AEO Contribution Margin: A CFO Framework for Defending the Budget When Cuts Hit

In April 2026, Notion's marketing analytics team published a remarkably candid post-mortem on their 2025 AEO program: the team had spent roughly $1.4M on AEO-optimized content, schema infrastructure, and citation engineering across the year, and the dashboard showed that branded search had risen 31 percent and demo requests from organic and AI-referred channels had risen 42 percent over the same period. When the CFO asked what fraction of that lift was caused by the AEO investment versus underlying brand momentum, product launches, and a competitor's churn event, the team could not answer. They had no holdout. They were reporting correlation as causation, and they knew it. The post-mortem, covered in the Stratechery interview and discussed widely in marketing analytics circles, became a wake-up call for an industry that had spent two years claiming AEO ROI without running the experiments that would prove it.

The pattern is everywhere now. AEO budgets ballooned from a rounding error in 2023 to a defined line item averaging 18 to 24 percent of marketing spend in mid-market B2B SaaS by Q2 2026, according to Forrester's marketing technology investment survey. And yet the measurement methodology that the discipline has converged on — correlation dashboards with before-and-after comparisons, share-of-citation tracking, and ad-hoc attribution claims — is structurally incapable of producing the causal evidence a CFO needs to defend the line item in the next planning cycle. The marketing teams that will survive the 2026-2027 budget compression are the ones running rigorous incrementality tests on their AEO investment. The teams that cannot prove incrementality will see their budgets reallocated to channels that can.

This piece is the operating playbook for running those tests. It draws on the methodology developed by Meta's Marketing Science team for Lift studies, Google's open-source GeoLift framework, decades of marketing-mix-modeling research from Kellogg's marketing department at Northwestern, and our own work helping six B2B brands design and run AEO incrementality experiments over the past 14 months. The mechanics are accessible; the discipline required to execute them correctly is the hard part.

Why Correlation Dashboards Fail in AEO

The default AEO measurement stack — Profound, Otterly, Bluefish, plus the company's GA4 instance and CRM data — produces a story that looks like evidence and is not. The story goes: in Q1, we published 30 AEO-optimized articles, our citation rate on ChatGPT rose 18 percent, branded search lifted 12 percent, demo requests from organic channels lifted 14 percent. Therefore, AEO is working and we should double the budget.

Every step in that chain is a correlation, not a causal claim. Citation rate rose — but was it because of the 30 articles, or because the entire category got more searchable through AI assistants as adoption grew? Branded search lifted — but was it because of the AEO program, or because the founder did a Lex Fridman podcast that week, or because a competitor announced a price increase that drove comparison searches? Demo requests rose — but in a quarter when the sales team also added two BDRs, the product released a major feature, and the macro indicator on B2B software spend ticked up, how much of that lift is the AEO investment causing and how much is everything else?

The honest answer is that without a counterfactual — what would have happened in the absence of the AEO treatment — none of those questions are answerable from the dashboard. The whole point of incrementality testing is to construct that counterfactual through experimental design, not after-the-fact regression on observational data. Kellogg's marketing measurement coursework has been hammering this point for thirty years, and the same lesson is now arriving in AEO with a generation's delay. Observational dashboards can describe what happened. They cannot tell you what caused it.

The cost of getting this wrong is significant. A team that doubles down on AEO because the dashboard showed a 31 percent lift, when the true incremental contribution of AEO was 4 percent and the rest was brand momentum and product factors, has just over-invested in a channel by an order of magnitude. The dollars that went into another 60 AEO articles could have gone into product marketing, paid acquisition, or sales enablement with much higher actual returns. The dashboard story protected the AEO budget. It did not protect the business.

The Three Experimental Designs That Work for AEO

Three experimental designs translate cleanly from the paid-media incrementality playbook into AEO. Each has different operational requirements, different statistical properties, and different failure modes. The right choice depends on the buying motion, the AEO surfaces being tested, and the analytical infrastructure available.

Geo-holdout is the design that translates most directly from Meta and Google's paid-media frameworks. You suppress the AEO treatment in a randomly selected subset of geographic markets and apply the full treatment in the others. The methodology comes from Google's GeoLift open-source library on GitHub, which was developed in collaboration with Meta to enable rigorous geo-experiments for marketing measurement. The advantage is that randomization at the geo level gives you a clean counterfactual without needing user-level identifiers — you compare aggregate outcomes in treated geos against the synthetic control built from untreated geos. The disadvantage in AEO specifically is that LLM citations and SEO surfaces do not respect geo boundaries cleanly. A piece of content that ranks in your treated geos will also surface in your control geos through global AI assistants. Without aggressive geo-targeted content suppression and platform-level controls, network spillover contaminates the cell.

