only the aggregate across a group of topics can be read.

What does a realistic test plan look like?

Twenty weeks, eight of them with no change at all. Illustration: Alba & Co is a fictional brand, and every parameter is an assumption to replace with your own. Alba tracks 120 prompts in 12 topics, across 4 engines, every week.

Power. The A/A test yields the number that decides everything: the standard deviation of the difference between pairs. Assume 6 points. With 6 pairs, the standard error is 2.45 points and the minimum detectable effect, at 5% two-sided and 80% power, is about 8.5 points. Alba has an 80% chance of seeing a move from 20% to 28.5% of answers. A real 4-point gain would most often come back inconclusive. Seeing 5 points would take about 14 pairs, or 28 topics.

Pre-registration. Before assignment, a one-page document signed by the CFO sets:

• One primary metric: the share of answers naming Alba, over all answers, including those naming no brand at all. A composite score or a ratio with a moving denominator can rise for other reasons, and prompts that already name the brand are excluded, because they inflate the score.
• The analysis: difference in differences per pair, with a paired t-test. A permutation test over the 64 possible assignments serves as a check: its minimum two-sided p, 3.1%, is reached only if all six pairs lean the same way.
• The decision rule: a gain significant at 5% (about 6.3 points under these assumptions) extends the program to the holdouts. A positive, non-significant gain earns one extension or more topics. A flat or negative gain sends the tactics back for review, without passing judgment on the channel.
• Exclusions: any known engine change date is logged, and any edit to a prompt invalidates its topic.

Reading discipline. Stopping at the first significant result produces false positives: in my simulation (6 topics against 6, 12 weekly reads, no real effect), over 20% of tests conclude wrongly, against 5% with a single read. Multiplying metrics does the same: 4 engines and 4 metrics make 16 tests, a 56% chance of at least one false positive if they were independent.

If leadership refuses to starve any topic. Use a staggered rollout: three groups start six weeks apart and all end up treated. Compare each wave only with groups not yet treated, because the classic estimator is biased when effects change over time (Goodman-Bacon, 2018). I found no reliable published lag between an action and its effect on answers: beyond six weeks, this design sees nothing, and the plan's 12 weeks remain a reasoned bet.

This test says nothing about sales, or about demand as a whole. When SparkToro collected 142 human-written prompts for a single need (January 2026), their semantic similarity was only 0.081. Your prompts are a sample you built, and to my knowledge no engine publishes the real distribution of questions.

How do you connect answers to demand?

With separate markets, and by stating the limit up front. Regions inside one country suit a content program poorly, since a published page can be read anywhere. Store networks raise a different question, covered in how to test local AI answers. You need countries with distinct languages: Spain against Italy rather than France against Belgium. Even then, English-language sources may feed both markets (my inference), and any leak pulls the result toward zero.

The number of markets caps the best result you can get, whatever the effect:

Author's calculation. The 1-in-N floor for one treated market is described by Lei and Sudijono (preprint revised in April 2025).

Few brands have 12 comparable markets. Most will get a direction, and the CFO has to accept that before the test starts.

The closest outcome is brand search, market by market. Since 20 November 2025, Search Console has offered a branded queries filter, opened to all eligible sites on 11 March 2026, and Google warns that its AI-assisted classification can make mistakes. Rand Fishkin (SparkToro, October 2026) argues that AI should be measured as a brand channel rather than a traffic channel. That is a reasoned opinion, and it moves the question without answering it: this brand signal has not yet been reliably detected.

Can revenue be tested directly?

Rarely, and the sums come before the spend. Compare the program's break-even with the smallest effect the test can see.

The test cannot confirm break-even, because it sees nothing below 5.9%. And that calculation is generous, since it assumes independent weeks. With an autocorrelation of 0.5, the threshold rises to 10.2%.

Measurable traffic weighs even less. In Ahrefs' panel (118,671 sites, October 2026), AI assistants make up 0.46% of tracked traffic, against 27.55% for search engines. If the program doubled those visits and each was worth an average visit, the gain would approach 0.46% of revenue, about €5,500, nearly thirteen times below the threshold. A positive result could only come from a large brand effect that referrals do not show.

