ROI (%) = (incremental contribution before programme costs − programme costs) ÷ programme costs × 100.
State whether the cost basis is additional cash or combined cash and capacity, and use a consistent evaluation period.
What investment are you measuring?
Define the programme before selecting a dashboard: improving discovery in AI-generated answers, earning relevant consideration and helping buyers reach a useful next step. Specify the products, markets, pages, services and period covered. Record the alternative use of the budget.
This differs from measuring whether an assistant makes marketing production more efficient. It also differs from an attribution model that happens to use machine learning. Buying an AI attribution tool does not establish which customers discovered you through AI search. Our separate AI-marketing ROI Answer covers workflow productivity and full-cost measurement.
Include research, writing, technical repairs, distribution, external fees, tools, internal review and maintenance. Record internal capacity separately from additional cash spending. Allocate setup over a declared decision horizon for planning, while preserving the actual cash dates. Include the work that failed, not just the pages that attracted visitors.
Which evidence belongs in the report?
Use four connected records without assuming they form a perfectly observed funnel:
- Exposure and representation. Which relevant questions surface the business, whether the description is accurate and which URLs are cited.
- Visits. Recognised unpaid AI referrals, landing pages and meaningful subsequent activity. Keep paid assistant placements separate.
- Commercial outcomes. Qualified enquiries, accepted opportunities, orders, refunds and contribution, reconciled with the CRM or order system.
- Incrementality. Evidence about what would have happened without the programme.
Google's dedicated generative-AI reports show impressions and page-level visibility, with country, device and date dimensions. Those figures are already included in overall performance; adding them again inflates exposure. They are not an order ledger. Google's reporting announcement
Bing's AI Performance report records visible citations across supported Microsoft and partner experiences. It cannot establish the resulting sales or why a change occurred. Its coverage is a defined slice of AI activity. Bing's report definitions
How do you connect visits with outcomes?
GA4 now has an AI Assistant channel for recognised assistant referrers. Check it alongside source/medium and the landing page, rather than treating every assistant visit as ordinary organic search. A missing referrer still leaves an unknown journey. Google Analytics' May 2026 product update
Use matching attribution scopes. A session's source, the source that first acquired a user and credit assigned to an outcome are different measures. GA4's AI Assistant channel excludes Google's AI Overviews and AI Mode; those non-ad search visits belong to Organic Search. GA4's channel definitions and scopes
Test the meaning of consequential events. A click towards a shop is an onward action; payment confirmation comes from a different record. Exclude identifiable internal tests, preserve legitimate consent choices and check cross-domain journeys where the sales site differs from the article site. Obtain aggregate reporting that answers the business question without collecting private assistant conversations.
For longer buying cycles, keep first discovery, later visits and the eventual opportunity or order as distinct observations. A voluntary “How did you hear about us?” answer can help, but memory and response bias limit it. Reconcile it to the same customer record so one buyer does not become two AI-generated customers.
What would a defensible calculation look like?
Consider this invented monthly review. The programme costs £900 in additional external spending. Internal work uses eight hours, valued at £40 an hour for capacity planning: £320. The combined resource cost is therefore £1,220; the cash cost remains £900.
The report contains 120 citations, 30 identified unpaid assistant sessions and four linked orders. One order is a known internal test and one is fully refunded. The remaining two orders each contribute £150 after the identified product, delivery and transaction costs, before this programme's charges. The recorded retained-order contribution is £300.
Two customer interviews also mention AI. One belongs to an order already counted; the other has not become an order. Neither adds another £150. Nor can 120 citations be divided by 30 sessions to produce a dependable visitor conversion rate: the reporting coverage and counting units differ.
The tracked £300 covers one-third of £900 cash spending, or about 25% of the £1,220 combined resource cost. Cost coverage is all this calculation establishes. Causal ROI remains unknown: either retained buyer might have purchased anyway, and unobserved assisted demand may exist.
At £150 contribution per genuinely additional retained order, six such orders would recover the £900 cash cost and seven would exceed it. Nine would exceed the £1,220 resource cost. Those thresholds require the stated margin and cost assumptions to hold; extra fulfilment capacity or maintenance could raise them. They are planning requirements, not a forecast that citations will produce those orders.
How would you derive an incremental estimate?
Extend the same invented review. Suppose a test was agreed beforehand using comparable topic groups, with related pages kept together, one group receiving the programme and another left unchanged. The tracked outcome is all-route retained orders associated with those predefined groups, not just assistant-referred orders. Before the test, define how purchased products or services map to the topic groups and set one allocation rule for mixed orders, so each order is counted once. Both groups recorded six orders in the prior comparable period. Equal counts in one prior period do not establish comparable trends; inspect several earlier periods where feasible and assess the design assumptions separately. Afterwards, the changed group records ten and the comparison group eight. The two retained assistant-linked orders above are already included in the changed group's ten.
The changed group's increase is four orders. The comparison group's increase is two. A simple difference-in-differences estimate attributes the remaining two to the programme: (10−6)−(8−6)=2. At the same £150 contribution, that is £300. It is not another £300 to add to the referral ledger; the methods examine overlapping business.
On that estimate, cash-basis ROI is (£300−£900)/£900, about −67%; combined-resource ROI is (£300−£1,220)/£1,220, about −75%. Counting all four additional orders in the changed group would instead assign £600 and ignore the comparison trend.
This calculation does not make the design credible by itself. The groups need a defensible comparable trend, stable measurement and similar opportunity to buy; unrelated changes and cross-group effects must be assessed. Random allocation can strengthen a suitable experiment. These tiny counts cannot support a precise causal conclusion. A content change may also improve ordinary SEO, so the estimate concerns the tested programme, not an isolated AI-only effect.
Stress-test the decision separately from statistical uncertainty. One genuinely additional order would supply £150; seven would supply £1,050 and exceed the £900 cash cost, while still falling short of £1,220 resource cost. Those are hypothetical scenarios, not a confidence interval. They show why the evidence must distinguish a plausible cash case from a resource case before expansion.
For lead generation, forecast qualified opportunities multiplied by a justified close-rate assumption and contribution per won customer, then replace the forecast with actual mature outcomes. Keep estimated lifetime contribution, refunds and cash dates explicit, and report pipeline separately from realised revenue. A simple before-and-after rise remains weaker evidence when price, paid media, demand or availability also changed.
What should the next investment decision be?
Set the commercial requirement before renewing. In the example, investigate whether the programme can plausibly produce the additional retained orders within an affordable period. A broken enquiry route needs repair before further visibility spending. If descriptions are inaccurate, their correction takes priority over treating additional mentions as success.
At very low volumes, a small bounded test and conversations with actual buyers can be more useful than a complex attribution subscription. Evidence and strategic value should justify renewal. Narrow or stop work that produces irrelevant exposure. Where return is still unknown, record “return not yet established” and specify what evidence is needed for the next spending decision.