Content-cohort holdout is the design we recommend most often for content-heavy AEO programs. The mechanic is to publish a batch of 30 to 60 articles within a tight time window and randomly assign each article to either a treatment cohort (full AEO optimization: schema markup, FAQ blocks, llms.txt inclusion, citation engineering, internal linking, distribution amplification) or a control cohort (baseline editorial production only). Measure the differential in AI citation rate, organic traffic, AI-referred traffic, and downstream conversions across the two cohorts at 4, 8, 12, and 24 weeks. The unit of randomization is the article, which means you can run a properly powered test on a single product line without splitting your sales territory. The disadvantage is that revenue attribution back to specific articles requires good last-touch and journey data — which most companies do not have for AI-referred traffic, given the broken referrer landscape covered in the dark-funnel attribution playbook.

Product/feature-page holdout is a narrower design useful for SaaS companies with discrete product or feature pages where the AEO treatment can be applied or withheld at the page level. Randomly split a set of comparable feature pages — say 40 pages within a single product category — into treatment and control arms. Treat the treatment arm with full AEO infrastructure: structured definitions, comparison tables, FAQ schema, internal linking to related concepts, llms.txt inclusion. Leave the control arm at the baseline marketing-page treatment. Measure differential citation rate, page-level AI-referred traffic, and downstream pipeline contribution from each cohort. The advantage is shorter run length and easier attribution; the disadvantage is that cross-page traffic recirculation muddies the result if a user lands on a treatment page and converts on a control page or vice versa.

Pre-Test Power Calculation: The Step Everyone Skips

The single most common mistake in AEO incrementality testing is skipping the pre-test power calculation. Teams design a test, pick a run length that feels right, launch it, and then discover at the end that the test was so underpowered that any effect smaller than 40 percent was undetectable — which means the test could not have found a realistic AEO lift even if one existed. The whole exercise produces a null result that is interpreted as no effect when it actually means insufficient data.

The math is not complicated. Given the baseline variance of your primary success metric (weekly branded search volume, weekly demo requests, weekly pipeline-qualified leads), the desired minimum detectable effect (the smallest lift you would care about, typically 5-10 percent for AEO), the desired statistical power (conventionally 80 percent), and the desired significance threshold (conventionally p < 0.05), you can compute the required sample size and run length.

For a representative B2B SaaS company with 800 weekly demo requests, a target minimum detectable effect of 7 percent, 80 percent power, and a 5 percent significance threshold, the required run length is approximately 11 weeks for a 50/50 geo-holdout split — assuming the geos are balanced on baseline and you have a clean synthetic-control construction. For a smaller company with 200 weekly demo requests, the same test would need 22 weeks or a larger minimum detectable effect to be powered. The companies running 4-week AEO tests with 100 weekly conversions and claiming they detected lift are reporting noise, not signal.

The power calculation also forces a useful conversation up front about what effect size would justify the AEO investment. If the AEO budget is $200K per quarter and the company needs the AEO program to generate $1M in pipeline to clear the ROI hurdle, that implies a specific minimum detectable effect against baseline pipeline volume. If that effect size is below the level the test can detect, either the test design needs to change (longer run, larger sample, different metric) or the investment thesis needs to be reconsidered before the test even starts.

The Treatment Definition Problem

In paid-media incrementality, the treatment is unambiguous: campaign X ran in treated cells and did not run in control cells. In AEO incrementality, the treatment is much harder to define cleanly, and the precision of the treatment definition is the second most common failure mode after underpowered tests.

What exactly are you treating? A reasonable AEO treatment definition might include all of the following: schema markup (FAQPage, HowTo, Article, Organization), FAQ blocks at the bottom of every article, inclusion in the llms.txt manifest, citation engineering (quotable statistics, declarative definitions, named-author bylines, methodology footnotes), internal linking to and from the article, and distribution amplification (LinkedIn, podcasts, newsletters). All six are legitimate AEO interventions, and most teams apply them together as a bundle.

But if your test is "AEO bundle versus no AEO bundle," and the bundle works, you do not learn anything about which components of the bundle drove the effect. The result tells you that AEO works in aggregate, which is useful for the CFO conversation but not useful for budget allocation across the AEO surfaces. A more sophisticated test design treats the bundle as a multi-arm experiment with each component tested separately or in factorial combinations. A 2x2x2 design with three components — schema, FAQ blocks, distribution — gives you eight cells and allows you to estimate the marginal contribution of each. The sample-size cost is real (eight cells require roughly 8x the per-cell sample of a two-arm test), but the analytical payoff is substantial.

