How It Works

Every time you approve a brief or publish content, ILLIXIS records your decision along with the recommendation features (keyword volume, difficulty, opportunity score, etc.). Seven days later, the system fetches performance data and calculates whether your decision led to good outcomes.

Data sources:

  • Google Analytics 4 - Pageviews, sessions, conversions
  • Google Search Console - Clicks, impressions, CTR
  • Social platforms - Engagement metrics (shares, comments, clicks)

Performance score calculation: The system compares your content's traffic to your tenant average. If a piece of content gets 2x your average traffic, the performance score is 1.0 (maximum). If it gets your average traffic, the score is 0.5. No traffic = 0.0.

What Gets Updated

Preference signals - Each decision you make creates a preference signal. Seven days post-publication, the system updates these fields:

  • Actual traffic - Total pageviews from GA4
  • Actual conversions - Conversions tracked in GA4
  • Performance score - Normalized score (0-1) based on traffic vs. your average

Example: Before feedback You approved a brief targeting a keyword with 2,400 monthly volume, 45 difficulty, and 78 opportunity score. Performance data is not yet available because the content hasn't been live for 7 days.

Example: After 7 days The same signal now includes a performance score of 0.85 (high traffic), confirming it was a good decision. This outcome is fed back into the model.

How Feedback Improves Recommendations

During nightly model retraining (3:15 AM), the preference model uses performance scores to weight recommendations.

Scenario 1: Model predicts approval, you approve, content performs well The model reinforces its feature weights. Next time it sees similar characteristics (high volume + moderate difficulty), it predicts approval with higher confidence.

Scenario 2: Model predicts approval, you approve, content performs poorly The model adjusts weights downward for those features. It learns that high predicted score doesn't always mean good outcome.

Scenario 3: Model predicts rejection, you approve anyway, content performs well The model learns you occasionally make contrarian decisions that succeed. It won't heavily penalize similar opportunities in the future.

Scenario 4: Model predicts rejection, you reject No performance data to collect (content wasn't created). Signal still teaches the model what you don't want.

Checking Performance Feedback

Navigate to Settings → Preference Learning → Training Logs to see how performance feedback affects your model.

Training log fields:

  • Signals processed - How many decisions were included
  • Previous accuracy - Model accuracy before training
  • New accuracy - Model accuracy after training
  • Accuracy change - Improvement or decline

If accuracy improves by 5%+ after a training run, performance feedback is working. The model is learning from outcomes, not just decisions.

Timeline

Day 0: You approve a brief, generate content, publish to WordPress Day 0-7: Content accumulates traffic from GA4, clicks from GSC Day 7: Weekly automated task runs Day 7: System updates the preference signal with actual traffic and performance score Day 8: Nightly training incorporates performance data Day 8+: Future recommendations weighted by past performance

Automation Schedule

Weekly performance update:

  • Schedule: Weekly on Mondays at 6:00 AM
  • What it does: Finds all preference signals created 7+ days ago with no performance data, fetches GA4/GSC metrics, and calculates performance scores

Nightly model training:

  • Schedule: Daily at 3:15 AM
  • What it does: Retrains preference models using all signals including updated performance scores

Integration Requirements

For performance feedback to work, you must have:

  1. Google Analytics 4 connected - Navigate to Settings → Integrations → GA4 to connect
  2. Google Search Console connected - Navigate to Settings → Integrations → GSC to connect
  3. Content published through ILLIXIS - Only content published via CMS connectors (WordPress, Shopify, Webflow, Payload) gets tracked

If GA4 is not connected, the system falls back to GSC data. If neither is connected, performance feedback cannot run.

Interpreting Accuracy Changes

After performance feedback updates run, check your training logs:

Accuracy increased 2-5%: Normal improvement. The model is learning from outcomes and adjusting weights.

Accuracy increased 5-10%: Strong improvement. Performance data significantly refined the model. Your content strategy aligns well with predictions.

Accuracy flat (±1%): The model was already well-trained. Performance data confirmed existing weights rather than changing them.

Accuracy decreased 2-5%: Your recent content performed unexpectedly (either much better or much worse than predicted). The model is recalibrating. Give it 2-3 more training cycles to stabilize.

Accuracy decreased 5%+: Your content strategy may have shifted significantly. Review your top preferences in Settings → Preference Learning to see if they still align with your goals. Consider resetting the model if the shift is permanent.

Common Questions

Why does it take 7 days to update performance? Content needs time to accumulate meaningful traffic. A piece published today may get most of its traffic in the first week. Waiting 7 days ensures we're measuring actual performance, not just launch-day spikes.

Does performance feedback run automatically? Yes. The weekly task runs every Monday at 6:00 AM. No manual action required.

What if I unpublish content before 7 days? The system checks publication status. If content is unpublished or deleted, the signal remains unchanged (no performance score is added).

Can I see which content generated which signal? Navigate to Settings → Preference Learning → Signal History (admin only). Each signal links to its associated brief or content ID.

Does this use API credits? No. GA4 and GSC syncs happen during nightly scheduled tasks, which don't consume user-facing API quotas. Performance feedback uses cached data from those syncs.

What if GA4 data is incomplete? The system uses whatever data is available. If GA4 shows 0 pageviews but GSC shows 50 clicks, it estimates pageviews as clicks x 10 (rough approximation). If neither has data, no performance score is assigned and the signal isn't weighted.

Can I disable performance feedback? Yes. Navigate to Settings → Preference Learning → Disable. This stops all preference learning, including performance updates. Existing signals are preserved but no new signals are created.

Related Features

  • Preference Learning Overview - See guide #21 for how the entire learning system works
  • GA4 Integration - See guide #12 for connecting Google Analytics
  • GSC Integration - See guide #11 for connecting Search Console
  • Model Training Logs - See Settings → Preference Learning → Training Logs to view accuracy over time

Questions? Email support@illixis.io or ask Maya (bottom-right chat icon).

Ready to lose the stack?

One platform. You approve. ILLIXIS executes. Marketing that just happens.

Join the waitlistNo spam, everUnsubscribe anytime
First 20 founding members: 50% off any plan for your first year.

Marketing, Unstacked.