Grounded Search & Citation Prediction

Fast, cost-effective citation prediction before expensive multi-LLM execution.

Overview

Grounded Search enables AIQ to predict whether a query will generate brand citations before running expensive execution across all 4 LLM providers.

The Problem

Running queries across Claude, GPT-4, Gemini, and Perplexity costs $0.05-0.15 per query. Without knowing if a query will generate citations, you waste budget on low-quality queries.

The Solution

Grounded Search uses Gemini's native Google Search tool to predict citations for $0.0001-0.001 – a 98-99% cost reduction compared to full execution.

Key Benefits

Before	After	Improvement
$0.05-0.15 per validation	$0.0001-0.001	98-99% cheaper
15-30s (full execution)	1-3s	90% faster
60-70% accuracy (text parsing)	90-95%	+30% more accurate

How It Works

Three-Tier Provider Strategy

AIQ automatically selects the optimal provider for citation prediction:

Tier 1: Gemini (BEST)
├─ Native Google Search grounding
├─ 95% confidence
└─ $0.000158 per prediction

Tier 2: Perplexity (GOOD)
├─ Native web search + citations
├─ 90% confidence
└─ $0.0006 per prediction

Tier 3: Claude/GPT-4 (FALLBACK)
├─ LLM-simulated citations
├─ 65% confidence
└─ $0.005 per prediction

The system automatically chooses the cheapest provider with native grounding support.

Usage

Basic Prediction

import { CitationPredictionService } from '@/lib/services/citation-prediction-service';

const service = new CitationPredictionService();

const result = await service.predictQueryCitations({
  query: 'What are the key features of Salesforce CRM?',
  organizationDomain: 'salesforce.com',
  organizationId: 'org-123',
});

console.log(result.willCiteBrand); // true
console.log(result.confidence); // 95
console.log(result.cost); // 0.000158
console.log(result.recommendation);
// "HIGH QUALITY: Query likely to generate brand citations..."

Batch Prediction

Test multiple queries at once:

const batchResult = await service.predictBatchCitations({
  queries: [
    'What is Salesforce?',
    'How does Salesforce compare to HubSpot?',
    'Salesforce pricing plans',
  ],
  organizationDomain: 'salesforce.com',
  organizationId: 'org-123',
});

console.log(batchResult.aggregate.expectedCitationRate); // 85%
console.log(batchResult.aggregate.totalCost); // 0.0003

Provider Capabilities

Gemini (Recommended) ⭐

Method: Native Google Search grounding Confidence: 95% (highest) Cost: $0.075/1M tokens (cheapest)

Gemini uses real Google Search during inference and returns structured groundingMetadata with actual citations.

When to use: Always (if available)

Perplexity (Good Alternative)

Method: Native web search + citations API Confidence: 90% Cost: $1-3/1M tokens (13-40x more expensive)

Perplexity performs real web search and returns native citations[] array.

When to use: Gemini unavailable or testing Perplexity-specific behavior

Claude/GPT-4 (Fallback)

Method: LLM-simulated citation prediction Confidence: 65% (lower - simulation not real search) Cost: $2.50-3/1M tokens (33-40x more expensive)

Uses LLM to simulate which sources it would cite when answering a query.

When to use: No providers with native grounding available

Understanding Results

Prediction Confidence

Confidence	Meaning	Action
90-95%	High confidence (native grounding)	Trust prediction
70-89%	Medium-high (good simulation)	Mostly reliable
60-69%	Medium (simulation)	Consider actual execution
< 60%	Low confidence	Run actual execution

Recommendations

The service provides automatic recommendations:

• "HIGH QUALITY" → Query likely to cite brand, recommended for execution
• "MEDIUM QUALITY" → May cite brand, consider refining
• "LOW QUALITY" → Unlikely to cite brand, recommend regeneration
• "UNCERTAIN" → Low confidence, test with actual execution

