AI search reporting often treats a displayed citation as a successful outcome. That is too generous. A source can appear beneath an answer while sending the searcher to a dead page, a subscription gate, an account prompt, a region-specific denial or an unrendered shell with no usable evidence.
That distinction matters for users and for measurement. Citation presence tells you that an answer engine selected a URL. Citation accessibility tells you whether a typical searcher—and, separately, a retrieval system—can reach and inspect the cited evidence. Those are related conditions, not the same condition.
This framework for AI search citation accessibility measurement makes that separation explicit. It gives SEO practitioners and content teams a repeatable way to audit cited sources, quantify accessible evidence share and avoid reporting a citation as equivalent to an evidence experience.
Why citation presence is not evidence accessibility
AI interfaces can cite URLs for reasons that do not guarantee a good click-through experience. The engine may have retrieved an earlier version, used a licensed index, accessed a cached representation, or selected a page whose state has since changed. A citation can therefore remain visible after the destination becomes unavailable.
Accessibility also has an audience dimension. A public-health page that works in one country but blocks visitors elsewhere is not universally accessible. A trade publication may be open to a crawler but gated for a reader. Conversely, a page may be readable in a browser but block the bot or JavaScript pathway needed for retrieval.
Do not collapse these cases into a single binary “working” label. The useful question is: accessible to whom, from where, through which path, and with how much of the cited claim visible?
This complements AI search citation share-of-voice measurement. Share of voice establishes who is cited; accessibility qualifies the practical value of those citations.
Define two accessibility layers before testing
Keep reader access and retrieval access in separate fields. Combining them obscures problems and can lead to bad recommendations, such as opening a page to bots without considering the reader’s post-click experience.
| Layer | Core test | Common failures |
|---|---|---|
| Reader accessibility | Can a normal, logged-out person open the cited page and locate the supporting content? | 404s, soft 404s, paywalls, login walls, consent loops, geographic blocks, app-only pages |
| Retrieval accessibility | Can an allowed answer-engine or search crawler request and interpret the supporting content? | robots restrictions, WAF challenges, bot-specific denial, blocked assets, client-side-only rendering |
| Evidence accessibility | Can the person or system see the specific claim, caveat or data point attributed by the answer? | Teaser-only copy, broken anchors, collapsed tables, missing context, inaccessible PDFs |
Reader accessibility is the default metric for a citation shown to a user. Retrieval accessibility is diagnostic: it helps explain citation volatility and informs technical remediation. Neither proves what a proprietary model saw internally, so record observed behavior rather than inferring model capabilities.
Use a status taxonomy that preserves the cause
Audit the final destination after following redirects, then assign one primary status and optional secondary notes. Primary status should describe the highest-impact observed barrier.
- A0: Fully accessible. The page loads without authentication or payment, and the cited supporting material is visible.
- A1: Accessible with friction. Content is available after ordinary consent or a non-blocking interstitial, but the evidence remains reachable.
- A2: Partially accessible. Some relevant text is visible, but material evidence is truncated, obscured or unavailable.
- A3: Restricted. Login, subscription, regional policy, age gate, app requirement or hard consent wall prevents meaningful access.
- A4: Technically inaccessible. Dead link, server error, redirect loop, challenge failure, malformed destination or unusable render.
- A5: Crawler-restricted. The reader can access it, but documented bot controls or observed bot denial block relevant retrieval paths.
A page may be A0 for readers and A5 for a particular crawler. Keep both labels. Google notes that eligibility and treatment in AI features can vary, and its AI features guidance is a useful reference for understanding the normal controls that affect search appearance. For Perplexity-specific bot controls, consult its published bot guidance rather than guessing from a generic user agent test.
Build the audit dataset around citations and claims
Start with a stable prompt set, engine, market, device context, collection date and full answer capture. Normalize citation URLs before testing; tracking parameters, mobile variants and redirecting short links otherwise distort source counts. The workflow in cleaning AI search citation data is a sensible prerequisite.
For every citation instance, capture the prompt ID, answer snapshot, cited URL, final URL, publisher, cited claim, citation position, access test timestamp, tester country, browser state, reader status, retrieval status and notes. Screenshot or archive the result where policy permits. Accessibility is temporal; a URL checked today may not behave the same next week.
Test in a clean, logged-out browser first. Record HTTP status and redirect chain, but do not stop there. A 200 response can still be a paywall, a bot challenge or a JavaScript shell. Check whether the cited passage, table, figure caption or qualifier is visible in the rendered page.
Test the difficult cases consistently
Dead links and soft 404s: Mark a hard 404, 410, persistent 5xx response or domain-resolution failure as A4. For soft 404s, assess the rendered page: a generic category page or “content unavailable” template does not preserve evidence merely because it returns 200.
