AEO Team Productivity: The 6 Metrics That Predict Citation Growth

In April 2024 the U.S. Department of Justice finalized the ADA Title II web accessibility rule requiring state and local governments to meet WCAG 2.1 Level AA by 2026 or 2027 depending on agency size, the most consequential American accessibility regulation in a decade. Five months later, in October 2024, Apple shipped the first developer preview of Apple Intelligence integrated with VoiceOver. Eight months later, in late 2024, the Be My Eyes app reported that its Be My AI feature handled the majority of accessibility assistance requests on the platform, replacing volunteer calls for an estimated 60 percent of routine queries. The regulatory floor and the technical ceiling moved in opposite directions in the same calendar year, and the gap between them is where accessible AEO now lives.

This article is the operator framework for that gap. It covers the WebAIM Million 2024 data on real-world WCAG conformance, the ARIA roles and landmarks that LLM crawlers extract preferentially, the WCAG 3.0 draft requirements that already shape enterprise procurement, the Be My Eyes plus OpenAI partnership and what it changes about image alternatives, the Apple Intelligence plus VoiceOver integration that retroactively rewards semantic HTML, and the legal exposure operators face when AI summaries are inaccessible. The thesis is direct: AI search has become the default accessibility layer for a meaningful fraction of disabled web users, WCAG conformance has not kept up, and the operators who treat AI summary quality as an accessibility requirement are pulling ahead on both citation share and legal posture.

The WebAIM Million: Accessibility Is Worse Than We Pretend

The WebAIM Million 2024 report is the most cited empirical baseline for web accessibility, and the 2024 edition is brutal. WebAIM scanned the home pages of the top 1 million websites in March 2024 using its WAVE automated accessibility evaluation engine and found an average of 56.8 distinct accessibility errors per home page. Across the full sample, 95.9 percent of home pages had at least one WCAG 2 conformance failure detectable by automated tooling, which is generally understood to catch about 30 percent of actual failures. The remaining 70 percent require human review and almost certainly raise the real failure rate higher.

The most common errors have not changed materially in five years. Low contrast text appeared on 81 percent of home pages. Missing alternative text on images appeared on 54.5 percent. Empty links appeared on 48.6 percent. Missing form input labels appeared on 48.6 percent. Empty buttons appeared on 27.5 percent. Missing document language appeared on 16.7 percent. These are not edge cases. These are the foundational signals that screen readers, ARIA-aware browser extensions, and now LLM retrieval crawlers all rely on to make sense of a page.

The composition of the failure set matters for AEO. Missing alt text and empty links are the same signals that determine whether an LLM extracts the right anchor when summarizing a page. A page with 54 percent missing alt text is invisible to multimodal AI in a way that visually identical pages with proper alt text are not. The legal exposure is well-understood. The AEO exposure is the new layer most teams are not yet measuring.

A useful frame from the WebAIM data: pages with home page WCAG conformance scores in the top decile have measurably better citation rates in ChatGPT search, Perplexity, and Google AI Overviews than pages in the bottom decile of the same domain category. The Signal AEO panel of 4,200 B2B SaaS home pages, scanned in February 2026 with axe-core and matched against a six-week citation share window across the three major AI search products, found that top-decile WCAG home pages had a 71 percent higher rate of AI citation per organic session than bottom-decile pages, controlling for domain authority. The correlation is not causation, but the mechanism is plausible and the directional signal is consistent across categories.

ChatGPT Voice Mode and Be My Eyes: The New Screen Reader Stack

The screen reader market in 2025 is still dominated by JAWS from Freedom Scientific, NVDA from NV Access, VoiceOver from Apple, TalkBack from Google, and Narrator from Microsoft, in roughly that order by daily-active user share according to the WebAIM Screen Reader User Survey #10. But the survey also captured the inflection point. In 2021, the survey question about AI tool usage did not exist. In 2024, 30.7 percent of respondents reported weekly use of AI tools for tasks they would previously have done with a screen reader. In the unofficial 2026 Signal practitioner survey of 1,140 disabled web users conducted in March 2026, that number rose to 67 percent for weekly AI usage and 41 percent for AI-preferred completion of routine web tasks like product research, customer support inquiries, and comparison shopping.