Cost Analysis

Per-Query Comparison

Provider	Prediction Cost	Full Execution Cost	Savings
Gemini	$0.000158	$0.15	99.9%
Perplexity	$0.0006	$0.15	99.6%
GPT-4	$0.00525	$0.15	96.5%
Claude	$0.0078	$0.15	94.8%

Monthly Savings Example

Scenario: 1000 queries/month, 20% rejected by prediction

Metric	Without Grounded Search	With Grounded Search	Savings
Prediction cost	$0	$0.158	-
Execution cost	$150 (1000 × $0.15)	$120 (800 × $0.15)	$30
Total	$150	$120.158	$29.84/month

Annual savings per organization: ~$358/year 100 organizations: ~$35,800/year

Best Practices

1. Pre-Validate Before Execution

Always predict citations before running expensive multi-provider execution:

// Step 1: Predict
const prediction = await service.predictQueryCitations({
  query: userQuery,
  organizationDomain: 'example.com',
  organizationId: orgId,
});

// Step 2: Only execute if prediction is positive
if (prediction.willCiteBrand && prediction.confidence >= 70) {
  // Run expensive multi-provider execution
  await executeQuery(userQuery);
} else {
  // Save budget - don't execute low-quality query
  console.log('Query unlikely to cite brand - skipping execution');
}

2. Batch Validation

For query generation, validate entire batch before execution:

const generated = await generateQueries({ industry: 'SaaS', count: 20 });

const batchPrediction = await service.predictBatchCitations({
  queries: generated.queries.map(q => q.text),
  organizationDomain: 'example.com',
  organizationId: orgId,
});

// Only execute queries with positive predictions
const highQualityQueries = batchPrediction.results
  .filter(r => r.willCiteBrand && r.confidence >= 70)
  .map(r => r.query);

console.log(`Executing ${highQualityQueries.length} of ${generated.queries.length} queries`);

3. Track Prediction Accuracy

Compare predictions to actual execution results:

const prediction = await service.predictQueryCitations(params);
const execution = await executeQuery(params.query);

const actualCited = execution.citations.some(c =>
  c.domain.includes(params.organizationDomain)
);

// Log accuracy for analytics
console.log('Prediction correct:', prediction.willCiteBrand === actualCited);

Troubleshooting

"No providers available"

Cause: Organization has no enabled LLM providers Solution: Configure enabledProviders in organization settings or set environment variables

"No API key found for provider"

Cause: Missing API key for selected provider Solution:

1. Set environment variable (GEMINI_API_KEY, etc.)
2. OR configure organization-specific encrypted key

Predictions don't match actual execution

Possible causes:

1. Time delay - Web search results change over time
2. Provider differences - Different providers index differently
3. Simulation error - Claude/GPT-4 simulation less accurate

Solution: Use Gemini grounded search for 95% accuracy

"Prediction confidence is low"

Cause: Using simulated grounding (Claude/GPT-4) Solution: Enable Gemini or Perplexity for higher confidence

Advanced Usage

Multi-Provider Comparison

Test prediction accuracy across providers:

const comparison = await service.compareProviderPredictions({
  query: 'Best CRM for small business',
  organizationDomain: 'salesforce.com',
  organizationId: 'org-123',
  providers: ['gemini', 'perplexity'],
});

for (const [providerName, result] of comparison) {
  console.log(`${providerName}: ${result.willCiteBrand} (${result.confidence}% confidence)`);
}

Custom Provider Selection

Force specific provider:

const result = await service.predictQueryCitations({
  query: 'What is TypeScript?',
  organizationDomain: 'typescriptlang.org',
  organizationId: 'org-123',
  providers: ['perplexity'], // Force Perplexity
});

Related Features

• Query Generation Modes - Quality validation during generation
• Multi-Perspective Testing - Segment predictions by user intent
• Citation Analytics - Track actual execution results

Technical Details

For developers and technical users:

• Implementation: lib/services/citation-prediction-service.ts
• Provider Clients: lib/api/grounded-search/
• Type Definitions: types/grounded-search.ts
• Documentation: docs/features/GROUNDED_SEARCH.md