Answers
How do you prove AEO work caused the AI citation lift, not just coincidence?
Short answer
How do you prove AEO work caused the AI citation lift, not just coincidence?
With a pre-registered, held-out control arm and enough queries to clear a measured noise floor, not with a single before-and-after reading. Assistant answers are unstable between identical asks, our own step-zero run found 7 of 12 queries changed outcome across five identical repeats, so a raw before-and-after comparison cannot separate real movement from the model changing its own mind. The fix is a difference-in-differences design: a treated group that gets the work and a control group that does not, both measured at the same two timepoints. Our own permutation test over 20,000 random splits with zero intervention applied found the null difference centers on a mean of roughly plus or minus 0.0016 with a standard deviation of 0.215, which means unworked noise cancels across a large enough sample and a real lift stands out from it.
The short version
Null-case check across 20,000 splits: unworked noise centers on roughly zero.
01
Why a before-and-after reading is not proof
The most common way AEO results get reported is a visibility score before an engagement and the same score after, presented as the lift. That comparison has no way to separate work-caused movement from the model itself changing its answers between the two reads, because nothing in a single before-and-after design accounts for the underlying instability. Our own instrument check found 58.3 percent of tracked queries, 7 of 12, returned a different outcome across five identical repeats with nothing else changed.
A number that moves for reasons unrelated to the work looks identical to a number that moves because of it. Without a comparison point that isolates the work, a rising score and a falling one are both compatible with having done nothing at all.
20,000 splits, zero intervention applied
Confirming the untreated case centers on zero is what makes a difference-in-differences result trustworthy.
02
The design that actually isolates the effect
A difference-in-differences design fixes this by adding a second, deliberately untouched arm: the same class of queries, measured at the same two timepoints, with no placement work applied. Whatever the untouched arm moves by is the model's own churn for that window, and subtracting it from the treated arm's movement leaves the part attributable to the work.
We validated the null case directly before trusting this design on real work: running 20,000 random splits with no intervention applied at all, the resulting difference centered on a mean of roughly plus or minus 0.0016 with a standard deviation of 0.215. That is the honest finding, the unworked noise is unbiased and it cancels at scale, which is exactly the property a difference-in-differences design needs to be trustworthy.
Difference-in-differences, two timepoints
The control arm's movement is subtracted from the treated arm's. What is left is attributable to the work.
03
What sample size the proof actually needs
Isolating a real effect from that noise floor costs samples, and the cost is steep at small scale. Computed from the variance in our own run, a 12-query by 5-repeat design can only detect a change of 51.1 percentage points, which is unusably coarse. A 20 by 20 design reaches 21.4 points, and a 40 by 40 design, roughly 3,200 calls and about six hours per timepoint, reaches 9.9 points.
That is the honest price of a provable answer rather than an asserted one. A vendor quoting a lift number from a small query set is either reporting something too coarse to mean anything or has not run the arithmetic that would tell them so.
Minimum detectable lift, our own design
A vendor quoting a lift number from a small query set is reporting something too coarse to mean anything.
Asked next
The questions that follow this one
Is a difference-in-differences design overkill for a small business?
It scales down, not away. The design principle, treated versus untouched, measured at two timepoints, holds at any size; what shrinks is the minimum detectable lift, which is why our entry tier reports the causal-lift number without attaching a guarantee to it.
Why does the noise cancel instead of biasing the result?
Because across 20,000 random splits with no work applied, the null difference centered on roughly zero rather than drifting positive or negative. An unbiased noise source averages out as the sample grows, which is the statistical property the whole design depends on.
Can I run this measurement myself before hiring anyone?
Yes, the design is not proprietary. Pick a query set, split it into a treated and control half, apply work to only one half, and measure both at two timepoints. The hard part in practice is discipline: not touching the control arm, and running enough queries to clear the noise floor.
Why is there no guarantee below the 10,000 dollar tier?
Instrument power, not confidence. At 20 to 30 queries the minimum detectable effect is 11.4 to 14 percentage points against a plausible real lift of 5 to 15, so a genuine win could read as statistically indistinguishable from zero. The causal-lift number is still reported at every tier.
Could PR, a launch or a review-site push have caused the lift instead?
Anything that reaches both arms at once, a launch, a pricing change, a model update, a wave of press, moves the control arm too and is subtracted out. The real risk is an event that only touches treated queries, such as earned media or a review-site push aimed at the same topics as the work. So outside events in the window have to be logged against the queries they could reach, and a treated query one of them hit is reported on its own rather than counted as the work's lift.
How do you turn a citation lift into a revenue or ROI number?
Not by multiplying citations by a guessed value per mention. A proved citation lift says the work changed what assistants answer; whether that reached pipeline is a second measurement, read from visits that arrive from assistant domains in your analytics, a how-did-you-hear-about-us field on signup and on sales calls, and the deals those leads became. Keep the two claims apart: a report that converts a visibility score straight into dollars has skipped the step that shows whether buyers acted on it.
Can an AEO platform prove that new content caused the lift?
