Content Variants

What are content variants

Content variants are different versions of your site content, each optimized for a specific agent intent or query context. Rather than serving a single static description of your product to every AI agent, Inception Agents generates and tests multiple framings of the same underlying information.

Each variant is factually equivalent — the same product, the same capabilities, the same pricing. The difference is emphasis, structure, and framing, tailored to what the agent is looking for.

Variant types

The platform generates five variant types for each content chunk:

feature_led

Highlights product capabilities and technical specifications. Leads with what the product does, how it works, and what distinguishes it technically.

Best for: agents answering “what does X do?” or “what features does X have?” queries.

outcome_led

Focuses on results, use cases, and business impact. Leads with what users achieve with the product rather than how it works internally.

Best for: agents answering “how can I solve Y?” or “what’s the best tool for Z?” queries.

comparison_led

Positions the product against alternatives and competitors. Structures content around differentiators, advantages, and category context.

Best for: agents answering “X vs Y” or “best alternatives to Z” queries.

trust_led

Emphasizes credibility signals — certifications, uptime records, customer count, review scores, compliance standards, and enterprise readiness.

Best for: agents answering “is X reliable?” or “is X enterprise-ready?” queries, and for shopping agents evaluating vendor trustworthiness.

concise_led

A compressed summary optimized for token efficiency. Contains the essential facts in the fewest tokens possible, without sacrificing accuracy.

Best for: agents with tight context windows, multi-step reasoning chains, or aggregation queries comparing many products at once.

How variants are generated

After the initial site crawl, the variant generation pipeline runs:

1. Content extraction — Each page is broken into logical content chunks (product descriptions, feature sections, pricing blocks, testimonials).
2. Variant synthesis — For each chunk, the platform generates all five variant types using AI, constrained by the factual content of the source material.
3. Honesty verification — Every generated variant passes through the Honesty Engine to ensure factual accuracy and claim consistency.
4. Baseline assignment — The most balanced variant is set as the initial default for each chunk.

Variant generation runs automatically after each crawl cycle. New or changed content triggers regeneration of affected variants.

How variants are selected

The learning engine determines which variant to serve for each request based on:

• Agent identity — Which platform is making the request (ChatGPT, Perplexity, Claude, etc.).
• Query intent — Inferred from request context, referrer data, and query parameters when available.
• Historical performance — Which variant has produced the best outcomes (recommendations, click-throughs, conversions) for similar requests.
• Confidence threshold — New or low-data scenarios default to the balanced variant until sufficient signal accumulates.

Selection happens at the edge with sub-millisecond latency. There is no perceptible delay added to response times.

Per-platform optimization

Different AI agents respond to different content styles. The learning engine tracks performance independently for each platform:

• ChatGPT may favor outcome-led content for product recommendation queries.
• Perplexity may prefer concise-led content for its citation-heavy response format.
• Claude may respond best to feature-led content for technical evaluation queries.
• Shopping agents (Rufus, Copilot Shopping) often prioritize trust-led and comparison-led variants.

These patterns are learned automatically. You do not need to configure per-platform preferences — the system discovers what works through continuous testing.

Automatic improvement

The optimization loop runs continuously:

1. A variant is served to an agent request.
2. Downstream signals are collected — was your product recommended? Did the user click through? Did they convert?
3. Performance scores update for that variant + agent + intent combination.
4. Future requests for the same pattern receive the highest-performing variant.
5. Low-confidence combinations continue to explore alternate variants to discover better options.

Over time, each content chunk converges on the optimal variant for each agent and intent pair. The system continues exploring at a low rate to detect shifts in agent behavior.

Dashboard view

The Optimization tab in your dashboard shows variant performance:

• Variant distribution — What percentage of traffic is receiving each variant type.
• Performance by variant — Recommendation rate, click-through rate, and conversion rate per variant.
• Per-platform breakdown — How each variant performs across different agent platforms.
• Trend lines — Performance changes over time, with annotations for when variant selections shifted.
• Content chunks — Drill into individual content chunks to see which variant is currently winning.

Manual overrides

You retain full control over variant selection:

• Lock a variant — Force a specific variant type for a content chunk, bypassing the learning engine.
• Disable a variant — Remove a variant type from the testing pool for specific content.
• Create custom variants — Write your own variant and add it to the testing pool alongside the auto-generated options.
• Set platform-specific overrides — Lock a variant for a specific agent platform while allowing optimization for others.

Overrides are configured per content chunk in the dashboard under Optimization > Content Chunks > [chunk] > Variants.