Citability Audit

1. What a citability audit measures

A citability audit examines one coherent surface, such as a page, template, or content cluster. It tests each passage to determine whether an AI engine can use the retrieved text in a synthesized answer without relying on surrounding context. The audit applies the seven structural signals defined in Citability: self-contained chunks, direct-answer blocks, Q&A, steps, citable tables, heading discipline, and liftable quotes. A manual chunk-extraction test (§4) then checks the actual text in the selected engine.

The result is a matrix that rates each sampled chunk and signal as passing, failing, or partial. Every finding receives a severity level and one line of rewrite guidance. Research on answer visibility supports the underlying principle that substantive content is more useful than keyword tricks; Aggarwal et al. 2024 discusses the scope and limits of that evidence. Microsoft described the operational shift in May 2026: “The unit of value shifts from documents to groundable information” (Bing: Evolving role of the index). For this audit, the relevant unit is the passage rather than the entire page.

“Citability audit” is a generic description, not a coined term. The combination of the seven signals in Citability §4 and a chunk-extraction test is GEO Wiki’s organizing method, not an established standard. Within a Full GEO Audit, the resulting findings provide detailed evidence for the content and trust assessment in Layer 4.

2. Set the surface, sample, engine, and baseline

Four decisions determine how to interpret every later finding. Set them before collecting data so the report does not combine results that cannot be compared. The Full GEO Audit and AI Citation Tracking use the same principle.

When to run it. Use this audit when Layer 4 of the Full GEO Audit identifies a page that was read but not cited, or run it quarterly on flagship and evergreen pages. Other triggers include a restructuring of the content, a CMS or template migration, a major rewrite of an answer block, or new tracking evidence that a competitor is cited for your queries while you are not.

Inputs to gather before you start. Collect the fetched HTML that a crawler receives, not only the DOM rendered in a browser (see AI Crawlers). You also need the page’s headings as a flat outline and a fresh incognito session in the engine selected above.

3. Review the seven audit signals

Review the seven structural signals in a consistent order and with consistent definitions. For each signal, ask whether the passage meets it and how a failure would affect the engine’s ability to ground an answer in that passage.

Signals 1, 2, and 7 usually have the greatest practical effect because they determine whether any passage can be used on its own. Signals 3 and 4 apply only when Q&A or HowTo content suits the page. Do not add either format when it does not belong; that is one of the counterproductive fixes in §7. Signal 5 matters more as the amount of tabular content increases, while Signal 6 is quick to assess and correct. Review the signals in order, but rank the findings by severity rather than by signal number (§8).

Six of the seven checks can be automated for a single page. The free Citability Checker evaluates Signals 1–6 from raw HTML without JavaScript and quotes the passage behind each failing verdict. Signal 7 and the judgment calls identified in the tables below remain manual.

4. Step 1: Run the chunk-extraction test

The manual chunk-extraction test provides the most direct evidence in the audit. Remove one passage from its context, paste it by itself into ChatGPT search or Perplexity, and ask the engine to summarize or answer using only that text. If the engine fills in missing information, asks for context, or completes the thought incorrectly, the chunk is not self-contained. The other checks estimate whether a passage can be extracted successfully. This test shows how the selected engine handles the actual text.

Follow this procedure:

1. Choose the sample. Select a representative set of 5–8 chunks based on the decision in §2: the TL;DR, the opening paragraph under one H2, one table row, one FAQ answer, and one quotable claim from the middle of the page.
2. Extract each chunk verbatim. Remove the surrounding paragraphs and headings, along with references such as “as above” or “see §3.” The isolated passage represents the text available to an AI engine after retrieval.
3. Paste the passage into a fresh engine session. Use a session with no prior turns, history, or system prompt because personalized sessions are not reproducible. Ask, “What is this passage saying?” or “Use only this passage to answer: [the page’s target query].”
4. Classify the chunk as one of three outcomes:
   • ✅ Liftable: The engine answers clearly using only the chunk.
   • ⚠️ Partial: The engine hedges, asks for context, or reconstructs the missing setup incorrectly. Record exactly what context was missing.
   • ❌ Broken: The engine cannot interpret the chunk because of a problem such as a pronoun chain, a table without a caption, or a hedged multi-clause sentence.
5. Record the failure pattern by signal. Map each partial or broken result to one of the seven signals in §3. That mapping becomes the finding.