Redirects: Follow the complete chain. A relevant permanent redirect can remain A0. A redirect to a homepage, unrelated product page, login form or geo-selector is partial or restricted because the originally cited evidence is no longer directly available.
Paywalls and login walls: Distinguish preview text from claim-level proof. A headline and first paragraph may support a broad statement but not a numeric assertion or a caveat appearing below the meter. Label the specific claim A2 or A3, not simply the domain.
Regional restrictions: Test from the priority market stated in your study design. If your brand serves multiple markets, sample each rather than treating one location as universal. Document VPN or proxy use because it can trigger different defenses from ordinary user access.
JavaScript rendering: Compare raw HTML with a standard rendered browser when feasible. If meaningful content appears only after scripts execute, it may still be reader-accessible. It becomes a retrieval concern only when you have documented evidence that relevant crawlers cannot render it or observed access failures under an approved test method.
Crawler blocks: Do not attempt to bypass controls. Check published crawler policies, robots directives and ordinary server responses. A block can be deliberate and appropriate; measurement should report it, not imply wrongdoing.
Calculate accessibility-adjusted citation share of voice
Use a claim- and query-weighted calculation so one low-value citation does not count the same as a prominent source supporting a central answer claim. First calculate each source contribution for a query:
Contribution = inclusion × prominence × claim coverage × query weight × accessibility factor
Use these fixed scales: inclusion is 0 or 1; prominence is 1.0 for the first or visually primary citation, 0.7 for a secondary citation and 0.4 for a tertiary or grouped citation; claim coverage is 0, 0.5 or 1.0, representing no support, partial support or full support for the evaluated claim; query weight ranges from 1 to 5, where 1 is low-priority and 5 is highest-priority; accessibility factor is 1.0 for A0, 0.8 for A1, 0.5 for A2, 0.1 for A3, and 0 for A4 or A5 when measuring reader-accessible evidence. Report crawler accessibility separately rather than using it to overwrite the reader factor.
Worked example: For a priority-4 query, Source A is cited once in the primary position, fully covers the claim and is partially accessible. Its contribution is 1 × 1.0 × 1.0 × 4 × 0.5 = 2.0. Source B is a secondary citation, fully supports the claim and is fully accessible: 1 × 0.7 × 1.0 × 4 × 1.0 = 2.8. Source C is tertiary, partially supports the claim and is restricted: 1 × 0.4 × 0.5 × 4 × 0.1 = 0.08.
The unadjusted opportunity denominator for that answer is the same calculation with accessibility fixed at 1: Source A = 4.0, Source B = 2.8 and Source C = 0.8, for a denominator of 7.6. Source A’s accessibility-adjusted SOV is 2.0 ÷ 7.6 × 100 = 26.3%. Across a prompt set, sum a publisher’s adjusted contributions and divide by the summed unadjusted opportunity denominator. Keep the denominator constant when comparing accessibility-adjusted SOV with conventional SOV.
This is deliberately a measurement convention, not an industry rule. The scales are useful because they are explicit and repeatable. Calibrate prominence rules to each interface, document changes and do not compare results across studies that use different definitions.
Report three metrics, not one
- Citation presence rate: the percentage of evaluated answers that cite the source.
- Reader-accessible citation rate: the share of citation instances classified A0 or A1, optionally reported with A2 separately.
- Accessibility-adjusted SOV: weighted source contribution after applying the accessibility factor.
Add a barrier breakdown: dead, redirected away, paywalled, login-gated, region-restricted, render-failed and crawler-restricted. This makes remediation actionable. A high A3 rate suggests editorial or distribution constraints; A4 points to link maintenance; A5 requires a policy and technical review, not a blanket instruction to remove access controls.
For sensitive sectors, assess the accessibility of caveats as well as the headline claim. A citation that exposes a reassuring conclusion but hides eligibility limits can create a misleading evidence path. Pair this work with AI search summary fidelity measurement to test whether the answer retained the source’s limits.
FAQ and conclusion
Is a paywalled source an inaccessible AI citation?
Not automatically. Classify access at the cited-claim level. If the visible preview proves the claim, it may be partially accessible. If the support sits behind payment or login, classify it as restricted and explain the barrier.
Should crawler blocks count as dead links?
No. A crawler block and a dead URL are different conditions with different remedies. Report reader access, crawler access and evidence visibility separately. Do not test by evading controls.
How often should teams audit?
Run a baseline before major reporting, then recheck high-impact prompts on a regular cadence and after migrations, paywall changes or notable answer-engine shifts. Time-stamp every observation.
What is the practical conclusion?
A citation is an attribution event, not proof that evidence is usable. Measure presence first, then test the destination, the cited claim and the relevant user context. Accessibility-adjusted SOV will usually be a more honest view of source value than raw citation counts, while the barrier taxonomy gives content, product and technical teams a clear list of issues they can actually investigate.
