Teams often begin AI search measurement with a citation report: was our URL cited, how often, and by which answer engine? That is useful, but it is downstream measurement. A page cannot be cited unless a system can access it, extract usable content from it, identify a relevant passage and consider that passage a credible candidate for a particular query.
AI search retrieval eligibility is the evidence-based assessment of whether a page could plausibly enter that candidate set before an answer system chooses whether to cite it. It is not proof that a model retrieved the page, and it is not a promise of citation. Most systems do not expose their retrieval logs or ranking thresholds. It is, however, a practical way to avoid diagnosing every missing citation as an authority problem.
The distinction matters. A technically blocked page needs an access fix, not another expert quote. A readable but vague page needs clearer answer passages, not schema markup alone. A strong passage that does not match the query’s actual subquestions may need editorial coverage, not a change to its title tag.
Why citation is the wrong first diagnostic
Citation is a source-selection outcome. It may reflect retrieval, passage relevance, corroboration, freshness, source diversity, interface rules and the answer engine’s own synthesis choices. A source can be retrieved without being shown, shown without a linked citation in every interface, or cited for a narrow claim while losing the broader answer.
Use the taxonomy in AI Search Measurement Taxonomy to keep retrieval, inclusion and citation as separate observations. Collapsing them creates misleading remediation plans.
In my view, eligibility should be treated as a pre-flight check. It cannot reveal a proprietary ranking system, but it can establish whether the obvious prerequisites and content signals are present. That makes citation tests more interpretable and keeps engineering, SEO and editorial work pointed at the right layer.
A four-layer model of retrieval eligibility
Assess a page at four distinct layers. Record a pass, fail or uncertain result for each test, along with the date, tested URL, query, locale and device or account context where applicable.
| Layer | Question | Useful evidence | What a failure usually means |
|---|---|---|---|
| Technical access | Can an eligible crawler fetch the canonical content? | HTTP response, robots directives, rendered-page checks, logs where available | Access or indexing barrier |
| Content extraction | Can the principal claims be recovered as meaningful text? | Rendered DOM, text extraction, source-versus-render comparison | Parsing, rendering or template problem |
| Passage relevance | Does a self-contained passage answer a specific need with support? | Claim audit, headings, citations, passage review | Weak answer unit or unsupported claim |
| Query-level eligibility | Is that passage a plausible candidate for this query and its subqueries? | Query decomposition, result comparison, controlled answer tests | Intent, scope, freshness or topical mismatch |
Do not turn this into a single universal score too early. A score can help prioritise work, but the layer-level evidence is the diagnosis. A page with an access failure is not meaningfully offset by excellent passages.
Layer 1: establish technical access without overclaiming bot support
Start with the URL an answer engine is most likely to encounter: the canonical, indexable version. Confirm that it returns a successful HTTP response to ordinary requests, resolves consistently, and does not depend on a login, location gate or fragile client-side sequence to reveal its main content.
- Check robots.txt and page-level robots directives for relevant disallow, noindex or nosnippet restrictions.
- Check the canonical target, redirects, duplicate variants and accidental parameter handling.
- Compare raw HTML with a rendered page. Important copy that appears only after interaction is a risk.
- Review server logs where lawful and available. Bot user agents and IP verification need careful interpretation; an unrecognised visit is not proof of non-crawling.
- Test status codes and content delivery from representative markets if the page is internationally targeted.
Google’s guidance on AI features in Search is useful here because it explains that the same foundational Search controls affect eligibility for applicable AI features. It does not establish how every answer engine retrieves or cites pages. For Perplexity-specific operational references, consult its published bot guidance, then verify your implementation rather than assuming a generic crawler rule applies everywhere.
Call this layer “access eligibility,” not “AI crawlability.” The latter implies evidence most sites do not possess. Record the exact directive and test result, especially when CDN rules, bot mitigation or country routing may differ.
Layer 2: test whether the page is extractable as content, not merely visible in a browser
A browser can display a polished page whose core information is difficult to isolate. Dense navigation, repeated promotional modules, accordions, image-only tables, cookie overlays and JavaScript-inserted copy can blur the main answer. Search and answer systems use different pipelines, so no local extraction test replicates them exactly. It can still expose avoidable weaknesses.
Export the rendered DOM and create a plain-text view. Ask four direct questions: does the page title remain attached to the main topic; do headings preserve hierarchy; do key claims survive outside scripts and interface controls; and can a reader identify where each important claim begins and ends?
Tables deserve special attention. Use real HTML table markup for comparative data, clear headers and nearby definitions. Do not put decisive facts inside an image. Likewise, accordions can be appropriate for usability, but do not hide the only concise answer behind an interaction that fails in a basic rendered capture.
Structured data may clarify entities and page types, but it is not a substitute for extractable prose. Google’s structured data introduction is clear that markup helps Google understand page information and can make a page eligible for certain Search features; it does not promise visibility. For an experimental approach, see Does Structured Data Improve AI Search Visibility?.
