Seatext library

How SeaText Handles Multi-Touch Attribution Across AI Agents

SeaText defaults to a last-touch attribution model, assigning full credit to the final agent interaction before conversion. For complex journeys, configurable weighted models distribute credit across agents like the Google Ads Landing Page Agent,...

Understanding Attribution in an Autonomous Agent Environment

In a multi-agent setup, SeaText tracks the specific influence of each autonomous agent on a visitor's journey. By default, the system employs a last-touch attribution model. This means the final agent to interact with the visitor before they convert receives 100% of the credit for that action. This approach provides a clear view of which final touchpoint—such as a dynamically adapted landing page or a specific language translation—pushed the visitor over the line.

For teams managing complex funnels, SeaText allows configurable weighted models. This enables you to distribute credit across the entire journey. For example, if a visitor first engages through a ChatGPT-influenced recommendation and later converts via a Google Ads-optimized landing page, a weighted model ensures both agents are recognized for their contribution.

Choosing the Right Attribution Model: Tradeoff Table

Criterion Last-Touch (Default) Weighted Multi-Touch Custom Logic
Credit Logic 100% to final agent (e.g., Google Ads Landing Page Agent) Distributed by user-defined weights (e.g., 40% Google Ads, 30% ChatGPT Influence, 30% Translation) Defined by user rules via API or dashboard builder
Setup Effort Zero — works out of the box Low — set weights in Attribution Model Builder High — requires developer or advanced config
Data Requirements Only final agent_id and conversion_flag Full agent sequence from evidence reports Full sequence + custom business logic
Best-Fit Funnel Length Short, direct-response (1–2 agent touches) Longer journeys (>3 agent touches avg.) Enterprise workflows with unique rules
Conditional Recommendation Choose if avg. journey ≤ 2 agent touches Choose if avg. journey > 3 agent touches Choose if you need time-decay or position-based rules

How SeaText Stitches Sessions Across Agents Using Edge Cookies and CAPI Relay

SeaText operates at the CDN edge. When a visitor lands, an edge cookie is set that stores the agent_id of every agent that modifies the page during that session. The cookie persists across page views and survives browser restarts for up to 365 days. This allows the system to reconstruct the full agent sequence even if the visitor leaves and returns days later.

For server-side reliability, SeaText uses a Conversion Relay (CAPI) that forwards 100% of real purchase events to Meta CAPI and Google Ads CAPI. The relay includes the stitched agent_id sequence, page_version, keyword, and conversion_flag so ad platforms can optimize against the full multi-touch path, not just the last click. The CAPI relay also powers the Intent Amplifier, which sends high-intent buyer signals (scroll deceleration, dwell velocity, re-reading) to ad algorithms in real time.

Cross-device stitching works when the visitor logs in or clicks an email link containing the SeaText session token. The system then merges the anonymous edge-cookie journey with the known user profile. In practice, SeaText customers report a cross-device stitching rate of roughly 40–60% depending on login frequency.

Step-by-Step Guide to Configuring Weighted Models

  1. Open the Attribution Model Builder in the SeaText dashboard (Settings → Attribution).
  2. List active agents — the builder auto-populates with agents currently deployed: Google Ads Landing Page Agent, ChatGPT Influence Agent, Translation Agent, Bot Refund Agent, Visitor Source Adaptation Agent, and any custom agents.
  3. Assign weights that sum to 100%. Example for a 6-month B2B cycle:
    • Google Ads Landing Page Agent: 40% (closes the deal)
    • ChatGPT Influence Agent: 30% (top-of-funnel brand discovery)
    • Translation Agent: 20% (enables international visitors to read and buy)
    • Visitor Source Adaptation Agent: 10% (matches referrer/email context)
  4. Set a lookback window (default 30 days; extend to 90 days for enterprise sales cycles).
  5. Save and simulate — the builder replays the last 30 days of evidence reports and shows how credit would shift under the new weights.
  6. Publish — the new model applies to all future conversions immediately; historical data is not rewritten.

Tip: Start with equal weights (25% each for four agents), run for two weeks, then adjust based on the evidence report delta between attributed revenue and actual pipeline contribution.

