How SIGNALS measures AI citation readiness
The SIGNALS score isn't based on intuition or conventional SEO wisdom. Every dimension and every weight is grounded in published, peer-reviewed research analyzing actual citation behavior across 2 million AI-generated responses. Here's exactly how it works.
The research behind the score
Each paper below contributed specific findings that shaped the SIGNALS framework. The weights aren't assigned subjectively — they're derived from measured effect sizes in peer-reviewed studies. When the research says something carries a 35% weight, it's because the measured effect size is 35% stronger than the next strongest signal, not because we thought it sounded right.
Discovered Labs — AI Citation Study, 2026
The largest study of AI citation behavior published to date. The central finding is what makes this study different from everything that came before: content alignment (vocabulary matching buyer search intent) is the only page-level signal that survives domain fixed-effects controls. Effect size β=+0.37 at q≈10⁻⁷³.
This is the Simpson's Paradox problem that most AEO practitioners miss. High-authority domains get cited more AND tend to have better content — so naive correlation studies make it look like content quality drives citation. The Discovered Labs methodology controls for domain, and when you do that, almost everything else disappears. Alignment doesn't.
Informs: Alignment dimension weight (35%), interpretation of all other dimensions as secondary to alignment
↗ View Discovered Labs researchPrinceton — Generative Engine Optimization (GEO), KDD 2024
Tested 22 content modification strategies across 10,000 queries in a controlled experiment. Key findings: adding statistics with citations increased AI visibility by 41%, adding expert quotes by 28%, and direct answer openings (BLUF) by 17%. The query fan-out analysis is particularly important — it showed that a single user prompt decomposes into multiple sub-queries internally, meaning a page needs to cover adjacent intents to be cited across all decompositions.
Informs: Grounding dimension (13%), Intent dimension (10%), Language dimension (10%), the BLUF rewrite format
↗ View on PrincetonAgentGEO — Pipeline Failure Analysis, March 2026
Mapped the four stages of Retrieval-Augmented Generation that determine whether a page gets cited: Retrieval, Parsing, Ranking, and Generation. The key finding was that most pages fail at Ranking — not because content quality is bad, but because of vocabulary mismatch with the query. This directly informed how SIGNALS prioritizes fixes: there's no point improving content quality if the page is failing at retrieval or parsing.
Informs: Pipeline diagnosis model, Structure dimension (12%), fix prioritization order
↗ View AgentGEO paperConvertMate — GEO Benchmark Study, 2026
Benchmarked AI citation patterns across 8,000 domains. The findings that most directly shaped SIGNALS: 83% of AI citations come from pages outside Google's top 10 (confirming SEO rank and AI citation are largely decoupled), 68.7% of cited pages use logical H1→H2→H3 heading hierarchy, and brands mentioned on third-party domains receive 6.5× more citations than brands that exist only on their own site.
Informs: Structure dimension (12%), Newness dimension (5%), Substantiation dimension weight (15%)
↗ View ConvertMate benchmarkThe 7 SIGNALS dimensions
Each dimension is weighted by its measured effect on AI citation frequency. The weights sum to 100%. Alignment carries 35% because it's the only signal with a documented causal effect that survives domain-level controls — not because we think content is more important than structure, but because the research is unambiguous about it.
On the Simpson's Paradox problem: Most AEO tools measure correlations between content features and citation rates without controlling for domain authority. High-authority domains get cited more AND tend to have better-structured content — so it looks like content structure drives citation. The Discovered Labs study uses fixed-effects regression and double machine learning to isolate true causal effects. When they do this, almost all signals disappear. Alignment survives.
| Dimension | Weight | What it measures | Source | |
|---|---|---|---|---|
S |
Structure | Heading hierarchy (H1→H2→H3), TL;DR boxes, self-contained sections, comparison tables. 68.7% of cited pages use logical heading structure. | ConvertMate 2026 AgentGEO 2026 |
|
I |
Intent | Coverage of adjacent buyer intents and query fan-out. A single user prompt decomposes into multiple sub-queries — a page must address several to be cited across all decompositions. | Princeton GEO 2024 | |
G |
Grounding | Verifiable claims: statistics with named sources, inline citations, expert quotes. Princeton measured +41% citation frequency from sourced statistics, +28% from named expert quotes. | Princeton GEO 2024 | |
N |
Newness | Visible Last Updated timestamps, specific year references, internal date consistency. Page age effect is real but small: β=+0.05 — meaningful but not a priority fix unless everything else is already addressed. | Discovered Labs 2026 β=+0.05 |
|
A |
Alignment
Dominant signal
|
Does the vocabulary on this page mirror how buyers actually search? Measures lexical overlap between page content and documented buyer search patterns. The only signal with a documented causal effect independent of domain authority. | Discovered Labs 2026 β=+0.37, q≈10⁻⁷³ |
|
L |
Language | Buyer-phrased title and H2s, fluency score, title-to-prompt similarity. Fluency optimization increases AI visibility 15–30%. Title-prompt similarity has a measured effect of β=+0.09. | Princeton GEO 2024 β=+0.09 |
|
S |
Substantiation | Third-party validation: G2/Capterra ratings, Reddit mentions, press coverage, named author credentials. Brands mentioned on third-party domains receive 6.5× more AI citations than brands existing only on their own site. | ConvertMate 2026 Discovered Labs 2026 |
How SIGNALS compares to other tools
Standard SEO tools were built for Google's PageRank. AI visibility trackers measure citation outcomes — they tell you whether you're being cited, not why or what to fix. SIGNALS is built to diagnose and fix the structural causes, not just report the symptom.