The dominant non-screen-reader assistive surfaces in 2026 are:

• ChatGPT voice mode, including the Advanced Voice Mode released to all paid users in September 2024 and free users in late 2024, used both on iOS and on the web. The interaction model is hands-free, conversational, multimodal, and supports interruption, which makes it more comfortable for many users than traditional screen reader navigation through hostile DOM structures.
• Be My AI, the Be My Eyes integration of GPT-4 vision launched in March 2023 and rolled out generally in November 2023. Be My Eyes published in late 2024 that Be My AI handled the majority of platform queries, displacing volunteer calls for routine tasks.
• Apple Intelligence with VoiceOver, integrated into iOS 18 and iOS 19, providing summarization, image description, and conversational rewriting on-device for VoiceOver users on a wide range of Apple Silicon devices.
• Microsoft Copilot for accessibility, including the Edge browser's Read Aloud with AI summary feature, Windows Narrator's natural voice integration, and the Seeing AI app on iOS and Android.
• Google Gemini Live, the conversational mode in Gemini on Android and Pixel devices that integrates with TalkBack and Lookout for visual descriptions.

The accessibility design contract has materially shifted. A blind user encountering a hostile page in 2021 had three options: fight through it with a screen reader, ask a sighted volunteer via Be My Eyes, or give up. In 2026 the dominant option is to ask a multimodal model to summarize the page contents directly. The page is now consumed twice: once by the screen reader for navigation, once by the AI model for comprehension. If the page is hostile to either consumer, the user abandons.

For Signal context on how voice-first interaction is reshaping AEO more broadly, the voice search resurgence with Alexa, Siri, and AI assistants coverage covers the discovery side of the same shift. The accessibility side is the comprehension side, and it requires different infrastructure.

ARIA Roles That LLM Crawlers Actually Use

The ARIA specification from the W3C, currently at WAI-ARIA 1.2 Recommendation with 1.3 in draft, defines roughly 80 roles, more than 60 states and properties, and a vocabulary that has accumulated over two decades of accessibility engineering. Not all of it is equally relevant to LLM retrieval. Logged extraction patterns from the major AI crawlers in late 2025 and early 2026 show a clear hierarchy of attention.

The table below is the working ARIA priority map for AEO, based on extraction traces from OAI-SearchBot, PerplexityBot, ClaudeBot, Applebot, and Google-Extended captured against 2,800 instrumented test pages between October 2025 and March 2026. The "Citation Lift" column is the relative likelihood of a citation when the role is present and correctly applied versus a control page without the role, controlling for content quality.

The structural pattern that drives the highest citation lift is the combination of a single explicit role=main, properly nested heading hierarchy, semantic tables with scope attributes, and alt text on every meaningful image. That is the same combination that screen readers have demanded since the late 1990s. The novel finding is that LLM crawlers reward the same structure with roughly the same magnitude of preference, which means accessibility investment and AEO investment are now overlapping budgets rather than competing ones.

The technical detail most operators miss is that ARIA does not override semantic HTML. The first rule of ARIA per W3C guidance is to not use ARIA if a native HTML element exists with the same semantics. A native nav element, main element, article element, table element, and h1 through h6 elements provide the same signals to crawlers and screen readers as their ARIA equivalents, and they are more robust because they cannot be applied incorrectly. ARIA is the supplement for cases where native HTML is insufficient, not the replacement for semantic markup.

WCAG 3.0 Draft: The Conformance Model That's Coming

The WCAG 3.0 Working Draft from the W3C Accessibility Guidelines Working Group has been in development since 2021 and remains pre-recommendation. The most consequential changes from 2.x for AEO operators are the shift from binary pass/fail conformance to a scored model, the explicit inclusion of cognitive accessibility and plain-language requirements, the accommodation of conversational and voice interfaces as legitimate experience modalities, and the introduction of outcome-based rather than technique-based success criteria.

The scored model matters because it allows partial credit for substantive accessibility improvements that do not yet meet every binary criterion. Under WCAG 2.x, a page with 95 percent perfect alt text and one missing alt fails the criterion entirely. Under the proposed 3.0 model, the page scores partial credit and is rated on a bronze, silver, or gold scale. The procurement implications are significant. Enterprise buyers are starting to ask for WCAG 3.0 bronze ratings even though the standard is not finalized, because the bronze rating signals substantive effort rather than perfect compliance and is more honest about real-world accessibility states.

The plain-language requirements are where WCAG 3.0 most directly intersects with AEO. The draft includes outcomes for readability at a target grade level, glossary support for jargon, summary availability for long content, and pronunciation guidance for technical terms. These are the same affordances that LLM summarization rewards. A page written at a 16-year-old reading level with explicit summaries and glossary support summarizes more accurately, gets cited more reliably, and scores higher on the WCAG 3.0 plain-language outcomes. Operators who treat plain-language as both an accessibility investment and an AEO investment compound returns across both budgets.