Use the following fields for each audit row:

audit_date          UTC date of the audit
page_url            URL of the audited page
chunk_excerpt       first 120 chars of the lifted passage
signal_n            which of the seven signals (1–7) the failure maps to
outcome             liftable | partial | broken
failure_shape       short note (pronoun ref / no caption / hedged / …)
severity            blocker | major | minor (see §8)

4.1 Worked example: three versions of one passage

The following example compares three versions of the same fact about robots.txt. Each version is pasted by itself into a fresh ChatGPT search session with the prompt “What is this passage saying?”

Version ✅ (liftable). “A robots.txt file is a plain-text file at the root of a domain that tells crawlers which URLs they may fetch. Each rule names a user-agent and a path.” The engine restates the definition, location, and structure without hedging or asking for more information. Verdict: Signal 1 passes.

Version ⚠️ (partial). “It tells crawlers which URLs they may fetch — see the earlier diagram for the path-matching rules.” The engine hedges: “This appears to describe a file that controls crawler access, but the subject is unclear — what file is being referenced?” The pronoun and the reference to “the earlier diagram” leave the subject unresolved. Verdict: Signal 1 is partial because the pronoun and external reference break self-containment.

Version ❌ (broken). “As above, it applies; but as we noted in §2, the precedence rules can override.” The engine cannot identify the subject or operation and asks for the original document. Verdict: Signal 1 is broken because the passage contains only a pronoun chain with no explicit anchor.

Repeat the exercise with the TL;DR, the opening paragraph under an H2, and a table row to create a set of per-chunk records. Tools that assign citability scores without this test infer outcomes from surface features; this test records how the selected engine handles the passage directly.

5. Step 2: Review each signal

For each signal, record the audit question, what good looks like, the failure pattern, and the rewrite guidance. Citability defines the signals; the chunk-extraction test shows how to apply them.

5.1 Signal 1: Self-contained chunk

5.2 Signal 2: Direct-answer / TL;DR block

5.3 Signal 3: Q&A / FAQ structure

Apply this signal only when Q&A suits the page. Inventing questions solely to satisfy the audit creates a counterproductive pattern. For the relationship between question-shaped headings and query fan-out, see Answer Loop §3.1.

5.4 Signal 4: Step / HowTo structure

5.5 Signal 5: Citable table / list

Microsoft is explicit: “Clear headings, tables, and FAQ sections help surface key information and make content easier for AI systems to reference accurately” (Bing AI Performance).

5.6 Signal 6: Heading-hierarchy discipline

5.7 Signal 7: Liftable quotable sentence

6. Account for differences among surfaces

The seven signals apply across surfaces, but the importance of each failure changes. Use these differences when choosing which engine to test.

A clean result on one surface does not guarantee the same result on another. Language also affects the outcome in practice: chunk and answer-block citability differ between Chinese and English. See Multilingual GEO.

7. Avoid fixes that create other problems

Some changes appear to fix an audit finding but create a different problem with spam or trust signals. Citability §6 explains the broader risks of over-optimizing for citability. The following changes can reduce trust even when they appear to address a structural signal.

Google addresses this directly in its May 2026 optimization guide: “There’s no requirement to break your content into tiny pieces for AI to better understand it. Google systems are able to understand the nuance of multiple topics on a page” (AI Optimization Guide). Dividing content too aggressively may appear to satisfy Signal 1 while removing the coherent context that the signal is meant to protect.

Citability is necessary but not sufficient. Systems can detect and penalize structure that lacks substance, and trust problems can still undermine otherwise well-structured content. Many of these rewrites also lose their effect when competitors adopt them; see C-SEO Bench.

