AIR Engine™

Pro feature, available in Referrer Attribution Pro v1.15.0+ and Refer App Pro

What the AIR Engine™ does

Most attribution tools tell you which channels are sending traffic. The AIR Engine™ goes one step further: it analyses the traffic that cannot be attributed. The Direct sessions cannot be individually attributed, and identifies the most probable explanation for where it came from.

Direct traffic is not a channel. It is a classification failure. Every session that arrives with no referrer and no UTM parameters is a gap in your attribution picture. Some of those sessions are genuinely direct: bookmarks, typed URLs. Most are not. They are dark social shares, branded search returns, cross-device revisits, and AI halo traffic that lost its referrer header in transit.

The AIR Engine™ runs a deterministic rule-based analysis against your existing session data to tell you which explanation is most probable for each page and each pattern it detects.

How it works

When you open the AIR tab, the engine:

  1. Collects all Direct first-touch sessions in the selected date range
  2. Groups them by page path, normalising trailing slashes so /page and /page/ are treated as the same page
  3. For each path, computes four predictive signals against the historical cross-channel data already in the database
  4. Evaluates the data against 20 pattern rules
  5. Generates a finding card for each rule that fires, with a confidence score and a recommended action

No data leaves your server. No external API is called. Zero LLM dependency: every output is derived from local session data.

The four predictive signals

Direct-to-known ratio

For each page, the engine calculates what proportion of all sessions on that path are Direct, then compares it against the site-wide average. A page where Direct is disproportionately concentrated above the site baseline is a stronger signal for dark social or branded revisit behaviour than a page where the ratio is at baseline.

Temporal correlation

The engine counts how many Direct sessions on a path arrived within 72 hours of a known-channel session on the same path. This identifies the "saw it through Google, came back directly the next day" pattern: consistent with both cross-device revisits and peer-to-peer sharing.

Viewport divergence

Compares the desktop/mobile split of Direct sessions against the known-channel sessions on the same path. A large gap (for example, Direct sessions are 80% desktop while organic sessions are 60% mobile ) is a strong cross-device signal.

Velocity trend

Splits the evaluation window in half and compares Direct session counts in the recent half against the earlier half. A rising trend increases the urgency of the finding.

The confidence score

Each triggered finding card shows a confidence score from 0 to 100. The score is derived from five weighted factors:

Findings are grouped into three priority tiers:

The five finding categories

Category A: Dark Social and Peer-to-Peer Networks

Direct traffic arriving on specific content pages from desktop browsers, with known organic or AI channel footprint on the same path. Consistent with enterprise peer sharing, LinkedIn DMs, and email forwards where the referrer header is stripped.

Category B: Cross-Device and Halo-Effect Attribution

Direct traffic to the homepage that correlates with active local search, AI referrer, or paid campaign activity in the same window. Consistent with a visitor seeing the site on one device and returning on another, or an AI tool citing the site and the user arriving directly rather than clicking through.

Category C: Legacy URLs and Content Audit

Direct traffic to pages with outdated URL patterns (year-stamped paths, archived directories) or informational content with no organic footprint. Flags pages that may need redirects or commercial conversion hooks.

Category D: B2B Authority and Ecosystem Credibility

Direct traffic to procurement, personnel, sector-specific, or leasing pages. Consistent with vendor vetting, procurement research, and industry directory referrals where the referrer is suppressed.

Category E: Automated Traffic and Infrastructure

Direct traffic to system paths, admin endpoints, or sitemap files. Flags paths that should be excluded from marketing reporting views.

The executive summary

The top of the AIR tab shows an executive summary card before any findings. It contains:

Marking findings as reviewed

Each finding card has a Mark as reviewed action at the bottom. Reviewed findings move to a collapsed Reviewed findings section at the bottom of the report, showing the tier, classification, page path, and the date it was reviewed. Findings can be restored to the active list at any time. The reviewed state persists across sessions.

Translating the report to plain English

The Copy prompt button copies a complete, self-contained prompt to the clipboard. The prompt includes the full finding data and instructs any LLM to produce a plain-English summary for a non-technical marketing professional, prioritised by confidence and ending with a list of the top three recommended actions.

The Open in Gemini button opens this prompt directly in Google Gemini (available when the report is short enough for a direct URL: typically fewer than eight findings).

Requirements

Differences between the WordPress and Refer App implementations

The AIR Engine is functionally identical across both products. The same 20 rules, the same five-factor confidence formula, the same four predictive signals, and the same Attribution Health Score apply in both.

One architectural difference: the WordPress plugin stores device type (desktop, mobile, tablet) in a separate Pro-only table keyed to the free plugin's session log. Device signals are available from the point of Pro activation, not retroactively. Sessions logged before Pro was activated show as Unknown for device-dependent rules. This matches the same approach used for the is_first_touch flag in the free plugin.

← All documentation Refer App changelog WordPress plugin changelog