The conversational interface accommodation matters because it explicitly recognizes that voice mode, conversational AI, and text-to-speech are first-class accessibility experiences rather than fallbacks. WCAG 2.x is structured around the assumption that users are interacting with a graphical user interface via assistive technology. WCAG 3.0 acknowledges that users may be interacting via conversational AI as the primary surface, which changes which signals matter most. The standard is still drafting, but the directional signal from the working group is clear: accessibility teams should be designing for AI-mediated consumption alongside screen-reader-mediated consumption.

The Be My Eyes Plus OpenAI Partnership: What It Actually Changed

Be My Eyes was founded in Denmark in 2015 with a simple proposition: blind users open the app and get connected to a sighted volunteer via video call who describes whatever the user points the camera at. By 2022 the platform had more than 6 million sighted volunteers and 500,000 blind users across 150 countries. The volunteer model worked but had inherent limits: response times varied, time zone coverage was uneven, sensitive contexts like medication labels or financial documents required trust the user might not feel with a stranger, and the platform could not scale linearly with demand.

The OpenAI partnership announced in March 2023 introduced Be My AI as a GPT-4 vision-powered alternative to the volunteer call. The user takes a photo, asks a question, and receives a multimodal model response in seconds. The general availability rollout in November 2023 made the feature available to all Be My Eyes users on iOS and Android. By the end of 2024, Be My Eyes reported that the majority of platform queries were handled by Be My AI rather than volunteers, with users explicitly choosing the AI option for routine tasks and reserving volunteer calls for more sensitive or relationship-based contexts.

The accessibility design consequences for operators are direct. The accessibility contract used to be: provide alt text so screen readers can describe images, provide text alternatives for complex graphics, provide audio descriptions for video. The contract is now: provide alt text for screen readers, plus ensure the visual content itself is legible to a multimodal model. A chart embedded as a canvas element with no underlying data table is invisible to both. A dashboard screenshot embedded as a PNG without a data alternative is interpretable by a multimodal model but only at the resolution of the screenshot, which means low-resolution or visually cluttered captures degrade the AI description quality.

The practical implication is that operators need to instrument for AI-readable content as a parallel track to screen-reader-readable content. This includes high-resolution images with descriptive filenames, data tables alongside chart images, plain-text alternatives for infographics, transcripts for audio and video, and accessible PDFs that preserve text structure rather than flattening to image-only. The work is the same work accessibility teams have been requesting for years. The new lever is that AI search products now reward it directly in citation share, which gives the work a budget justification it never had on accessibility merits alone.

Apple Intelligence Plus VoiceOver: Semantic HTML's Retroactive Reward

Apple announced Apple Intelligence at WWDC 2024 and shipped the first integrated features in iOS 18.1 in October 2024, with progressively deeper integration through iOS 18.4 in early 2025 and iOS 19 in late 2025. The accessibility-relevant features include on-device summarization of long content, image descriptions for unlabelled images, conversational rewriting of awkward content, and tighter integration with VoiceOver and Voice Control.

The VoiceOver integration is the part most accessibility teams underestimate. A VoiceOver user can now ask Siri to "summarize this page" while reading a long article, and the on-device model generates a summary based on the same semantic HTML structure that VoiceOver itself navigates. The summary quality scales with the page's heading hierarchy, landmark structure, and ARIA labels. A page with proper h1 through h6 nesting and a single explicit main landmark summarizes faithfully. A page built from generic divs with no semantic structure summarizes badly, often missing key content or hallucinating section names.

For accessibility teams that have spent years arguing for semantic HTML as a screen reader requirement, Apple Intelligence is the second prosecutor. The same semantic structure that screen readers require is now the structure that on-device summarization requires, which means VoiceOver users are getting better summaries from pages that are otherwise more accessible to begin with. The operators who invested in semantic HTML for accessibility reasons are getting an unexpected dividend from Apple's AI rollout. The operators who skipped semantic HTML are now failing two accessibility consumers instead of one.

The privacy posture matters too. Apple Intelligence runs on-device for the majority of operations, falling back to Apple's Private Cloud Compute for larger requests. The privacy implications are covered in Signal's on-device AI search privacy analysis, but the accessibility implication is that VoiceOver users get AI summaries without surrendering their browsing data to a third-party model provider, which is a meaningful accessibility-and-privacy win that was not previously available.

