> ## Documentation Index
> Fetch the complete documentation index at: https://docs.asva-ai.com/llms.txt
> Use this file to discover all available pages before exploring further.
# How Asva's AI Commerce Attribution Model Works
> Understand the signal tiers, confidence scoring, attribution windows, and multi-touch options Asva uses to classify AI-sourced sessions and attribute revenue.
Asva does not rely on a single signal to decide whether a session was AI-sourced. A referrer domain alone is unreliable — ChatGPT strips it. A product-page landing alone is ambiguous — direct bookmarks land the same way. Instead, Asva's attribution engine combines multiple session signals, weights them by reliability, and produces a confidence score for each session. This page explains what signals Asva looks for, how they are combined, and what ends up in your dashboard and GA4 reports.
## Signal hierarchy
Signals are evaluated in order of reliability, from direct evidence down to behavioral inference.
### Tier 1: Direct signals
These signals provide explicit evidence that a session originated from an AI surface. When present, Asva classifies the session at high confidence without requiring additional signals.
| Signal | Example | Confidence |
| ------------------------ | --------------------------------------------------- | ---------- |
| Explicit `utm_medium=ai` | ChatGPT link with UTM params attached | 95%+ |
| Known AI referrer domain | `chatgpt.com`, `gemini.google.com`, `perplexity.ai` | 90%+ |
| Asva SDK source flag | Session tagged by your ACP or UCP checkout endpoint | 95%+ |
### Tier 2: Inferred signals
These signals are consistent with AI-sourced traffic but are not conclusive on their own. Asva combines them to reach a confidence score.
| Signal | What it indicates | Confidence |
| -------------------------------------- | -------------------------------------------------------- | ---------- |
| Direct + non-homepage landing | Arrived at `/products/running-shoes` with no referrer | 65–80% |
| Behavioral fingerprint match | Session timing matches known AI-referral patterns | 55–75% |
| Product-specific landing from new user | First-ever session that lands directly on a product page | 60–70% |
### Tier 3: Negative signals
These signals reduce confidence that a session is AI-sourced. When a negative signal is present, Asva downgrades the session's score or reassigns it to a different channel.
| Signal | What it means |
| ----------------------------- | -------------------------------------------------------------------------------------- |
| Bookmark detected | Browser signals a bookmark visit — lower AI probability |
| Paid click parameters present | Google Click ID (`gclid`) or Meta Click ID (`fbclid`) present — credit to paid channel |
| Internal navigation | User arrived from another page on your own site |
| Known bot user agent | Crawler or headless browser, not a human shopper |
## Confidence thresholds
Each session receives a confidence score from 0 to 100 based on its combined signals. The score determines how the session is classified and reported.
| Confidence | Classification | Reported as |
| ---------- | -------------------------- | ---------------------- |
| ≥ 80% | High-confidence AI session | AI Commerce |
| 50–79% | Probable AI session | AI Commerce (probable) |
| \< 50% | Insufficient evidence | Direct (unclassified) |
By default, high-confidence and probable sessions are combined and reported together as "AI Commerce". To split them — for example, to report only high-confidence sessions in your KPI dashboards — go to **Dashboard → Attribution → Display settings** and enable **Show confidence breakdown**.
For sessions with a direct signal (UTM or known referrer), attribution accuracy is 90%+. For sessions classified using only inferred signals, accuracy is typically 70–85% depending on your site's traffic mix.
## Attribution window
Asva uses a **30-day last-touch** model by default. A conversion is attributed to the AI Commerce channel if the session was classified as AI-sourced within 30 days before the conversion event.
You can change the attribution window in **Dashboard → Attribution → Model**:
| Window | Best for |
| ------- | ---------------------------------------------------- |
| 7 days | Conservative reporting; short consideration cycles |
| 30 days | Default; suits most DTC and mid-market brands |
| 90 days | Extended; high-consideration or high-ticket products |
## Multi-touch attribution
For brands with complex customer journeys, Asva supports multi-touch attribution. A typical AI-assisted journey might look like this:
```
User journey:
Day 1: Arrives via ChatGPT → browses (AI Commerce session)
Day 5: Returns via Google Organic → adds to cart
Day 7: Returns via Direct → purchases
Default (last-touch): Direct gets 100% credit
Multi-touch (linear): AI Commerce, Organic, Direct each get 33%
Multi-touch (first-touch): AI Commerce gets 100%
```
Configure the attribution model in **Dashboard → Attribution → Model**. Linear and first-touch models are available on Growth and Enterprise plans.
## What gets reported
Once sessions are classified, Asva surfaces attribution data in three places:
AI Commerce channel breakdown, per-surface traffic, product-level AI referrals, and month-over-month trends.
`ai_source`, `is_ai_session`, and `ai_confidence` are passed to GA4 on every classified page view. See [Integrate Asva Attribution with Google Analytics 4](/attribution/ga4-integration).
Pull session-level attribution data and revenue figures programmatically. See [Attribution API](/api-reference/attribution).
Daily or weekly exports are available from the Asva dashboard for use in spreadsheets or BI tools.
## Accuracy and limitations
For sessions with direct signals (UTM parameters or a recognised AI referrer domain), accuracy is 90%+. For sessions classified using only inferred behavioral signals, precision is typically 70–85% on sites with significant AI traffic volume. Sites with very low AI traffic may see more noise in the inferred tier.
Yes. Enable **Session-level export** in **Dashboard → Attribution → Export** to download each session's signals and confidence score. The same data is available via the [Attribution API](/api-reference/attribution).
No. All session classification is first-party and server-side. Asva does not use third-party cookies or cross-site tracking, and does not store personally identifiable information.
The Asva snippet runs client-side. For headless stores, you can also send attribution events server-side via the Attribution API when your UCP or ACP checkout endpoint is called. Server-side classification produces the highest-confidence attribution because it is triggered by a confirmed AI-initiated transaction.
## Related
Log attribution events and pull session-level data programmatically.
Surface AI Commerce as a named channel in your GA4 reports.