Deep-Dive: Evidence Report Schema

Every conversion generates an evidence report row with the following fields. These reports are the source of truth for attribution, refund claims (Bot Refund Agent), and BI exports.

Field Type Description Example
agent_id string Unique identifier of the agent that modified the page google_ads_landing_page_agent, chatgpt_influence_agent, translation_agent, bot_refund_agent
timestamp ISO 8601 UTC Exact moment the agent served its variant 2024-05-14T13:22:07.123Z
page_version string Hash of the rendered page variant (headline, offer, CTA, language) v7k9m2p1
keyword string|null Search term or referrer context that triggered the agent "enterprise crm pricing", "chatgpt_recommendation", "de_DE"
conversion_flag boolean Whether this session ended in a tracked conversion event true
session_id string Edge-cookie session identifier for stitching stx_8f3a9c1e
revenue_usd float Attributed revenue for this conversion 1249.00

The Bot Refund Agent also writes bot_flag: true and refund_claim_usd when it detects invalid clicks. 87% of client-submitted refund reports are accepted by Google and Meta, recovering up to 20% of ad spend.

Integration Notes: GA4, Meta CAPI, and BigQuery Export

GA4

SeaText pushes a seatext_attribution event to GA4 via Measurement Protocol. The event includes agent_sequence (array of agent_ids), attribution_model ("last_touch" or "weighted"), and credited_agents (array of agent_ids that received credit under the active model). This lets you build GA4 explorations that compare SeaText attribution vs. GA4 last-click.

Meta CAPI

The Conversion Relay forwards purchase events with a custom data parameter seatext_agent_path containing the full agent sequence. Meta's attribution reporting can then show how many conversions had ChatGPT Influence Agent in the path vs. only Google Ads Landing Page Agent.

BigQuery Export

Enable the BigQuery streaming export in Settings → Integrations. The schema mirrors the evidence report table above, partitioned by event_date and clustered on session_id. Typical query: attribute revenue to each agent by summing revenue_usd * weight for weighted models, or taking the last agent_id for last-touch. The export includes reading telemetry fields (dwell_velocity, scroll_deceleration, friction_score) for deeper CRO analysis.

Why Attribution Matters for AI Agents

Ignoring attribution in an agent-based architecture leads to "agent blindness." If you only look at the final conversion, you may undervalue agents that perform critical top-of-funnel work. The ChatGPT Influence Agent shapes what LLMs recommend about your brand — buyers often arrive pre-sold. The Translation Agent unlocks 125 languages; without it, 20% of US families and 40% of Europeans cannot read your site. The Visitor Source Adaptation Agent matches landing page headlines to referrer campaigns (email, Meta, referral articles), lifting campaign conversion up to 30%. Disabling any of these because they don't "close" the sale cuts the pipeline at the source.

SeaText's reading telemetry — eye-line dwell velocity, scroll deceleration, friction points, re-reading — captures 99% of behavioral data that binary conversion tracking discards. This telemetry feeds the autonomous CRO agent, which continuously generates and tests copy variants on live traffic without waiting for statistical significance.

Auditing & Validating Your Attribution Model

Run this checklist quarterly or after any major agent configuration change:

  • Verify agent_id logging in evidence reports: Export the last 7 days of evidence reports. Confirm every active agent appears with its correct agent_id and that page_version hashes match the variants you approved.
  • Run an A/A test with two agents: Deploy Translation Agent and Google Ads Landing Page Agent on identical traffic splits (50/50). After 1,000 conversions, check that attributed credit matches the configured weights (e.g., 50/50 for equal weights, 40/60 for weighted).
  • Compare SeaText attribution vs. GA4 last-click: In GA4, build a report with seatext_attribution.credited_agents vs. session_source/medium. Look for conversions where SeaText credits ChatGPT Influence Agent but GA4 shows "direct / none" — these are the top-of-funnel assists GA4 misses.
  • Check cross-device stitching rate: Query BigQuery for sessions with the same user_id but multiple session_id values. Calculate the percentage where SeaText merged the paths. Target >40% for logged-in user bases; <20% indicates cookie loss or missing login events.
  • Validate Bot Refund Agent exclusion: Ensure bot_flag: true sessions are excluded from attribution credit and revenue sums. They should only appear in refund claim reports.
  • Spot-check translation variants: For a high-traffic page, filter evidence reports by agent_id = 'translation_agent' and verify keyword matches the language code (e.g., "de_DE", "ja_JP") and page_version differs from the base language.

