The Honesty Audit: Which AI-Visibility Claims Are Actually Backed by Evidence

What it takes to audit a visibility claim

The problem is not that AI-visibility advice is all wrong. It is that a peer-reviewed result, a single unreplicated study, and an unsupported sales line are all sold in the same confident register, so a buyer has no way to tell one from another. A rigorous audit restores that distinction by putting the same three questions to any claim before believing it.

These are not academic niceties. They are the difference between spending a budget on something the evidence supports and spending it on a tactic that a primary source has already quietly retired.

• What primary source supports it? Not a blog citing a blog, but a study, a standards document, or a platform's own statement that can be read directly.
• What evidence tier is it? Established (replicated or standards-backed), emerging (a real but single, unreplicated finding), or contested (an industry narrative with no primary source behind it).
• What was it measured against? A number with no baseline, no control, and no place of publication is an assertion wearing the costume of a finding.

Backed by evidence: content changes can raise how often you are cited

Start with the claim that survives audit, because it sets the bar. The idea that you can influence generative-engine visibility through content is not vendor folklore; it is the founding, peer-reviewed result of the field. In the paper that coined generative engine optimization, Aggarwal and colleagues ran a controlled benchmark of roughly 10,000 queries across nine datasets and measured which content-level interventions changed whether a source was surfaced inside a generated answer.

The strongest levers were concrete: adding citations to credible sources, including direct quotations, and replacing vague statements with specific statistics. Together these produced a 30 to 40 percent relative lift on the study's visibility metric against an unoptimized baseline, and citing authoritative sources was consistently the single strongest move.

Read that verdict precisely, because precision is the point of an audit. This is a relative lift on one visibility measure, in the engines the researchers tested, versus content that had done nothing. It is not an accuracy figure, and it is not a promise that any single query will cite you. What it is, is real: published, controlled, and replicable, which is more than almost any other claim in this space can say.

Not backed by evidence: the llms.txt AI-visibility claim

Now the claim that collapses. A well-marketed tactic holds that placing an llms.txt file at your site root, a curated summary meant to hand AI systems a clean version of your content, improves how you appear in AI answers. The primary data does not support it.

Ahrefs analyzed 137,210 domains and found that 97 percent of valid llms.txt files received zero requests in the month studied. Of the small remainder that were fetched at all, 96 percent of fetches were generic bots, and only about a fifth of that residual traffic came from named AI tools. A second study, from SE Ranking across 300,000 domains, put adoption at roughly 10 percent, with no evidence of a visibility effect. Google's John Mueller has stated publicly that llms.txt is "not done for search," and Google's own guidance on optimizing for its AI features is explicit that no special file or markup is required to appear in them.

The verdict: llms.txt is not read at meaningful scale and is not treated as a ranking or citation signal by any major AI platform. It may still be a convenience for AI coding tools reading developer documentation, which is a narrow and different job. As a lever on whether a buyer's question names your business, the evidence is that it does nothing.

Half true, oversold: the structured-data accuracy jump

Some claims are harder because a real thing sits underneath an inflated number. Structured data is the clearest case. Schema.org is a genuine, standards-body vocabulary, founded jointly by Google, Bing, Yahoo, and Yandex in 2011 so that machine-readable markup means the same thing across every engine that consumes it. Marking up your pages has documented value: it lets engines parse your facts unambiguously and makes a page eligible for rich results.

The oversell is the statistic bolted on top. Claims of a specific "accuracy jump" or a percentage "lift in AI citations" credited to adding schema almost never arrive with a baseline, a control group, or a place of publication. The 30 to 40 percent figure that does exist in a controlled, peer-reviewed study is the one from the generative-engine-optimization paper above, and it is attributed to content-level changes, not to markup. Google, for its part, states plainly that no special markup is required to appear in its AI features, which run on the core index.

The audit's verdict is a split decision. Structured data is worth engineering for its parsing and eligibility value, and it is one of the few parts of the discovery stack you can inspect end to end through open validators. The quantified promise usually attached to it is a separate object, and that object generally fails the "what was it measured against" test.