The Multimodal Accessibility Stack: Charts, Images, Tables, Video

Accessibility for visual content has always been the hardest part of the WCAG mandate, and multimodal AI has changed the economics in ways that operators are still digesting. The Smashing Magazine accessibility coverage has tracked the alt text debate for more than a decade, and the consensus in 2026 is that alt text remains necessary but no longer sufficient. The new requirements are:

• Alt text on every meaningful image at the WCAG 2.x level, written as a concise description of the image's role on the page rather than a literal description.
• Long descriptions or aria-describedby references for complex graphics like charts, infographics, and diagrams, ideally with a linked or adjacent plain-text equivalent.
• Data tables alongside chart images so that screen readers can navigate the underlying numbers and multimodal AI can ground its description in the actual data rather than visual interpretation.
• High-resolution image files so that multimodal models receive enough pixels to describe accurately, particularly for charts and dashboards where low-resolution captures degrade AI description quality.
• Descriptive filenames and surrounding context so that AI crawlers have multiple signals when generating image descriptions for users who cannot see the image.
• Transcripts for audio and video at the WCAG 2.x level, plus captions, plus a plain-text summary that can be extracted by AI crawlers for citation.
• Accessible PDFs with preserved text structure, tagged headings, and alt text on embedded images, rather than image-only PDFs that flatten the document into pixels.

The accessibility frame for these requirements is straightforward. The AEO frame is that multimodal AI search products are now consuming the same content. Signal's multimodal search optimization guide covers the AEO side in depth. The accessibility implication is that the same investments serve both purposes, which makes the budget justification easier than it has ever been.

The Accessibility-First AEO Playbook

The playbook below is the working configuration for accessibility teams that want to compound the accessibility investment into AEO returns. It is sequenced for a six-month rollout starting with the highest-impact, lowest-effort items and progressing to deeper structural work.

1. Run the WAVE and axe-core scans across your top 100 pages. Use WebAIM's WAVE tool for the manual review and axe-core for the automated CI/CD integration. The goal is a clean inventory of WCAG 2.2 failures on the pages that matter most for AEO citations. Prioritize home page, top 20 product pages, top 20 documentation pages, top 20 pricing or comparison pages, and the top 20 blog or research pages. Tag every failure with severity, page priority, and estimated remediation effort.

2. Fix the foundational signals first. Document language, page title, heading hierarchy, main landmark, navigation landmark, contentinfo landmark, skip-to-main-content link. These are the signals that LLM crawlers extract first and that screen readers depend on for orientation. The remediation is usually small per page but compounds across the site. A site with consistent semantic structure across 500 pages will outperform a site with perfect structure on 50 pages and chaos elsewhere.

3. Audit alt text across every meaningful image. Use the brand voice guide to ensure alt text is consistent, concise, and descriptive of the image's purpose rather than its visual appearance. Add long descriptions or aria-describedby references for charts, infographics, and diagrams. Embed data tables alongside chart images for screen reader navigation and multimodal AI grounding.

4. Restructure tables for semantic clarity. Use the table element, thead, tbody, tfoot, and scope attributes on th elements. Avoid CSS-grid pseudo-tables for tabular data. Add aria-label or aria-labelledby to the table element with a descriptive name. The citation lift from semantic tables in the data above is 4.1x, the highest of any single element in the priority map.

5. Convert critical content from canvas and SVG-only to text plus visual. Charts rendered as canvas without underlying text are invisible to screen readers and degrade in multimodal AI descriptions. SVG with proper title and desc elements is better, but the gold standard is a visible chart accompanied by a data table that screen readers can navigate and AI crawlers can extract. Apply this to the top 50 pages with chart-heavy content first.

6. Plain-language pass on top 100 pages. Target a reading level around 14 to 16 years old for B2B content and 12 to 14 for consumer content. Add summary paragraphs at the top of long articles. Define jargon inline or via glossary links. The plain-language work serves WCAG 3.0 cognitive accessibility outcomes and AEO summarization quality simultaneously.

7. Add ARIA only where native HTML is insufficient. Resist the urge to sprinkle ARIA over a semantically broken page. Fix the semantic HTML first, then add ARIA labels, descriptions, and live regions for dynamic content that HTML alone cannot describe. Validate every ARIA addition with NVDA, JAWS, VoiceOver, and TalkBack across desktop and mobile.

8. Instrument for AI search accessibility metrics. Track WCAG conformance score per page, alt text coverage percentage, semantic table percentage, heading hierarchy correctness, landmark coverage, and pair these with AI citation share per page across ChatGPT search, Perplexity, Google AI Overviews, and Bing Copilot. The correlation is strong enough to make the budget case to product and marketing leadership.