For most teams running their first AEO incrementality test, we recommend starting with the simpler two-arm bundle test to establish that AEO works at all in aggregate. Once that result is in hand, the second round of testing can decompose the bundle into components. Trying to run a sophisticated multi-armed test before establishing baseline incrementality often produces null results across all arms that are uninterpretable because the bundle itself was not validated.

Measurement: Beyond the Citation Rate

The primary success metric for an AEO incrementality test cannot be citation rate alone, even though citation rate is the most obvious AEO-native KPI. Citation rate is a leading indicator of revenue but not a substitute for it. A test that shows a 40 percent lift in citation rate but no measurable lift in demo requests or pipeline is either too underpowered to detect the downstream effect, suffering from broken attribution between citation and conversion, or revealing that the citations are not converting — which is itself an important finding.

The measurement framework we use for AEO incrementality tests has four metric layers, ordered from leading to lagging.

Layer 1: AI citation rate. Measured weekly across ChatGPT, Claude, Perplexity, and Gemini for a fixed query set of 200 to 500 head-term and long-tail queries relevant to the test cells. Treatment cells should show a measurable citation rate lift starting in week 2-4. If they do not, the AEO treatment is not being picked up by the models — likely a content quality, schema rendering, or crawler accessibility issue that needs to be diagnosed before the test continues.

Layer 2: Branded search and unbranded search. Measured weekly via Google Search Console, segmented by treatment cell where possible. Branded search lift is the canonical second-order signal of AI citation impact — when AI assistants mention your brand more often, downstream branded searches increase as users seek to validate or learn more. Unbranded search lift is the riskier indicator because category-level search is influenced by too many factors to attribute cleanly to AEO.

Layer 3: Demo requests and pipeline contribution. Measured weekly via CRM, segmented by attribution source where possible. This is the metric layer the CFO actually cares about, and the layer where AEO attribution gets messiest. Layer 3 lift typically lags Layer 1 by 60 to 120 days because of buyer journey latency, which is why the test run length matters. A 4-week test will catch Layer 1 lift but miss Layer 3 lift entirely.

Layer 4: Customer survey lift on source attribution. Measured via post-purchase or post-demo survey asking customers where they first heard about the company. This is the methodology that finally cuts through the broken referrer attribution problem — covered in detail in the multi-touch attribution playbook for the AI search era — and it is increasingly the only way to capture AI-search-influenced acquisition that does not show up in any deterministic tracking. Run the survey in both treatment and control cells, compare the percentage of customers citing AI assistants or AEO-content channels, and use the differential as the survey-based incrementality estimate.

The four layers should converge directionally. A test where Layer 1 shows strong lift but Layer 2-4 show none is suspicious — possibly an artifact of the test or evidence that citations are not driving downstream behavior. A test where Layer 1 shows weak lift but Layer 3-4 show strong lift is also suspicious — possibly indicating that the treatment is driving something other than AEO citations (better content, better distribution) or that the citation tracking is undercounting.

The Bot Traffic Contamination Problem

Every AEO incrementality test in 2026 runs into the same analytical pitfall: AI crawler traffic from GPTBot, ClaudeBot, PerplexityBot, Anthropic-Search, Google-Extended, and a dozen others inflates the apparent organic traffic in treated cells without producing any real buying signal. The crawlers are doing exactly what they should be doing — discovering, indexing, and re-crawling AEO-optimized content at a much higher rate than baseline content. But to the analytics dashboard, that traffic looks like sessions, and if you do not filter it out, you will overstate the apparent traffic lift in your treatment cell by 15 to 40 percent.

The remediation has three components.

1. Server-log-level bot filtering. The GA4 default bot filter does not catch the modern AI crawler fleet. You need server-log analysis or a CDN-level filter that identifies and excludes the user agents and IP ranges of the major AI crawlers before the data hits your analytics layer. Most teams underestimate the volume — for a content-heavy site running an AEO program, AI crawler traffic can easily reach 25 to 40 percent of total raw sessions by Q2 2026.

2. Separate reporting of human and bot traffic. Even with filtering applied to the primary dashboard, you want a separate view of crawler activity because crawler volume is itself a meaningful AEO leading indicator — a piece of content getting hit hourly by GPTBot is signal that the content is being actively used in citation lookup, which is the precursor to citation lift in user-facing responses. Filtered out of the primary dashboard, surfaced in a secondary one.