Layer 3: audit passages as answer units
Retrieval frequently operates on chunks, passages or sections rather than a page as a single indivisible object. A topically relevant URL may therefore remain a poor candidate if its useful statement is split across several paragraphs, qualified only in a footnote, or surrounded by unrelated copy.
Build a claim-level inventory for priority pages. For each passage, capture the claim, supporting evidence, scope, date sensitivity, named entities and the question it answers. A well-formed passage usually states the answer early, defines any necessary condition, and links or attributes the evidence close to the claim.
This does not mean every paragraph should be an overconfident mini-summary. Important limits belong in the same answer unit. “Tool X supports export” is weak if export is available only on a higher plan, in selected countries, or through an integration. An extractable but materially incomplete passage can create the wrong kind of visibility.
Teams that need a repeatable audit can adapt the method in How to Run a Claim-Level Content Evidence Audit for AI Search. The goal is not stylistic uniformity. It is a clean mapping between claim, evidence and intended question.
Layer 4: measure query-level eligibility, not generic topical relevance
A page can be eligible for “how does employee monitoring software work?” and ineligible for “which employee monitoring tools offer on-premises deployment in Germany?” The second query carries constraints: comparison, feature, deployment model and market. Generic topic coverage is not enough.
Create a small query set by intent and annotate each query with its likely answer requirements. For complex prompts, map the probable subqueries: definition, eligibility condition, comparison criterion, current policy, method or local constraint. This is especially important when answer engines fan out a prompt into several searches; this guide to AI search query fan-out provides a practical planning model.
Then compare your candidate passage with the answer requirements and with pages currently surfaced in conventional results or cited answer outputs. Do not mistake those outputs for a complete retrieval trace. Use them as competitive evidence: what claim form, source type, recency and specificity appear to be winning?
A worked claim-level example
Consider a software company testing the query: “Does Acme Monitor offer on-premises deployment for German companies?” Its product page contains the sentence, “Acme Monitor offers flexible deployment options.” A competitor’s documentation says, “On-premises deployment is available for Enterprise customers in the EU, including Germany,” followed by a dated implementation guide.
Audit the Acme page as one candidate passage. Technical access passes: status 200, indexable canonical, main copy visible in rendered HTML. Extraction passes: the deployment sentence appears in the plain-text capture. Passage relevance scores 1 out of 4: it names deployment but not on-premises, Germany, plan eligibility or evidence. Query-level eligibility scores 1 out of 4 because it cannot satisfy the geographic and product-condition constraints.
For a simple prioritisation indicator, use a gated calculation: access (pass = 1) × extraction (pass = 1) × passage score (1/4) × query score (1/4) = 0.0625, or 6.25%. This is not a probability of retrieval. It is an internal diagnostic score showing that the page is technically reachable but editorially unfit for this query.
The editorial action is not “add more keywords.” Verify the actual commercial and legal position with the product and regional teams. If on-premises deployment is available, publish a dated, scoped section that states availability, customer requirements, German applicability and links to primary implementation documentation. If it is unavailable, say so plainly and create content for the deployment option that does apply. Retest the new passage against the same query set after publication.
Build a measurement workflow your team can maintain
- Select high-value pages and 10 to 30 representative queries; preserve the original wording and locale.
- Run technical and extraction checks before observing citations.
- Inventory the best candidate passage for every query, including its evidence status and freshness date.
- Score passage and query fit with a documented rubric, retaining “uncertain” where evidence is absent.
- Run answer-engine observations in controlled rounds, logging date, product surface, account state, query and cited URLs.
- Classify outcomes: access failure, extraction failure, passage gap, query mismatch, retrieved-but-not-cited where observable, or unresolved source preference.
- Change one meaningful variable where practical, then rerun the same test rather than crediting several simultaneous edits.
The unresolved category is legitimate. If a page passes the first four layers yet remains absent while comparable sources appear, source preference, corroboration, freshness, brand/entity association or unknown system behavior may be involved. That is when source-selection analysis becomes more useful than another technical audit.
FAQ and conclusion
Is retrieval eligibility the same as being indexed?
No. Indexing evidence can support discoverability, but it does not show that a system can extract the relevant passage or considers it a fit for a specific query. Different answer systems may also use different sources and retrieval paths.
Can a page be retrieval-eligible and still receive no citation?
Yes. Citation is a later selection and presentation decision. The answer may use another source, synthesize without visible citations, choose a more current document, or require corroboration your page does not provide.
What is the minimum useful reporting unit?
Use a query-page-passage record, not a domain-level score. Include test date, locale, access result, extraction result, evidence status, relevance scores, observed answer outcome and the next action.
Conclusion: Measure AI search retrieval eligibility before treating citation absence as a verdict on content quality or authority. Start with access, verify extraction, audit a claim-sized passage, then test its fit against the actual query and subqueries. This sequence will not expose proprietary retrieval systems, but it produces a defensible diagnosis: fix a barrier, strengthen a passage, broaden or narrow query coverage, or investigate genuine source preference. That is a far more useful basis for AI search visibility work than a citation count alone.