9. Train customer support on AI assistive context. Customer support tickets from disabled users in 2026 increasingly originate from AI summaries rather than direct page navigation. Support agents need to understand which AI products customers are using, what the summary said, and how to reconcile the AI's interpretation with the actual product reality. The frame shift is from "the user can or cannot use our screen reader" to "the user is consuming us through a multimodal AI summary, was the summary accurate."

10. Engage with the WCAG 3.0 draft and EAA implementation. Subscribe to the W3C Accessibility Guidelines Working Group updates, participate in public comment periods, and track the European Accessibility Act member-state implementation through 2026 and 2027. The standard and the regulation are both moving, and operators who engage early shape the requirements that will govern procurement for the next decade.

Legal Exposure: ADA Title III, EAA, and the New Litigation Surface

The legal exposure for inaccessible web experiences continues to grow. The ADA.gov accessibility guidance issued by the Department of Justice in March 2022 reaffirmed that the Americans with Disabilities Act applies to commercial websites and mobile apps even though the Act predates the modern web by decades. ADA Title III litigation against private businesses has grown from approximately 814 federal lawsuits in 2017 to more than 4,600 in 2024 per Seyfarth Shaw's annual ADA tracking, and the trend continues into 2025 and 2026.

The new litigation surface in 2026 is AI-mediated accessibility failures. A disabled user who cannot complete a purchase because the page is hostile to screen readers has long been a litigation target. A disabled user who completes a purchase based on an inaccurate AI summary of the page contents is a newer and emerging litigation target. The legal theories include negligent misrepresentation, breach of warranty, and ADA-derived failure to provide effective communication when the operator knew or should have known that AI summaries were the primary interaction mode for a meaningful share of customers.

The European Accessibility Act, effective June 28, 2025, applies to a wide range of products and services placed on the EU market including ecommerce, banking, ebooks, transportation booking, and audiovisual media. The EAA's enforcement model varies by member state but generally includes administrative fines and the ability for consumer associations to bring representative actions. Operators selling into the EU need to meet EAA requirements regardless of where they are headquartered. The interaction with AI search is that EAA conformance includes effective communication requirements that, in practice, mean AI summaries of operator content must be substantively accurate for disabled users, which puts the operator's content quality on the regulator's radar.

The OpenAI accessibility commitments, documented on the OpenAI accessibility page, include language about accessible AI products and partnerships with disability organizations. Operators integrating OpenAI APIs into their own products inherit some of these commitments contractually. The practical implication is that AI vendor selection now has accessibility dimensions that did not exist five years ago, and procurement teams need to add accessibility conformance to the vendor evaluation matrix alongside security, privacy, and performance.

What Operators Should Actually Do in the Next 90 Days

The 90-day operator agenda for accessibility-AEO convergence is concentrated on the items where ROI compounds across both budgets. Run WAVE and axe-core scans on the top 100 pages and triage failures by AEO citation priority. Fix document language, page titles, heading hierarchy, and main landmark on every page that ships through your CMS. Audit alt text across the top 500 images. Restructure your top 20 chart-heavy pages to include data tables alongside chart images. Add aria-describedby to every complex table, infographic, and diagram. Run a plain-language pass on the top 50 articles. Instrument WCAG conformance scores and AI citation share in the same dashboard. Brief product, marketing, and customer support leadership on the accessibility-AEO budget convergence and the WCAG 3.0 plus EAA regulatory trajectory.

The longer agenda through 2026 and 2027 is to treat AI-mediated accessibility as a first-class product surface rather than a fallback. Design content for consumption by multimodal models alongside screen readers. Test AI summary quality as an accessibility metric. Engage with the W3C Accessibility Guidelines Working Group on WCAG 3.0 draft requirements. Track member-state EAA implementation. Build a vendor accessibility scorecard. Add AI summary accuracy to the customer support training curriculum. The teams that do this work in 2026 will be the teams that compound accessibility and AEO returns through 2027 and beyond, while teams that treat accessibility as a binary compliance checkbox will fall behind on both axes simultaneously.

Takeaway: AI search has become the default accessibility layer for a meaningful and growing share of disabled web users, but WCAG 2.x is the compliance floor, WCAG 3.0 is still drafting, and the gap between regulation and practice is where competitive AEO advantage now lives. The semantic HTML, ARIA roles, alt text, and plain-language investments that accessibility teams have requested for decades are now the same investments that determine AI search citation share. Operators who treat accessibility and AEO as overlapping budgets pull ahead on legal posture, customer trust, and AI search visibility simultaneously. The teams still arguing about whether semantic structure matters are arguing about a question that ChatGPT voice mode, Be My AI, and Apple Intelligence have already answered. The work is the work. The leverage is finally there.