Every AI answer is the product of a silent scoring process. When a user asks ChatGPT, Claude, Perplexity, or Gemini a question, the model retrieves candidate passages, scores them, and cites the winners. That scoring process is what we call Signal Weighting — the third phase of Adaptive RAG — and it decides, in practice, whether your brand appears in the answer.
This guide explains how Signal Weighting works, what it means for the content elements you control (headings, schema, citations, freshness), and includes a downloadable reference table for practitioners.
The Core Idea: Not All Content Is Weighted Equally
Retrieval systems do not read your page like a human. They convert content into vectors, retrieve chunks by semantic similarity, and then score those chunks against the query. The score is a weighted combination of signals. A heading that states your brand's core claim does more retrieval work than a paragraph buried in the middle of a 2,000-word article — not because headings are "better writing," but because they carry more structural weight in retrieval.
RAG Signal scores brand signals across seven dimensions. The weights reflect how much each dimension influenced citation outcomes in our 2025 controlled experiment (63 prompts × 4 LLMs):
| Dimension | Weight | What it measures |
|---|---|---|
| Source Authority | 25% | Trustworthiness of the sources around your claims |
| Factual Consistency | 20% | Internal consistency + corroboration across sources |
| Entity Linkage | 15% | Density and clarity of entity definitions |
| Cross-Model Persistence | 12% | Cited consistently across all 4 models |
| Temporal Freshness | 12% | Recency of content relative to the query |
| Citation Frequency | 10% | How often the brand is referenced elsewhere |
| Competitive Diff | 6% | How specifically the content answers this query vs competitors |
How Content Elements Are Weighted: Concrete Examples
1. Headings — Entity Linkage & Competitive Diff
Consider a prompt: "What is the best corporate video production agency in London?"
- Weak heading: "Our Services" — carries no entity or category signal. Low retrieval value.
- Strong heading: "Corporate Video Production in London — 35+ Enterprise Clients, 97% Retention" — states the entity (corporate video), the geography (London), and the proof (client count, retention). The model can retrieve this chunk and answer the query from it alone.
Headings are where Entity Linkage (15%) and Competitive Diff (6%) are won or lost. A heading that names your category and differentiator converts a generic page into a citable answer.
2. Schema — Source Authority & Entity Linkage
Structured data tells the model what your content is. A page with Organization schema (name, logo, sameAs, knowsAbout) and Article schema (headline, author, dates) gives the model machine-readable certainty about your entity. In the 2025 experiment, content with explicit entity markup and structured context achieved an 81% citation rate vs a 22% baseline — schema was the single highest-leverage structural change.
- Weak: no schema, or schema that doesn't match the visible content (a "FAQ" block with no questions on the page).
- Strong: Organization + Article + FAQPage schema that mirrors what a human sees. Models can extract entity relationships directly.
3. Citations — Citation Frequency & Source Authority
"Citations" here means third-party references: industry directories, press mentions, credible external pages that name your brand. Each external reference is a retrieval anchor. In our companion audit, brands with zero backlink authority but RAG-ready content were cited at 34% — while brands with DA 70+ but poorly structured content were cited at only 12%. Authority transfers in AI through structured, consistent references — not raw link equity.
- Weak: your brand appears only on your own site.
- Strong: your brand appears consistently, with the same facts, on 5-10 credible third-party sources. Each one raises the retrieval confidence of every future query.
4. Freshness — Temporal Freshness
Models favor recent, dated content — especially for queries with a temporal component ("best X in 2026", "latest AI trends"). A page last updated in 2024 scores lower on Temporal Freshness (12%) than an identical page updated this month, even if the core content hasn't changed.
- Weak: undated content, no update history, stale statistics.
- Strong: dated publication + visible last-updated dates + current statistics with sources. Date-stamped updates are a cheap, compounding freshness signal.
Weighting Is Per-Model, Not Global
Critically, the weights are not static. ChatGPT's retrieval behavior differs from Gemini's — we re-calibrate the weighting per model and per prompt cluster in every engagement. A brand can be strong on ChatGPT and invisible on Perplexity because the underlying scoring differs. This is why Cross-Model Persistence is tracked as its own dimension: single-model wins can mask a persistence problem.
Download the Reference Table
For practitioners, we publish the full scoring reference: signal-weighting-reference.csv — all seven dimensions with weights, what each measures, and concrete improvement actions per dimension. Use it to audit your own content before you spend a single dollar on more publishing.
Where to Go Next
- Read the full methodology (MAP → BUILD → WEIGHT → REINFORCE → MEASURE)
- See the 2025 controlled experiment that produced these weights
- Understand Brand Memory — what the weighting is applied to
- Get your own Citation Baseline — the same framework applied to your brand