Website analytics for clear acquisition and conversion decisions
Understand the website metrics, attribution rules, bot controls and conversion evidence needed for a concise but commercially useful analytics dashboard.
Website analytics should answer a compact set of questions quickly: how much qualified attention arrived, where it came from, which pages earned engagement, what visitors tried next, where conversion failed and which outcomes created real value.
More dimensions do not automatically produce better decisions. A useful dashboard keeps core definitions visible, separates human and automated traffic, lets every table become a filter, and connects acquisition to explicit calls to action and trusted outcomes.
MetricFold combines that concise web view with product behavior, friction, performance and revenue lifecycle evidence. It is designed for marketing sites that lead into a SaaS or transactional product, where “page viewed” is only the beginning of the commercial journey.
The core metrics and their definitions
Visitors are distinct observed people under the system's identity model. In a privacy-minimized anonymous model, identity may rotate daily rather than persist for months. The dashboard should state that limit.
Visits group activity separated by an inactivity timeout. A visit is not necessarily a browser tab or authenticated session. Cross-midnight behavior and inactivity settings can change counts.
Page views count accepted page-view events after duplicate, preview, origin and bot controls. Single-page applications must send navigation changes without counting the initial render twice.
Views per visit is page views divided by visits. It can indicate exploration, but a focused one-page conversion may succeed with a low value.
Bounce rate usually means the share of visits that fail the system's engagement rule. The rule may include a second page, a meaningful action or minimum active time. Compare the definition before comparing products.
Visit duration should be based on observed event timing and capped to avoid an abandoned tab becoming an all-day session. It is directional, not a stopwatch of human attention.
These metrics become useful when the current window includes an equal previous-period comparison and annotations for launches, campaigns, outages or tracking changes.
Sources, channels and campaigns
A source report begins with the referring origin and campaign parameters. A channel groups those facts into useful categories such as organic search, paid search, social, referral, email and direct.
Attribution needs deterministic rules. Normalize known search and social domains, preserve explicit UTM parameters, remove sensitive query data and document the direct-traffic fallback. Do not overwrite a known first touch merely because a later visit is direct.
For commercial decisions, show both first-touch and last-touch context where possible. First touch answers how demand was originally acquired. Last touch answers what immediately preceded the outcome. Neither proves that one interaction caused the conversion.
AI assistant referrals deserve their own source grouping. Referrals from ChatGPT, Gemini, Claude, Perplexity and other answer surfaces can be shown separately from AI crawler traffic. A crawler fetching a page is not a human referral and must not inflate visitors.
Pages are roles, not just paths
Top pages show volume, but entry and exit roles often reveal more. An entry page acquires a visit. An exit page is the last observed page, which may represent success or abandonment. A confirmation page should have a high exit rate; a signup step may require investigation.
Useful page evidence includes:
- visitors, views, entries and exits;
- active engagement and scroll depth;
- CTA impressions and clicks;
- next and previous actions;
- conversion and revenue by entry page;
- errors and Core Web Vitals by device;
- campaign and source mix.
Normalize dynamic identifiers in paths where they would create thousands of meaningless rows. Keep route templates when the entity id does not support a report.
CTA analytics without manual event handlers
Calls to action are more useful when the system knows both exposure and activation. MetricFold uses a declarative data-analytics-cta identifier. The lightweight tracker observes when the element becomes meaningfully visible and records a click against the same stable target.
Anchor: href=/register data-analytics-cta=pricing.primary label="Start free"
This produces impressions, clicks and click-through rate without reading visible button copy or requiring a custom JavaScript handler. Use identifiers that describe the role, not the current text. When a CTA moves or its copy changes, the historical series stays interpretable.
Autocapture should remain bounded. Tracking every DOM click creates noise, high cardinality and accidental sensitive data. MetricFold observes a small safety-focused set automatically and uses semantic attributes for commercial interactions.
Bot filtering and traffic quality
No browser analytics product can perfectly classify every automated request. Known user-agent registries catch search, preview, uptime, SEO, AI and training bots. Behavioral rules catch impossible navigation rates, missing browser signals and repeated data-center patterns. Server-side crawler tracking can record legitimate crawler activity in a separate report rather than discarding it.
The dashboard should expose excluded traffic counts and reasons. Hidden filtering makes discrepancies impossible to diagnose. Treat the human dashboard, crawler dashboard and rejection diagnostics as separate datasets.
Bots that execute full browser JavaScript and imitate humans remain a hard boundary. Use rate limits, network reputation and product-side abuse controls when adversarial traffic matters. Analytics classification should not be your only security layer.
Geography and technology
Country, region and city are derived from network context before the raw address is discarded. A privacy-minimized map should show coarse country or region centroids, not pretend to locate a person precisely.
Device, browser and operating system reports are valuable when joined to conversion, friction and performance. A mobile share alone is descriptive. A mobile signup drop combined with failed requests on a browser version is actionable.
