How ASEO Works and How to Measure It Properly

If you've been following the GEO (Generative Engine Optimisation) space, you've probably noticed a pattern: a lot of tools that claim to measure AI visibility either return empty results, break on anything beyond a demo query, or give you a number with no explanation of what caused it.

This post is the explainer for what's actually happening under the hood when AI platforms decide whether to cite your content - and how to measure it properly, not just track it.

The retrieval mechanism most "AI SEO" tools ignore

When ChatGPT, Perplexity, or Gemini answers a query, it doesn't rank your page. It runs a retrieval-augmented generation (RAG) pipeline.
The pipeline works roughly like this:

1. The query gets embedded into a vector
2. Your content has been pre-chunked into segments (typically 134-167 words each) and embedded into the same vector space
3. The retrieval system does a nearest-neighbour search and pulls the highest-scoring candidate chunks
4. The generator model uses those chunks to compose its answer, citing the sources they came from

The selection decision happens at chunk level, not page level. This is the detail that breaks most "AI visibility" tools - they measure whether your brand appeared in AI responses, but they don't measure whether your specific content was retrieved or why.

That distinction matters. A brand mention can come from training data, third-party coverage, or a Reddit thread. A citation means a specific URL was retrieved at query time and used as source material. Different mechanism. Different intervention.

The five signals that determine chunk retrieval

These are the variables the CPS® (Citation Probability Score®) framework measures. Each one is derived from RAG system behaviour, not from a checklist:

1. Content Structure Is the chunk 134-167 words? Does it open with a declarative answer rather than scene-setting? RAG systems embed chunks and score them for query relevance. A chunk that spends its first two sentences restating context before making a point loses embedding density on the actual signal.
2. Fact Density How many named entities, statistics, percentages, and verifiable claims appear per 100 words? AI retrieval models consistently weight fact-rich passages 2-3x higher than descriptive prose. The GEO study (KDD 2024) found that adding statistics to content improved AI visibility by approximately 41%.
3. Answer Structure Does the chunk open with the declarative pattern: [Topic] is/means/works by [specific mechanism]? Retrieval systems are built to match content to query intent. A chunk that buries its answer after three sentences of context loses ground against one that states the answer on line one.
4. Self-Containment Can this chunk be read in isolation - without the paragraph above it, the section heading, or the image next to it? RAG systems extract chunks without surrounding context. "As mentioned above" or "refer to the diagram" are silent retrieval failures. The chunk gets deprioritised because it's incomplete as a standalone unit.
5. Freshness Signals Does the chunk carry date markers - "as of Q1 2026", updated statistics, schema timestamps? Perplexity and Bing-powered AI search weight recency heavily. An undated chunk competes at a disadvantage against the same information with a temporal anchor.

What a proper measurement stack looks like

Most tools give you one of these. A proper ASEO measurement stack needs all four:

Brand mention rate : does your brand name appear in AI responses? Tracked by most monitoring tools. Tells you about training data recall and third-party coverage, not retrieval performance.
Citation rate : was a specific URL from your domain used as a source? This is the retrieval signal. Requires running structured prompts, parsing responses for URL citations, and matching against your domain.
Hallucination rate : when AI mentions your brand, what percentage of factual claims are incorrect? Most tools don't track this at all. It requires cross-checking AI-generated brand claims against a verified fact schema. We call this Verified Brand Facts.
Funnel-stage SOV : are you appearing on Awareness queries ("what is X"), Consideration queries ("X vs Y"), and Decision queries ("best X for Y") separately? Aggregate visibility masks the fact that most brands dominate Decision-stage queries (where buyers already know them) and are invisible at Awareness and Consideration (where preferences form).

The practical audit sequence

Before writing a word of new content:

1. Run 75-100 structured prompts across ChatGPT, Perplexity, Gemini, Claude, and Copilot - covering your category's Awareness, Consideration, and Decision query patterns
2. Record whether your brand appeared, at what position, as a mention or a citation, and with what sentiment
3. For every cited URL, check whether that specific page would score Grade B or above on the five CPS® pillars
4. Flag every response where AI made a factual claim about your brand and cross-check against ground truth

That sequence tells you which pages are actually generating citations, which chunks on those pages are doing the work, and whether what AI says when it cites you is accurate.

Free tools worth running first

If you want to test your own content before committing to a full audit:

CPS® Block Scorer : paste any paragraph, get a 0-100 score across all five retrieval pillars in under 30 seconds. No signup.
citedbyai.info/cps-scorer

AI Crawler Simulator : choose GPTBot, ClaudeBot, or PerplexityBot and see exactly what each crawler can read on your site, what it's blocked from, and what signals it's missing. citedbyai.info/ai-crawler-simulator (scroll to free tools)

llms.txt Generator : crawls your sitemap, classifies pages, writes contextual AI descriptions automatically. Outputs spec-compliant llms.txt and llms-full.txt.
citedbyai.info/llms-generator

AI Brand Accuracy Check : 5 prompts against Claude, claims cross-checked against your homepage. Flags any factual errors AI is currently making about your brand. citedbyai.info/ai-accuracy-check

The research foundation
The CPS® framework isn't a heuristic. It's built from peer-reviewed retrieval research (including the GEO study accepted at KDD 2024), Ahrefs' analysis of 17 million citations, Wix's analysis of 75,000+ AI answers, and Seer Interactive's log file data on AI crawler behaviour.
The full research foundation with sources and caveats is documented here: citedbyai.info/cps-research-foundation

What this isn't

This isn't "add FAQ schema and you'll get cited." Schema helps. It's necessary at the infrastructure layer. But it doesn't determine which chunk gets selected when two competing pages both have clean schema and decent domain authority.

This also isn't "write more content." Publishing new pages that score below Grade D on Answer Structure and Fact Density adds to your crawl budget, not your citation rate.

The retrieval decision happens at the chunk level. Measuring and improving at the chunk level is the only intervention that directly moves the signal.

Cited By AI® is a UK-based ASEO consultancy. The CPS® (Citation Probability Score®) framework is built on peer-reviewed retrieval research and validated against observed citation outcomes across ChatGPT, Perplexity, Gemini, Claude, and Copilot. Citation Probability Score® and CPS® are registered UK trade marks.