None of this is unique to GEO. Lewis and Rao (Quarterly Journal of Economics, 2015) analyzed 25 US advertising experiments, most reaching millions of customers: the median confidence interval on return was over 100 percentage points wide. And they could randomize individuals.

A revenue test starts to make sense with enough markets. With 6 pairs and proportional cost, the threshold falls to about 3%, below break-even, and climbs back to 5.2%, break-even itself, with the same 0.5 autocorrelation. That takes twelve markets, with the program running in six.

What should you tell your CFO?

Three statements, in this order:

• What can be shown. Whether the work moves answers relative to a holdout, and by how much, within a margin announced in advance.
• What gives a direction. Brand search and AI traffic in treated markets against holdouts. Referrals in your analytics are a last-click floor.
• What stays unknown. The effect on revenue: I found no published study that measures it, and your own test probably will not either.

The budget consequence fits in one rule: fund in tranches, with the next one depending on the first link. Budgeting across GEO, SEO and paid covers the trade-off between channels. Then keep a few holdout topics permanently: in C-SEO Bench (NeurIPS 2025, simulated engines), gains shrink as others adopt the same methods.

That leaves the repetitive part: reading two groups of prompts, run after run, without mixing them. This is what Nanga's GEO module does. Prompts are organized into topics and tags; a "treated" tag and a "holdout" tag let you filter the views, group results by tag in the Overview and compare two periods, engine by engine. Tracking is weekly by default, daily as an option. A marker on the trend chart dates each completed action, and the latest run's answers export to CSV, to archive at each run so you can recompute your metric over your own denominator.

The Nanga team has seen why the control matters: after a model version upgrade, answers grew longer and cited more sources, with no brand having changed anything. Without a holdout, that jump reads as a win.

To read your treated and holdout topics side by side: request a demo.

Frequently asked questions

Sources:

• Olivier Martinez, Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026), preprint, July 2026.
• Keisuke Watanabe and Kazuki Nakayashiki, Disentangling Answer Engine Optimization from Platform Growth, preprint, June 2026.
• Faye Zhang et al. (Pinterest), Generative Engine Optimization: A VLM and Agent Framework for Pinterest Acquisition Growth, preprint, February 2026.
• Kevin Indig, Growth Memo, Does AI drive more direct traffic?, October 2026.
• Thomas Blake, Chris Nosko and Steven Tadelis, Consumer Heterogeneity and Paid Search Effectiveness: A Large Scale Field Experiment, NBER, 2014, published in Econometrica, 2015.
• Randall Lewis and Justin Rao, The Unfavorable Economics of Measuring the Returns to Advertising, Quarterly Journal of Economics, 2015.
• Lucas Uberti-Bona Marin et al., "If I Had to Buy Just ONE: Galaxy S26 Ultra": Auditing AI-Generated Product Recommendations, preprint, September 2026.
• Julius Schulte, Malte Bleeker and Philipp Kaufmann, Don't Measure Once: Measuring Visibility in AI Search (GEO), preprint, April 2026.
• Rand Fishkin, SparkToro, AIs are highly inconsistent when recommending brands or products, January 2026.
• Rand Fishkin, SparkToro, Is AI visibility a marketing boondoggle?, October 2026.
• Andrew Goodman-Bacon, Difference-in-Differences with Variation in Treatment Timing, NBER, 2018, published in Journal of Econometrics, 2021.
• Lihua Lei and Timothy Sudijono, Inference for Synthetic Controls via Refined Placebo Tests, preprint, January 2024, revised April 2025.
• Google Search Central, AI features and your website, accessed 7 November 2026.
• Google Search Central, Generative AI performance reports, June 2026.
• Google Search Central, Introducing the branded queries filter in Search Console, November 2025, updated March 2026.
• Ahrefs, AI vs Search Traffic Analysis, October 2026 data, updated 7 November 2026.
• Haritz Puerto et al., C-SEO Bench: Does Conversational SEO Work?, NeurIPS Datasets and Benchmarks, 2025.