Performance is conversion context
Collect real-user LCP, INP, CLS and TTFB using the browser's performance APIs and report percentiles, sample counts and affected paths. Averages can hide the slow tail. The Core Web Vitals thresholds published by web.dev provide investigation boundaries, not guarantees of ranking or conversion.
Keep performance events sampled and bounded. The analytics tracker must not become a performance problem itself: load it with defer, avoid synchronous dependencies, compress it, coalesce repeated signals and isolate dashboard-only libraries from customer pages.
The concise weekly report
A useful weekly website review fits on one operating surface:
- visitors, visits, views, engagement, signups and trusted revenue;
- equal-period changes and annotations;
- channel, source and campaign quality;
- entry, top and exit pages;
- CTA exposure and conversion;
- country and technology segments;
- funnel drop, friction and performance evidence;
- one reproducible next check.
Export and API access matter because the dashboard is not every possible analysis. They should be scoped and bounded, with clear definitions and no secret credentials in client payloads.
MetricFold's purpose is not to recreate every Google Analytics report. It is to connect the website evidence most teams use with what happens after the click, then make the next investigation obvious.
Use a dashboard hierarchy that preserves context
| Surface | Must answer | Drill-down evidence | Failure mode to avoid |
|---|---|---|---|
| Overview | Did qualified attention and value change? | Equal period, annotations and data quality | A wall of unrelated KPIs |
| Acquisition | Which sources and campaigns create value? | Entry page, CTA, activation and revenue | Optimizing visitor volume alone |
| Content | Which page roles help or lose the journey? | Entry/exit, scroll, CTA, next action and performance | Treating every exit as failure |
| Technology | Is a device/browser cohort blocked? | Funnel loss, error and RUM distribution | Reporting share without outcomes |
| Realtime | Is a launch or incident behaving now? | Active pages, actions, countries and stream health | Animated map without definitions |
Plausible's overview documentation is a useful benchmark for concise web reporting. MetricFold adds the downstream funnel analytics, product analytics and revenue attribution needed when the marketing site leads into a SaaS product.
Reconcile the visitor number
Test controlled traffic
Create deterministic direct, campaign, referral, SPA, duplicate and bot visits. Record expected page views, visit boundaries, source and exclusion reason. Test the day boundary and inactive timeout. Compare hosted and same-origin delivery against the same collector contract.
Compare server-owned outcomes
Reconcile signups, activations, paid outcomes and refunds against the systems that own them. If website visitors fall while verified signups stay flat, investigate collection or traffic quality before declaring conversion improved. Expose excluded and unattributed totals.
State uncertainty
Ad blockers, disabled JavaScript, network failure, bot evasion and rotating identity mean accepted browser events are a measurement, not objective reality. Accuracy comes from explicit definitions, negative data tests and outcome reconciliation rather than a promise of perfect humans.
The bot filtering guide distinguishes accepted, excluded, validated-human and unknown traffic. The web.dev vitals reference provides primary definitions for the performance observations.
Frequently asked questions
What should website analytics track automatically?
Page/SPA navigation, bounded engagement, outbound and download actions, form lifecycle, errors, failures and Core Web Vitals are useful defaults. Use stable semantic attributes for commercially important CTAs.
Is every accepted visitor definitely human?
No. Known and high-confidence automation is removed with visible reasons, but ordinary accepted traffic remains accepted rather than mathematically proven human. Server-owned activation can provide positive validation.
Why is a country map approximate?
MetricFold retains a coarse country code and places markers at country centroids. It does not store or display precise visitor coordinates. The adjacent table is the exact count.
What is the first useful conversion?
Choose the next durable value step after the website CTA—completed signup, qualified lead, activated account or verified payment. A click alone measures intent, not the business outcome.
Which external definition should be documented?
GA4's traffic acquisition dimension reference illustrates how another platform names acquisition fields. MetricFold does not claim identical channel or session rules; it publishes its own source, first/last touch, direct and visit definitions beside the report.
How should the dashboard be released?
Seed dense deterministic events across sources, pages, devices, browsers, countries, CTAs, funnel steps, friction, payments and bots. Exercise every tab, filter and export with real interaction tests. Inspect desktop and mobile in light/dark and reduced motion, test keyboard/focus and accessibility, and prove live data changes without reload. Empty screenshots alone are not a release gate.
What evidence belongs in every release?
Keep the tracker gzip and execution budget, storage/network inspection, collector contract tests, known-bot/real-browser corpus, deterministic report fixture, accessibility results and representative screenshots. Reconcile one server signup and one verified money outcome. A release that cannot prove these basics should not claim accurate, lightweight or decision-ready analytics.
Review source/channel rules and the crawler registry on a schedule. Those dependencies change traffic categorization even when the dashboard component does not. Preserve the version or annotation so a reporting shift can be explained.