| Capability | SEO tools (Ahrefs, Semrush) | AI trackers (Profound, Trackr) | SIGNALS |
|---|---|---|---|
| Measures AI citation readiness | ✗ | ✗ (outcomes, not causes) | ✓ |
| Diagnoses why AI ignores a page | ✗ | ✗ | ✓ Pipeline diagnosis |
| Generates page-specific fixes | ✗ | ✗ | ✓ BLUF, FAQ, H2s, schema |
| Applies fixes to your CMS | ✗ | ✗ | ✓ WordPress, Shopify, Webflow |
| Based on peer-reviewed research | Partially | Partially | ✓ 4 papers, 2M+ citations |
| Works for small sites | ✗ (favors high-DA domains) | ✗ (only tracks top brands) | ✓ Page-level signals, no DA bias |
| How you pay | Subscription only | Subscription only | Free audit → one-time fix plans (readable) → agency retainer (ongoing) |
Frequently asked questions
How is the overall SIGNALS score calculated?
Weighted average of 7 dimension scores, each rated 0–10. Weights: Alignment 35%, Substantiation 15%, Grounding 13%, Structure 12%, Intent 10%, Language 10%, Newness 5%. Score 70+ is good, 45–69 is fair, below 45 is critical. The thresholds are calibrated against the distribution of scores across pages that do and don't appear in AI citation studies.
Why does Alignment carry 35% of the score?
Because the Discovered Labs 2026 study — the largest AI citation analysis published — found it's the only page-level signal with a causal effect independent of domain authority. Effect size β=+0.37 at q≈10⁻⁷³. Every other signal they tested collapsed to near-zero when domain was controlled. We weighted Alignment at 35% because that's what the evidence supports, not because we thought alignment sounded important.
Does a higher SIGNALS score guarantee AI citations?
No, and any tool that claims guarantees is lying. What SIGNALS does is ensure your pages have the structural and content characteristics that the research shows predict citation frequency. It removes the structural reasons you'd be skipped. Think of it like technical SEO — ensuring crawlability doesn't guarantee a #1 ranking, but it removes a category of reasons you'd be excluded.
How does the pipeline diagnosis work?
Based on the AgentGEO framework, AI citation requires passing four stages: Retrieval (can crawlers access the page?), Parsing (can content be extracted from the HTML?), Ranking (does the page beat competitors in the retrieved set?), Generation (is the content quotable as standalone units?). SIGNALS assesses each stage and identifies your primary failure point — because fixing the wrong stage wastes time.
How is SIGNALS different from traditional SEO scores?
Traditional SEO scores measure signals Google's PageRank algorithm rewards: backlinks, click-through rate, page speed. SIGNALS measures signals for AI citation via RAG: vocabulary alignment, structural clarity, quotable content, sourced claims. A page can rank #1 on Google and score 28/100 on SIGNALS. They're optimizing for different algorithms with different reward functions.
Which AI engines does SIGNALS optimize for?
ChatGPT, Claude, Perplexity, and Google AI Overviews. The underlying research covers all four. The signals are consistent across engines because they all use RAG architectures with similar retrieval and ranking mechanisms — the structural signals that predict citation in ChatGPT also predict citation in Perplexity.
How often should I re-audit my pages?
After applying fixes, re-audit immediately to verify the structural changes registered. After that, quarterly is reasonable unless you're making significant content changes. AI citation frequency takes 2–6 weeks to shift after structural changes, so monthly re-auditing during an active optimization period gives you the best signal on what's actually moving.
Related reading
See your SIGNALS score in 60 seconds
Paste any URL. Get a full audit — score, 7-dimension breakdown, pipeline diagnosis, and page-specific fixes. Free for 1 page, no account required.
Audit my page free →No account · No credit card · Results in under 60 seconds