SI & AgenticAnalytics
We bring web and app interactions, SI-sourced referrals and the agent data you can actually access into one measurement system. We verify the data, examine how it relates to business outcomes and produce reports you can make decisions from.
Is this for you?
If one of these sounds familiar, you are in the right place.
- Traffic from ChatGPT and Perplexity disappears inside direct or referral in your reports.
- You cannot separate bot and agent visits from human visits.
- You have no management view that shows what the SI channel contributes to revenue.
What is SI & Agentic Analytics?
SI & Agentic Analytics brings web and app interactions, SI-sourced referrals and accessible agent transaction data together in one measurable data system. It verifies data quality, examines the relationship with business outcomes and produces reports that can be acted on.
Three different jobs are often given the same name: analysing with SI, measuring the customers SI sends you, and measuring how an SI agent performs. We write explicitly which one is in scope. Without reliable conversion and customer data, SI impact cannot be measured reliably either, which is why every engagement starts with an audit of the existing analytics stack.
Five measurement areas that must not be conflated
| Area | What is examined | Data source | Question it answers |
|---|---|---|---|
| Web and SI referral analytics | Visits from SI platforms and what happens next | GA4, app events, CRM | Which entries turn into qualified demand? |
| SI visibility and reputation | How the brand is described and cited on selected questions | Brantial and similar sampling tools | Where do we appear and how are we represented? |
| External agent access analytics | Crawler, fetcher and agent requests to the site | CDN, WAF and origin logs | Who reaches which content, with what result? |
| Owned SI application analytics | Execution of the assistant or agent you run | Application instrumentation and evaluation | Is the task completed correctly, safely and economically? |
| Agentic commerce transaction analytics | Offer, order, payment and refund | Commerce backend, verified webhooks and APIs | Where did a started transaction end, with what net result? |
These areas are related in the same report but never summed as if they were the same event. A page crawled a thousand times does not mean a thousand users, a thousand recommendations or a thousand purchase opportunities.
Every number carries its evidence level
Direct observation
An event a system actually recorded, with a known source and scope. Even so, every source passes a quality test; wrong instrumentation is a reality too.
Verified linkage
Two records joined through a permitted transaction or business identifier. The matching rule and window are documented.
Analytical inference
An assessment made with a model, time proximity or behavioural rule. The method is stated openly in the report.
Unknown
No data, insufficient scope or an unverifiable relationship. These rows are not hidden; they appear as "not measured".
When the denominator is zero the result is not 0%, it is not measured. A match is not causality, and attributed revenue is not the incremental revenue the service caused. We keep that distinction in every report.
The measurement ground shifted this year
The default channel definitions in GA4 now include SI Assistant, covering sources such as ChatGPT and Gemini, while AI Overviews and AI Mode stay inside Organic Search. Older reports built on a referral-only filter need re-checking. We do not start a project by copying a fixed regex list from a blog post.
The Search Console generative SI performance report provides impression data for AI Overviews and AI Mode, not conversations, sentiment or a separate click metric. Bing Webmaster Tools offers citations, cited pages and citation share views. Metrics from the two products cannot be added into one total SI reach.
There is now more than one route for sending server events into GA4. For a new build we choose based on access, diagnostics, batch error behaviour and maintenance capacity. We do not replace a pipeline that works simply because a new option exists.
The semantic conventions for generative SI observability are still in development. We pin the versions of the libraries, exporters and field mappings we use, so a moving standard does not silently break historical reporting.
Three core packages, two add-on modules
Measurement and data quality audit
System and data inventory, KPI dictionary, sample flow verification, source and actor separation, critical error list and target architecture. Acceptance: it is clear which business question can be answered with which data, and every gap has evidence and an owner.
Implementation and integration
A versioned measurement plan, event and data layer contract, SI channel and log normalisation, permitted customer-data joins, data models, dashboards and quality checks. Acceptance: the selected flow is traceable from source to business outcome.
Continuous measurement and decision support
Data health monitoring, classification maintenance, explained anomalies, cohort and funnel analysis, a monthly decision report, experiment proposals and an action log. Monitoring frequency is set by contract; there is no default round-the-clock operation.
SI application observability and evaluation
An add-on module for the SI application you run: tracing, task outcome, cost and latency, tool behaviour, a versioned test set and rollout controls.