8. Score findings and prepare the report

The report’s primary result is a pass, fail, or partial verdict for each signal, recorded for every audited chunk and paired with a severity level. Broken Signals 1, 2, and 7 are major by default. Signals 3, 4, and 6 are minor by default unless the page uses the wrong form entirely. The severity of Signal 5 depends on how much of the page is tabular. As in Full GEO Audit §5, severity depends on the failed signal rather than a single global rubric.

Do not reduce the audit to a score from 0 to 100. The market examples reviewed here do not publish their formulas. Topify offers “a 0–100 grade of how AI-ready your website is” without disclosing its method. Citability.ai reports a “Combined Score: 62” from three subscores but does not publish their weights. Mangools’ AI Search Grader says the score is “weighted by market share” without identifying those weights. A number based on an opaque method is not an interpretable measurement. The provenance rules in GEO Metrics apply even more strongly here. Per-signal verdicts can be reproduced; a single undisclosed score cannot.

Every report should include a header with the date, audited surface, sampling decision, and tested engine; a matrix of outcomes by chunk and signal with a liftable, partial, or broken verdict in each cell; findings ranked by severity, including the failure pattern from the chunk-extraction test; one line of rewrite guidance for each finding that points to Writing for AI Citation; and, when a previous audit exists, a comparison with the baseline. During a re-audit, rerun the chunk-extraction test only on the changed chunks and carry forward the results for unchanged signals.

9. Check the audit’s validity

Review each item before publishing the report.

• Auditing the browser-rendered DOM instead of the fetched HTML: Client-side-rendered content can hide signals from the crawler. Test what a client without JavaScript receives (SSR for AI Crawlers).
• Sampling only the top of the page: Signal 1 fails most often in the middle of long pages. Sample the opening paragraphs under H2 headings throughout the document.
• Running the chunk-extraction test in a personalized session: Sessions that are logged in or retain history produce results that cannot be reproduced reliably.
• Treating six passing signals as a passing page: A page can pass six of seven signals and still fail to earn citations when Signal 1 is broken.
• Generalizing from a single engine: A clean result in Perplexity does not necessarily apply to AI Overviews. Name the tested engine in the report header.
• Combining locales: Chinese and English pages may produce different results. See Multilingual GEO, and do not apply an audit of one language to the other.
• Reporting an unverified citation as a pass: A quoted passage may not support the claim to which the engine attached it (Liu et al. 2023). Verify the citation with AI Citation Tracking §4.1.

10. Further reading

• Concepts: Citability defines every signal in the audit, while E-E-A-T explains the trust requirements for grounding.
• Related playbooks: Full GEO Audit uses this process for its Layer 4 assessment; Writing for AI Citation provides rewrite methods for each signal; and AI Citation Tracking measures outcomes.
• Engine-specific surfaces: Perplexity and ChatGPT search.
• Research: Aggarwal et al. 2024: GEO: Generative Engine Optimization, followed by the bounded discussion in C-SEO Bench.

References

Academic:

1. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K. & Deshpande, A. (2024). GEO: Generative Engine Optimization. KDD ‘24. arXiv:2311.09735 · ACM DL · paper summary
2. Puerto, H., Gubri, M., Green, C., Oh, S. J. & Yun, S. (2025). C-SEO Bench: Does Conversational SEO Work? NeurIPS ‘25 Datasets & Benchmarks. arXiv:2506.11097
3. Liu, N. F., Zhang, T. & Liang, P. (2023). Evaluating Verifiability in Generative Search Engines. Findings of EMNLP 2023. arXiv:2304.09848

Official platform documentation (verified 2026-05):

• Google Search Central: A new resource for optimizing for generative AI in Google Search · AI Optimization Guide · AI features and your website · Succeeding in AI search
• Microsoft Bing: Evolving role of the index: From ranking pages to supporting answers · AI Performance in Bing Webmaster Tools
• OpenAI: ChatGPT search Help Center
• Perplexity: What is an answer engine, and how does Perplexity work as one?