Not on its own. A visibility platform measures how often an assistant names you across a prompt set, and that is the right raw material, but a rising line on its chart is a before-and-after reading with the same weakness as any other: assistant answers move between identical asks, and the model, its retrieval index and your competitors all change in the same window as your new content. Proof needs a design around the measurement. Split the prompts into a treated arm, the questions the new content targets, and a control arm the work leaves alone; measure both arms at the same two timepoints; and read the lift as the treated change minus the control change. If a platform can export per-prompt results, you can run that comparison on its data. If it can only show a blended score across every prompt, it cannot separate your content from the model's own drift, however good the dashboard looks.
What counts as a meaningful brand mention?
One that a buyer would act on. A meaningful mention is the assistant recommending you, by name, in answer to a question someone asks when they are choosing a tool: a best-of question, an alternatives question, a how-do-I-solve-this question in your category. A passing reference in a list of twenty, a mention only when the prompt already contains your name, or a citation of your docs for a definition with someone else recommended above it all count on a raw mention tally and move nobody. So the prompt set is built from money questions taken from sales calls, support tickets and community threads, each prompt is disambiguated so it cannot be read as a different question, and each is asked five identical times, with the outcome scored on whether you are recommended in the majority of those repeats. Branded prompts are tracked separately, because an assistant naming you when asked about you proves recall, not recommendation.
What should a share-of-voice and brand mention trend dashboard show?
The arms, not just the line. A share-of-voice trend across every tracked prompt blends the questions your work touched with the ones it never reached, so a real gain on the treated questions can be hidden by churn elsewhere, and a model update can lift everything at once and look like a win. A dashboard that supports a causal claim shows the treated prompts and the control prompts as two separate series, marks the date the work shipped, and shows the noise band, the movement you would expect from repeat asks with nothing changed, so a reader can see whether the gap between the arms clears it. It also keeps branded and unbranded prompts apart, and lists which competitor took each answer you did not win, because a share-of-voice percentage alone never says who you lost to or on which question.
How long after publishing should we measure brand mention lift?
Long enough for the assistants to read the new material, and decided before the work starts. Answers that retrieve from the live web can pick up a new page or a new third-party mention within weeks, while anything that depends on a model's training data waits for the next model release, so a fixed date on the calendar tells you less than the source of each citation does. What matters for proof is that the second timepoint is chosen in advance and applied to both arms, because choosing it after looking at the numbers lets you stop on a good day. Our engagements pre-register the read for day ninety and report the causal-lift number then, whatever it shows. The page on how long AEO takes to work covers the timeline question on its own.
What data should we export before choosing an AEO platform?
Enough to redo its arithmetic without it. Before you commit to a platform, export a sample of its raw results and check that each row carries the exact prompt text, the engine and model it ran against, the timestamp, the full answer, whether and where your brand appeared, and the URLs the answer cited. Check that repeats of the same prompt come through as separate rows rather than one averaged score, because the repeat-to-repeat spread is the noise floor every lift claim has to clear. Then check you can tag prompts into a treated and a control arm and pull both at two dates. A platform that only exports a weekly blended score is a monitoring tool, which is useful, but it cannot answer whether your work caused a change, and you will find that out at the moment you most need the answer.
What is the lowest-cost way to measure AEO lift credibly?
Spend on queries and repeats, not on seats. The cost of a credible measurement is set by how small a lift you need to detect: from the variance in our own run, a 12-query by 5-repeat design can only detect a change of 51.1 percentage points, a 20 by 20 design reaches 21.4, and a 40 by 40 design, roughly 3,200 calls and about six hours per timepoint, reaches 9.9. Those calls are the real bill, and they cost the same whichever platform sends them. So the lowest-cost credible setup is the smallest design whose detectable lift is below the lift you expect, run twice, with the control arm left untouched. A cheaper plan that tracks fewer prompts once a week is not a cheaper version of the same measurement; it is a measurement too coarse to tell a real lift from noise.
What should a causal-lift report contain so we can check it ourselves?
Everything you would need to reach the same number without us. That means the prompt list split into its treated and control arms, fixed before the work began, with the date it was fixed; the engines and models measured and the dates of both timepoints; the per-prompt outcome at each timepoint across all five repeats, not a summary score; the treated change, the control change and the difference between them; and the minimum detectable lift the design could see, so a small result is read against what the measurement was able to find. It should also say what happened in the window that could have moved the numbers on its own, such as a model release, a launch or press coverage, and which prompts each event could reach. A report that hands over less than that is asking you to trust the conclusion, which is the thing a causal design exists to make unnecessary.
Does it matter which page gets cited, your product page or someone else's?
For the proof, no: the outcome measured is whether the assistant recommends you for the query, wherever its citation points. For the plan, yes, because most citations point off-site. Across 901 citation links in 60 answers we measured (one engine, July 2026), 91.5 percent pointed at pages the brand does not own, so work aimed only at your own product page leaves most of what assistants read untouched.
Building your own design
Free gap report
When the model answers,
be the one it names
Tell us your category and the questions your buyers ask. We run them against live AI answers and walk you through what came back. If you are already winning, we will tell you that too.
12 queries · 5 repeats each · 5 working days
Free · Walked through live · Five working days
Queries we run
Cited instead of you
reddit.com38%
g2.com26%
news.ycombinator.com21%91.5% of citations point off-site