The difficult part is attribution. Many people read an AI answer without clicking. A later branded search, direct visit, sales conversation, or purchase may have been influenced by the answer, but the evidence rarely supports a perfect user-level chain.
Use a layered model that reports what is directly observed and labels what is inferred.
The AEO value chain
Published action → page or source becomes available → AI answer mentions or cites it → user exposure or click → qualified behavior → pipeline or revenue outcome
Each arrow is a separate measurement problem. A page publication is not proof of adoption; a citation is not proof of a click; and a click is not proof of revenue.
Layer 1: execution evidence
Track the work delivered:
- pages created, consolidated, or updated;
- source contributions published;
- technical issues resolved;
- Prompt Panel and measurement configuration;
- publication dates, versions, and acceptance checks.
Output volume is not ROI, but it establishes what changed and when.
Layer 2: answer visibility
Measure successful, evaluable responses using stable definitions:
- Mention Rate;
- Visibility and TOP3;
- Sentiment;
- description accuracy;
- Citation Rate and Official Citation Rate;
- competitor answer and source gaps.
These are leading indicators. They show whether public evidence is being used, not whether revenue has already followed.
Layer 3: traffic and engagement
Combine analytics and search data to review:
- identifiable AI referrals;
- cited landing-page sessions;
- branded search and direct traffic trends;
- engagement and assisted conversions;
- market, device, and page mix.
Document changes to referrer labeling, consent, channel grouping, and tracking configuration before interpreting a trend.
Layer 4: commercial outcomes
Use business measures appropriate to the funnel:
- qualified leads;
- product signups or trials;
- assisted pipeline;
- influenced revenue;
- customer acquisition cost;
- retention or support deflection where relevant.
Do not assign a revenue amount to an AI mention unless the attribution method supports it.
An honest ROI formula
AEO ROI = (attributable or modelled value − AEO cost) / AEO cost
Report the value class:
- directly attributable: a supported event chain exists;
- assisted: AEO exposure or referral participated in a broader journey;
- modelled: the estimate uses documented assumptions;
- directional: the signal is useful but cannot support a financial claim.
Publishing one blended number without these labels makes the result look more precise than the evidence.
Design a useful AEO experiment
- Select a defined question cluster and page or source owner.
- Establish a baseline with a frozen prompt set.
- Change one meaningful content or evidence unit.
- Record the publication version and other known events.
- Allow an appropriate observation window.
- Repeat the same collection conditions.
- Review answer, citation, traffic, and commercial layers separately.
- Decide whether to scale, revise, or stop.
A non-result is valuable if the test was attributable. It prevents continued investment in an unsupported tactic.
Measure AEO with UnderAI
UnderAI GEO Workspace provides the answer-evidence layer: Prompt Groups, daily results, Visibility, TOP3, Sentiment, Answer Snapshots, competitors, and Citations. UnderAI Monitoring & Retesting preserves panel versions and compares compatible observations after implementation.
Your analytics, CRM, and finance systems remain the source of truth for visits, pipeline, cost, and revenue. UnderAI should be connected to that decision process without pretending to replace it.
Build an AEO measurement plan or review AI Search Visibility Metrics.
Worked example: why visibility progress is not yet ROI
UnderAI’s July 2026 baseline created a traceable answer-and-source starting point. The complete counts belong in the Citation benchmark; the ROI page uses the baseline only to show where financial attribution begins and ends.
Suppose the next comparable run produces 12 correct UnderAI mentions and four official-site Citations. That would be answer-level progress. It would not yet prove return on investment.
To calculate ROI, connect the change to commercial evidence:
| Layer | Example measure | What it can prove |
|---|---|---|
| Execution | Four priority pages shipped | The planned work happened |
| Answer | 12 correct mentions; four official Citations | AI representation changed in the measured Panel |
| Traffic | Qualified sessions from AI or Search landing pages | Users visited after exposure, subject to attribution limits |
| Commercial | Leads, opportunities, or revenue associated with those sessions | Business contribution under the chosen attribution model |
| Cost | Software, research, content, technical, and source-development cost | Investment denominator |
If the program creates answer gains but no measurable traffic, the team should investigate discoverability, click appeal, and journey design. If traffic rises without qualified outcomes, the next problem is conversion or audience fit—not automatically AEO execution.
A practical reporting boundary
Google reports AI feature performance through Search Console, while OpenAI adds utm_source=chatgpt.com to referral URLs from ChatGPT search. These sources improve attribution but do not establish perfect causality. Use Google’s AI feature guidance and the OpenAI publisher FAQ when defining the measurement plan.
Frequently asked questions
How soon can AEO ROI be measured?
Leading indicators may change before commercial outcomes. Set separate observation windows for publication, answer adoption, traffic, and revenue rather than promising one universal deadline.
Is visibility improvement an ROI result?
It is evidence of answer-level progress. ROI requires a value model and cost comparison.
How should no-click influence be valued?
Use assisted or modelled analysis with explicit assumptions. Do not present it as direct attribution.
What costs belong in the denominator?
Include relevant software, research, strategy, content, technical work, external source development, and ongoing measurement costs.