Architecturally unsupported: getting cited proves the model read you

A subtler claim treats an AI citation as evidence that the model reasoned from your page. The retrieval-augmented-generation literature says that inference is unsafe. In the generate-then-retrieve and retrieve-then-generate pipelines that dominate production systems, a citation is attached to an answer largely independently of the evidence that actually produced it, so the source credited is not reliably the source the model used. Researchers argue that only inline, generation-time citation can guarantee faithfulness, and most deployed systems do not work that way.

The scale of the gap is documented. An attribution survey reports that more than 95 percent of answers from the open-source language models it tested contained at least one unattributed sentence. This is an established finding about the faithfulness problem; which mitigation actually closes it remains an open, emerging question.

The verdict: winning a citation is a real and useful signal, but it is not proof that your content shaped the answer, and it is certainly not a mechanism you can reverse-engineer with confidence. Any claim that treats "we got you cited" as a controllable, faithful outcome is overstating what the architecture currently delivers.

Unsupportable: the guaranteed AI ranking

This is the claim that fails hardest, and it fails on two counts at once. First, no primary source establishes that any intervention guarantees a citation or a fixed position inside a generated answer. Second, two documented properties of these systems make such a guarantee unsupportable rather than merely unproven.

The first property is non-determinism. Answer engines sample their output, so the same question, asked twice, can return a different set of named sources; a single check tells you almost nothing, and a guarantee about the next check would claim more certainty than the mechanism allows. The second is the decoupling and faithfulness gap above: rank and citation are not the same job, and the link between what a model reads and what it credits is unreliable. Stack these together and a promised, repeatable AI ranking is not a bold claim, it is a claim the evidence contradicts.

This is why measuring the AI-answer surface means reporting a presence rate with a confidence band, sampled many times per engine and stamped with the engine, locale, and date it was run against. That range reflects how these systems actually behave: sampled, non-deterministic, never fixed. A single guaranteed number would claim a certainty the field does not have.

Mislabeled: raise your E-E-A-T score

Some claims are not false so much as category errors, and "optimize your E-E-A-T score" is the common one. Experience, Expertise, Authoritativeness, and Trust are real concepts, but Google's own documentation is explicit about what they are: a framework that thousands of human raters use to evaluate the quality of the search algorithm's output, a feedback signal on the system rather than a score computed for any individual page.

Google has repeatedly clarified that E-E-A-T is not itself a direct ranking factor a page can optimize into a document. A large share of commercial GEO and AEO advice nonetheless treats a rater heuristic as if it were a machine-readable target with a dial you can turn. The underlying qualities genuinely matter for how content is judged over time; the "score" you are being sold the ability to raise does not exist as a thing an algorithm reads off your page.

Handle with care: AEO vs GEO as interchangeable terms

The last item is a terminology claim, and it earns a contested tier rather than a clean pass or fail. Industry accounts describe answer engine optimization as the older label, inherited from the featured-snippet and voice-search era of the mid-2010s, when the job was structuring content to be lifted as a direct, extractable answer. Generative engine optimization is the newer, academically coined term for being retrieved and cited by a large language model that synthesizes an answer.

In everyday usage the two labels have collapsed into near-synonyms. That is a terminology-hygiene problem, not a settled fact, and this narrative is an industry consensus rather than a peer-reviewed or standards-body claim. The more accurate handling is to keep the distinction the mechanisms deserve: AEO vs GEO share many tactics, but a pre-LLM featured snippet and an LLM-synthesized citation are produced by different machinery, and any content that flattens them into one thing should be read as convenient shorthand, not as established science.

A four-question test for any AI-visibility claim

The value of an audit is not the individual verdicts, which will age as the systems change. It is the method, which does not. The next confident claim you are handed, from us or from anyone, can be run through the same four questions before it earns your budget.

• Where is the primary source? If the trail ends at a blog citing another blog, treat the claim as marketing until a study, a standard, or a platform statement is produced.
• What is the evidence tier? Established, emerging, or contested are not the same and should never be priced the same.
• What was the number measured against? No baseline, no control, and no publication means no finding, however precise the figure looks.
• Does it survive how the systems actually work? A claim that ignores non-determinism, the ranking-citation split, or the faithfulness gap is describing a machine that does not exist.