3. Conversion-funnel sanity checks. A treatment cell that shows a 30 percent traffic lift but only a 5 percent demo-request lift is suspicious — either the traffic lift is bot-contaminated, the traffic is from low-intent queries, or the conversion path is broken. The diagnostic is to compute the per-session conversion rate in both cells and look for divergence. Healthy human traffic should convert at similar rates across treatment and control. Diverging conversion rates almost always indicate measurement contamination.

A 7-Step AEO Incrementality Test Playbook

The following playbook is what we use with every team designing their first AEO incrementality test. Each step has been a recurring failure point in tests we have seen run without it.

1. Define the investment thesis and target ROI before designing the test. Write a one-page memo that states exactly what AEO investment is being tested, what revenue or pipeline outcome would justify continuing the investment, and what minimum detectable effect on the primary success metric is consistent with that outcome. This memo forces the team to commit to a specific success criterion before the data starts coming in, which prevents the post-hoc rationalization that destroys experimental discipline.

2. Run a pre-test power calculation. Given the baseline variance of the primary success metric, the target minimum detectable effect from step 1, 80 percent power, and a 5 percent significance threshold, compute the required sample size and run length. If the required run length exceeds 24 weeks, reconsider the test design — either the effect size is too small to detect with available sample, or the test needs to be redesigned with a more sensitive primary metric.

3. Pre-register the experimental design. Document the holdout cell selection method, the treatment definition, the primary and secondary success metrics, the planned run length, the planned analysis method, and the stopping rules. Save the document with a timestamp before the test launches. Pre-registration is the single most effective discipline against the post-hoc analytical choices that inflate false-positive rates by 3 to 5x against the nominal significance threshold.

4. Launch the treatment and instrument measurement. Apply the AEO treatment to the treatment cells, suppress it from the control cells, and confirm at the end of week 1 that the assignment is being honored — no leakage of treatment into control or vice versa. Run a sample-ratio mismatch check on the traffic distribution between cells; if the observed split diverges from the planned split by more than 2 percent, halt and diagnose before continuing.

5. Monitor leading indicators weekly without making decisions. Watch Layer 1 (citation rate) and Layer 2 (branded search) on a weekly basis to confirm the treatment is being picked up by the AI assistants and surfacing in user behavior. Resist the temptation to declare success or failure based on early data — Layer 3 and Layer 4 effects lag by months, and early peeks on noisy data lead to bad decisions.

6. Run the full pre-registered analysis at the pre-registered end date. Compute the lift in each metric layer, the confidence interval, and the p-value for the primary success metric. If the result is significant, the lift estimate is the incrementality finding. If the result is not significant, the test is null — which is also a finding, and a more honest one than the alternative of cherry-picking metrics until something is significant.

7. Extend or replicate before changing the budget. A single significant result is the start of an evidence base, not the end. The teams running rigorous AEO measurement programs treat every incrementality test as one data point in a sequence and run replications across product lines, time periods, and treatment definitions to build the evidence base that supports a budget conclusion. One test with a 15 percent lift is interesting. Three tests with consistent 10 to 18 percent lift is a budget defense.

Pitfalls in the Wild: What Goes Wrong

We have seen each of the following failure modes in AEO incrementality tests over the past 14 months. Each is preventable with the right design, but each is endemic in tests run without explicit attention to the failure mode.

Network contamination across geos. A B2B SaaS company we worked with ran a clean geo-holdout test in EMEA, with the UK and Germany as treatment cells and France and the Netherlands as control. Three weeks into the test, the company's US PR team published a press release about a new product that was picked up by global trade media. The press release was cited by AI assistants in both treatment and control geos, contaminating the control cell with effective AEO treatment. The test was unsalvageable; the team had to relaunch with stricter cross-functional coordination and a longer pre-test communication freeze on the marketing calendar.

Sample-ratio mismatch from broken bot filtering. Another team ran a content-cohort holdout where the treatment cohort received llms.txt inclusion and the control cohort did not. The week-1 sample-ratio check showed that the treatment cohort was receiving 2.3x the bot traffic of the control cohort — exactly as expected, because llms.txt inclusion brings the crawlers — but the team's analytics layer was including bot traffic in the session count. The apparent traffic lift in the treatment cohort was almost entirely bot traffic, and the team initially declared the test a success before catching the bug in week 4. After re-filtering, the actual human traffic lift was 6 percent — still positive, but a tenth of the apparent lift.

