How to read an 'accessibility improves rankings' claim

A claim circulates in optimization material that making a website more accessible also makes it more visible — higher in search results, and more often cited by AI answer engines. It is not one claim but three, and each fails a different test, so the first task is to sort a given claim into the right class. None of the three, as usually stated, supports a specific number such as “accessible pages are cited some percentage more.” The full audit is the companion whitepaper on how search optimization, generative engine optimization, and accessibility relate; this note is the shorter tool for a single claim.

Direct-factor claims: the engine’s own documentation

A direct-factor claim asserts that an accessibility property is read as a ranking input; the test is whether the engine’s own documentation says so. The documentation answers property by property and surface by surface, not for accessibility as a whole — and for the clearest page-ranking cases it says the opposite. In a March 25, 2022 Search Central office-hours session, a Google spokesperson, asked about accessibility-related link presentation, said accessibility is important for a site’s users but “not something that we would pick up and use as a direct ranking factor when it comes to search,” adding, “Maybe that will change over time” [1]. Two qualifications attach. Google published no official transcript of the session — its archive begins later in 2022 — so the wording is a contemporaneous trade-press transcription of the public recording [2]; and the “difficult to quantify” phrase often paired with this exchange is the outlet’s framing, not the spokesperson’s words. A statement by one engine on one date is also not general, and must be checked against that engine’s current documentation.

Google’s SEO starter guide narrows the target further: headings in semantic order are “fantastic for screen readers, but from Google Search perspective, it doesn’t matter if you’re using them out of order” [3]. Read exactly, that decouples heading order from ranking; it does not say headings are worthless — the same documentation treats them as useful for title-link generation and user navigation. A claim resting on heading structure must survive this precise statement, not a paraphrase inflating it into “headings help rankings” or deflating it into “headings don’t matter.”

One documented case cuts the other way, and an honest sorting must include it: Google’s image documentation states that “Google uses alt text along with computer vision algorithms and the contents of the page to understand the subject matter of the image” [7]. Alt text is an accessibility property, and here it is a genuine documented input — for understanding images in image search, not a page-ranking factor. The property and the surface are what changed, not the engine’s candor. A direct-factor claim must therefore name which accessibility property and which search surface it means; “accessibility” as an undifferentiated whole is not something the documentation either endorses or rejects.

Correlational claims: n, window, controls, and direction

A correlational claim reports that accessible sites rank or are cited more often, without saying why. Four questions decide whether the association means anything: how many observations (n), over what window, with what controls, and in which causal direction. The population makes the last two acute. In the 2026 WebAIM Million — an automated scan of 1,000,000 home pages in February 2026 — 95.9% of pages had at least one detected WCAG 2 failure [4]. That is not “95.9% of pages are inaccessible”: it counts pages with at least one automatically detectable failure, and automated tools catch only a subset of conformance issues. Fully accessible pages are therefore a small minority, so a comparison drawn from them pits unusual sites against typical ones — and sites that invest in accessibility also tend to be better resourced and more actively maintained, which itself predicts ranking. A raw association cannot separate an accessibility effect from a site-quality effect without controls.

The same census cautions about direction: pages using ARIA carried more detected errors on average (59.1) than pages without it (42), and the report states this “does not necessarily mean that ARIA introduced these errors” — those pages were also more complex [4]. The correlation points the wrong way because a confound, page complexity, drives both variables; an accessibility-and-visibility correlation inherits the same hazard. The arithmetic for turning an observed rate into an interval, so a small-sample figure is not read as a precise one, is in the companion note how to read a visibility percentage.

Mechanism claims: plausible but unmeasured

A mechanism claim proposes a pathway rather than a correlation: accessible, well-structured markup is parsed more faithfully, chunked more cleanly, retrieved more reliably, and therefore cited more often. The pathway is plausible, and parts of it are documented in general retrieval tooling. But two load-bearing links are, in the record sampled here, verified documentary absences rather than under-cited work. No public source in that record documents a distinct parse-quality stage — how faithfully a crawler’s extraction preserves page structure — as separate from chunking, for any deployed answer engine. And no public study in that record measures the effect of retrieval quality on answer inclusion for a deployed generative engine: the one controlled study of answer inclusion, the KDD 2024 paper that introduced generative engine optimization (n = 10,000 queries against a simulated engine), supplies its sources to the generator directly and never varies retrieval quality as a manipulated variable [6].

The engines’ own documentation stops at the endpoints. On eligibility for its AI Overviews and AI Mode, Google says a page must be indexed and eligible for a snippet, then adds: “There are no additional technical requirements” [5] — the input, and nothing about the internal parse, chunk, and retrieval stages the mechanism depends on. Such a claim is an inference across a gap, not a measured chain; it may still be true, but it remains a hypothesis. None of the three classes examined here licenses a specific quantified expectation, and a claim that attaches one asserts more than its evidence class can carry.

Limitations

This note sorts claims; it does not settle the underlying question. The documentation cited describes specific engines on the dates read — the Google pages were last updated 2025-12-10 (the image documentation, 2026-03-02) and read on 2026-07-09 — and should be assumed perishable: an engine can change how it treats a signal without notice, so “not a direct ranking factor” is a dated finding, not a permanent property. The census figures come from an automated tool on home pages only and undercount true failures; they bound the population’s measured accessibility, not any individual site’s. And none of the three questions, answered well, shows that accessibility has no effect on visibility — only that a given claim has not yet shown one.

Checklist: what a claim must disclose

1. Which specific engine and version, and on what date — since documentation and rankings change without notice?
2. Which class is it — direct-factor, correlational, or mechanism — and does the evidence match?
3. For a correlational claim: n, the window, the controls, and could a confound such as site resources or page complexity produce the same association?
4. Which way does the causal arrow run, and has the reverse or a common cause been ruled out?
5. Is the estimate an interval rather than a single point, following the arithmetic in the companion visibility-percentage note?
6. For an AI-answer-citation claim: is the query construction disclosed — one fixed phrasing, or a distribution of phrasings for the same need? For an organic-ranking claim: is the keyword sample disclosed, along with the rank tracker’s geography and personalization settings?
7. For a mechanism claim, is every link in the pathway documented, or is one assumed?

These questions restate, for a single claim, the minimum disclosure requirements this publication applies to any visibility measurement. A claim that answers them is evaluable, even if the effect turns out small or absent; a claim that answers none is a slogan, not a measurement.

@techreport{BA-MN-5,
  author       = {{Barkhausen AI}},
  title        = {How to read an 'accessibility improves rankings' claim},
  institution  = {Barkhausen AI},
  year         = {2026},
  url          = {https://barkhausen.ai/notes/how-to-read-an-accessibility-ranking-claim/}
}