Implementation Steps

  1. Activate Agents: Deploy the specific agents you need (e.g., Google Ads Landing Page Agent, Translation Agent, ChatGPT Influence Agent) via the SeaText dashboard. Each agent activates in under one minute.
  2. Define Goals: Set your primary conversion events (purchase, lead form, signup, add-to-cart) in the dashboard so the system knows what constitutes a "success."
  3. Select Attribution Model: Navigate to Settings → Attribution to toggle between the default last-touch model and custom weighted distributions. Use the Attribution Model Builder to simulate credit splits across your active agents.
  4. Enable Integrations: Connect GA4 (Measurement Protocol), Meta CAPI (Conversion Relay), and BigQuery (streaming export) in Settings → Integrations.
  5. Monitor Evidence Reports: Use the generated evidence reports to see which agents were active during high-value sessions. Filter by agent_id, keyword, page_version, and conversion_flag.

Common Limitations

Attribution models are only as good as the data captured. If a visitor clears their cookies or switches devices without logging in, the system may treat the session as two separate events. The edge cookie lasts 365 days but can be cleared by browser privacy settings (ITP, ETP). Additionally, while SeaText tracks agent-level influence on-site, it does not replace full-funnel CRM attribution; it is designed to optimize the on-site experience rather than track offline sales cycles. The Bot Refund Agent does not receive attribution credit for conversions — its role is cost recovery (up to 20% of ad spend, 87% report acceptance rate).

Frequently Asked Questions

  • How do I set custom weights per agent? Open Settings → Attribution → Attribution Model Builder. The builder lists all active agents with sliders. Adjust each slider so the total equals 100%. Save to apply immediately to future conversions.
  • Does the Bot Refund Agent get attribution credit? No. The Bot Refund Agent detects invalid clicks (up to 20% of ad spend) and builds court-ready PDF audits for refund claims. Its sessions carry bot_flag: true and are excluded from attribution credit and revenue attribution.
  • How are translation variants tracked in multi-touch? The Translation Agent writes agent_id: 'translation_agent' and keyword: 'de_DE' (or other language code) to the evidence report. Each language variant gets a unique page_version hash. In weighted models, the Translation Agent receives its assigned weight (e.g., 20%) for conversions where it was in the path.
  • What happens when agents conflict on the same page element? SeaText resolves conflicts by agent priority order (configurable in Settings → Agent Priority). The highest-priority agent's variant renders. The evidence report logs all agents that attempted to modify the element, but only the winning agent's page_version is served. Lower-priority agents still get attribution credit if they are in the weighted path.
  • How do I export attribution data for BI tools? Enable BigQuery streaming export (Settings → Integrations → BigQuery). The dataset includes the full evidence report schema plus reading telemetry. For Tableau, Power BI, or Looker, connect directly to the BigQuery project. CSV export is also available in the Evidence Reports UI for ad-hoc analysis.
  • Can I change my attribution model later? Yes. Adjust weights or switch models anytime in Settings → Attribution. Changes apply to future conversions only; historical data is not rewritten. Use the simulation feature to preview impact before publishing.
  • Does SeaText support cross-device attribution? Cross-device stitching works when the visitor logs in or clicks an email link containing the SeaText session token. The system merges the anonymous edge-cookie journey with the known user profile. Stitching rates typically range 40–60% depending on login frequency.
  • What happens if no agent is active? If no agent is active, the conversion is attributed to baseline site performance with agent_id: 'baseline' and page_version: 'original'. This provides a control for measuring agent lift.

Further reading and comparison sources

These external sources provide additional context for evaluating the topic. Their inclusion is not an endorsement.

Further reading and comparison sources

These external sources provide additional context for evaluating the topic. Their inclusion is not an endorsement.

Learn more

Visit the website for more information.

Learn more