Post-hoc metric switching to chase significance. A third team pre-registered demo requests as the primary success metric. The test ran 12 weeks; demo requests showed a 3 percent lift with a wide confidence interval that crossed zero. The team then computed lifts across 14 other metrics — branded search, page views, scroll depth, email signups, podcast downloads, etc. — and found that one of them (newsletter signups) showed a 22 percent significant lift. The team reported the newsletter lift as the headline finding. This is the classic garden of forking paths problem in marketing measurement; with 14 metrics tested at p < 0.05, you would expect 0.7 false positives just by chance. The post-hoc reporting destroyed the experimental discipline that justified the test in the first place. Pre-registration prevents this.

Ignoring lagged effects. Two teams we worked with ran 6-week and 8-week AEO incrementality tests against pipeline-qualified leads as the primary metric. Both tests showed null results and the teams declared AEO to have no measurable incrementality. Both teams then extended the measurement window to 16 weeks for the same cohorts (without changing the treatment) and found significant Layer 3 lift that had not yet emerged at the original test endpoint. The lesson is that the buyer journey latency for AEO-influenced pipeline is long enough that short tests systematically understate the true effect. Plan for it in the run length, or accept that you will measure leading indicators only.

Confounding from concurrent treatments. A common mistake is launching an AEO incrementality test in the same quarter as a major paid acquisition campaign, a brand launch, or a product release. The treatment in the AEO test is no longer isolated — the geo-holdout cells are also seeing differential exposure to the other concurrent initiatives. The cleanest tests run in quiet operational periods or use sufficiently aggressive randomization to balance the concurrent effects across cells, but most teams cannot create the conditions for a quiet period. The next-best alternative is to model the concurrent treatments as covariates in the analysis, which adds complexity but salvages interpretability.

The CFO Conversation

The whole point of running an AEO incrementality test is to enable a defensible budget conversation with the CFO and the rest of the executive team. The form that conversation should take, once you have a result in hand, is structurally different from the dashboard conversation that preceded it.

The dashboard conversation says: AEO investment was $X, AI citations rose Y percent, branded search rose Z percent, demo requests from organic rose W percent. The implicit claim is that the AEO investment caused all of W percent of the demo request lift. The CFO discounts the claim by a factor of 2 to 5x — which is roughly the right discount given the lack of counterfactual — and the conversation ends with a budget cut.

The incrementality conversation says: AEO investment was $X. We ran a pre-registered geo-holdout test with $Y of that spend across Q1 and Q2, with N treatment geos and M control geos. The test was powered to detect a 7 percent lift in pipeline-qualified leads at 80 percent power. The observed lift was 11 percent with a 95 percent confidence interval of 4 to 18 percent. The implied incremental pipeline contribution from the tested AEO spend is $Z, which produces a payback period of P months against the tested investment. We are recommending continued investment at the current level with the next test focused on decomposing which AEO surfaces drove the lift.

The second conversation is roughly 10x more defensible than the first because it produces a causal estimate, a confidence interval, and a payback math that can be inspected. It is the conversation the marketing teams who keep their AEO budgets are having. It is also the conversation that the CMO AEO dashboard playbook for the board deck is built around, and the analytical foundation for the AEO ROI payback period calculation framework for CFOs.

The discipline required to run the second conversation is substantial. It requires giving up the comfortable dashboard story that makes the AEO program look good in the short term. It requires accepting the possibility that a clean test will return a null result, and that null result will trigger a hard conversation about whether the AEO investment is justified. It requires the cross-functional coordination to suppress treatment in the control cells, the analytical sophistication to filter bot traffic and run the correct statistical tests, and the operational patience to wait 12 to 16 weeks for the result to mature.

The teams who do this work are the ones whose AEO budgets compound through 2027 and 2028. The teams who refuse to do this work — who hide behind correlation dashboards and post-hoc rationalizations — are the ones whose AEO budgets get cut in the next planning cycle when the CFO finally asks the question the dashboard cannot answer.

Takeaway: AEO incrementality testing is not optional infrastructure for any marketing team with a meaningful AEO budget in 2026. The correlation dashboards that defined the early AEO era are running out of credibility, and the CFOs who funded the experiment are starting to ask for causal evidence. The methodology to produce that evidence — geo-holdouts, content-cohort holdouts, product-page splits — is well understood from decades of marketing-mix-modeling research and has been formalized in Meta's Lift framework and Google's GeoLift library. The mechanics are accessible, but the experimental discipline they require is exactly the discipline most marketing teams have spent the last decade avoiding. The teams that learn it now will compound their measured AEO advantage through 2028. The teams that do not will see their AEO budgets reallocated to channels whose ROI they can actually prove.