An API that tells you if an ad will work — before you spend

cat > ad.json <<'EOF'
{
  "ad_copy": {
    "primary_text": "Still paying an agency $3k/mo for ads a robot writes better?",
    "headline": "Fire your ad agency"
  },
  "platform": "meta",
  "product_context": "AI Meta-ads automation tool, $99/mo",
  "target_audience": "DTC founders spending $2-20k/mo on Meta ads"
}
EOF

curl -sS -X POST https://www.spendict.com/api/v1/assess \
  -H "Authorization: Bearer spd_live_YOUR_KEY" \
  -H "Content-Type: application/json" \
  --data @ad.json

The response scores seven dimensions and names the failure mode:

{
  "overall_score": 78,
  "launch_recommendation": "run",
  "dimension_scores": {
    "hook": 78, "angle": 80, "clarity": 82,
    "audience_resonance": 80, "platform_fit": 85,
    "cta": 0, "compliance_safety": 95
  },
  "predicted_failure_mode": "No material weakness — would run as-is.",
  "prioritized_fixes": []
}

Most “AI ad scoring” is a prompt away from telling you what you want to hear. Spendict’s server recomputes the overall score and the launch recommendation from fixed gates after the model has judged the ad — a dead hook or a likely policy rejection always caps the verdict at fix_first, and two failing fundamentals always mean kill. The model provides the judgment; the server derives the authoritative verdict. That’s what makes it safe to automate ad spend against.

• ✓Spend gates for AI agents — the agent generates variants, the API kills the weak ones, only run verdicts reach the ads manager.
• ✓Batch scoring — score 50 variants in a loop and rank by overall_score before creative review, so humans only look at the top five.
• ✓Compliance pre-flight — catch likely policy rejections (the compliance_safety dimension) before the platform does.
• ✓Agency QA — a consistent, client-defensible second opinion on every ad that ships.

Free tier: 100 calls/month, no credit card. Paid: $19/mo for 1,500 calls, $49/mo for 5,000, $199/mo for 25,000 — every tool call counts equally, quota is checked before inference, and failed calls are refunded. Errors are always structured JSON ({ error, message }), never a stack trace. Full reference in the docs, plans on the pricing page.