How to measure an AI website consultant: enquiries, conversion and ROI

An AI consultant helps a business when visitors receive accurate answers, suitable enquiries reach staff, and incremental profit or actual savings justify the costs. Message volume alone does not show this. This guide provides a measurement framework for your website, analytics and business records. Every calculation is hypothetical: these are not NEXSTA customer results or a promise of returns.

Define one business outcome

For a service website, the outcome might be a qualified enquiry; for a shop, an order; for support, an issue resolved without a repeat contact. Define a useful enquiry in advance: a relevant service, an area you serve, an understandable need and an agreed contact method. Spam and requests for unavailable services do not qualify.

Agree on the reporting period, time zone, traffic sources and the window between an enquiry and a sale. Orders can arrive after the conversation. Do not compare an incomplete week with a full month or credit every later purchase to the chat.

Build a funnel from session to sale

• Sessions where the consultant was available: your starting audience. Track widget loading failures separately.
• Sessions with a conversation: a visitor sent a message. Opening the window without asking anything is a separate action.
• Sessions with a confirmed enquiry submission: the system actually accepted or stored it, rather than merely recording a button click.
• Unique qualified enquiries: staff checked the agreed criteria and merged repeat submissions of the same request.
• Sales and contribution margin: outcomes from business records after staff processed the enquiry.

Document each step, its source and its owner. A session, person, conversation and enquiry are different units. One enquiry can contain many messages and one customer can return several times. Do not add these counts as if they represented the same thing.

Calculate conversion with explicit denominators

Conversation rate = sessions with a conversation / sessions where the consultant was available × 100%. Conversation-to-enquiry rate = sessions with an enquiry following a conversation / sessions with a conversation × 100%. Website enquiry conversion = sessions with an enquiry / all comparable website sessions × 100%.

Hypothetical monthly example: 10,000 sessions, with the widget available in each; 600 sessions with a conversation; 90 sessions with one new enquiry each. The rates are 6%, 15% and 0.9%. Of the 90 unique enquiries, 60 qualify and 12 become sales: the qualification rate is 66.7% and qualified-enquiry-to-sale conversion is 20%. The last two rates use enquiries, not sessions. They do not demonstrate incremental impact from AI.

Separate attribution from incremental impact

A customer might have called or completed a form without the consultant. Compare enquiries across all channels: more chat enquiries accompanied by an equal fall in forms can mean contacts simply moved. People who choose chat may already have stronger purchase intent.

With sufficient traffic, consider a properly configured experiment that randomly assigns visitors and keeps each person in a consistent group. Compare total qualified-enquiry conversion and sales between groups. With limited traffic, observe comparable periods while recording advertising, prices, seasonality and page changes; this does not establish causality.

For example, 100 sessions with a useful enquiry out of 5,000 equal 2%; 130 out of 5,000 equal 2.6%. The difference is 0.6 percentage points, or 30% relative to the starting rate. This illustrates arithmetic, not statistical significance. Observation time depends on traffic, conversion variability and the sales cycle; there is no universal duration.

Check answer quality and staff follow-up

• Review successful and unsuccessful conversations for accuracy against company knowledge, appropriate questions and invented terms.
• Track qualified enquiries, repeat contacts, complaints and cases needing a human.
• Measure time to the first staff reply and resolution. Median and 90th percentile are useful alongside the average.
• Investigate differences by page, device and source. Many conversations without enquiries warrant checking the answers and the next step.

Busy chat activity with inaccurate answers can harm performance. See the knowledge-base guide and the guide to collecting enquiries.

Calculate ROI using the full cost

Include costs for the same period: the service, setup with a clear allocation of one-off expenses, knowledge updates, answer reviews, integrations and additional enquiry handling. Contribution margin is sales revenue minus variable costs associated with those sales. Do not substitute total revenue.

ROI = (incremental contribution margin + substantiated savings − total costs) / total costs × 100%. This formula does not apply when costs are zero. Additional sales needed to break even = costs / contribution margin per additional sale, rounded up; this simplified version assumes equal margins and no savings.

A separate hypothetical example: an evidence-based estimate identifies 5 additional sales with a contribution margin of RUB 4,000 each, against RUB 12,000 of costs. Net benefit is RUB 8,000, ROI is about 66.7%, and break-even requires 3 additional sales. These 5 sales cannot be inferred from the earlier funnel: incremental impact needs separate evidence. These figures are neither NEXSTA pricing nor NEXSTA results. At RUB 24,000 of costs, the same contribution would give an ROI of about −16.7%.

Show freed-up time separately. Ten hours valued internally at RUB 500 per hour represent RUB 5,000 of time, but not necessarily cash savings. Add this to financial results only when the economic benefit is substantiated and the same work is not counted again in costs or incremental profit.

Set up measurement without unnecessary data

Yandex Metrica distinguishes goal completions from sessions containing a goal; session conversion uses the latter. In Google Analytics, events important to the business can be marked as key events. Match naming and counting rules to your funnel: event occurrences do not equal new enquiries.

Test an enquiry from submission to the business record and exclude tests from production reports. Use UTM parameters for social links. Keep phone numbers, email addresses and full conversations out of URLs and analytics parameters. Blockers, visitor consent and cross-device journeys can leave gaps, so reconcile analytics with business records.

This is a recommended measurement framework, not a list of automatic NEXSTA reports. Check the events and integrations in your configuration; analytics setup may require separate work.

Start with a focused NEXSTA scenario

Choose one use case, prepare company knowledge and test typical and difficult questions. Explore industry examples and the setup guide. Record your baseline and assign someone to handle enquiries. Try the NEXSTA AI consultant on your website’s tasks, then decide based on answer quality and measured outcomes.

Primary sources

• Yandex Metrica: goals and goal metrics.
• Google Analytics: key events and reporting.

Sources checked on 2 November 2026. The editorial team supplies formulas and hypothetical scenarios for your own calculations.