Agentic commerce measurement
A conditional add-on. Where a real integration exists: checkout, order, payment and refund events, authorisation status, webhook deduplication and reconciliation. Without technical access, the package is not presented as if it were installed.
The management view reads in this order
| Panel | What it shows | Drill-down |
|---|---|---|
| Business outcome | Qualified leads, opportunities, net revenue and the attribution method used | Permitted, narrow-scope customer records |
| Acquisition | SI Assistant, organic search and other channels | Source, landing page, date and scope |
| External agent access | Verified or claimed identity, route and result | Redacted request detail and verification method |
| Internal agent | Task, quality, latency and cost | Authorised trace and evaluation records |
| Data health | Latency, loss, matching, duplicates, model changes | Test and incident log |
| Actions | Issue, decision, owner, date and re-check | The related work list |
We do not produce a single large SI score. Units that cannot be converted into one another are never summed in the same chart, and missing data is never drawn as performance falling to zero. The monthly decision note answers five questions: what can we trust, what changed, how does it relate to a business outcome, what is still uncertain, and who does what next.
Metric definitions that go into the contract
| Metric | Calculation | Interpretation limit |
|---|---|---|
| SI-sourced session share | Measurable sessions in the SI Assistant scope over all sessions in the same scope | Not total SI impact or market share |
| Goal completion rate | Unique SI sessions with at least one selected goal event over all SI sessions | Different from dividing event counts by sessions |
| Customer-data match coverage | Eligible leads with a permitted deterministic join over leads expected to have a web touch | Unmatched records are retained, not deleted |
| Verified agent access | Deduplicated requests with a verified identity | Not a count of users, tasks or sales |
| Content access success | Requests verified to have received the expected content over requests examined | Not generalised from a sample to all requests |
| Verified task success | Tasks confirmed complete by a business record over eligible tasks | A retry is not counted as a separate task |
| Cost per successful task | Total cohort cost over verified successful tasks | The cost of failed attempts stays in the numerator |
How the work is divided with our other services
Visibility side
Visibility Intelligence and Citation Optimization produce sampled visibility and citation data. Analytics takes it as context and never auto-matches it to sales.
Experience side
Agent Experience resolves where an agent or user fails to complete a task. Analytics supplies the failure and success data.
Measurement foundation
Our web analytics and data analysis service establishes GA4, tag management and server-side measurement. This service adds the SI and agent layer on top.
Prerequisites and the promises we do not make
Three things are needed for the service to work: permitted and functioning web measurement, a business record where conversion can be verified, and access to logs or the application. Without them the first engagement is an audit, not a report.
An SI referral does not mean the visitor was an agent; more often it is a person clicking the link in an answer. We cannot measure the real internal decision mechanism of an external platform. We do not guarantee merchant acceptance on a given platform or a given volume of agent-driven sales. Anything we cannot measure is marked in the report as not measured.
Frequently asked,
carefully answered.
01 Does this replace our existing analytics service?
No. GA4, tag management and server-side measurement remain the foundation. This service adds SI referral, agent access and, where needed, application observability on top of it.
02 We already see SI traffic in GA4, what more is there?
The SI Assistant channel in GA4 shows measurable sessions only. Agent access sits at log level, transactions in the commerce backend and visibility in sampling tools. This service joins those sources in one framework with explicit evidence levels.
03 How do you separate agent traffic?
We store identity, behaviour, purpose and evidence level in separate fields. Requests with provider verification, a signature or CDN verification never land in the same bucket as requests based on a user-agent claim alone.
04 Do you also measure our own SI application?
Yes, as an add-on module. That work requires access to the application code and runtime, so it runs under a separate scope with a named technical owner.
05 When will I see results?
The audit output is ready within a few weeks. After implementation, the first reliable monthly decision report waits for data maturation, which depends on your conversion cycle.
06 Can you tell us the incremental revenue SI brought?
We report attributed revenue and the observed relationship. Causal incremental revenue is only established through an experiment; where it fits we propose the design and label the result separately.
07 Which tools do you work with?
It depends on the existing stack: GA4, a tag manager, a data warehouse, the customer data system, CDN and log platforms, and open-standard observability tooling on the application side. The goal is data ownership, not tool lock-in.