LEARN

Widely Repeated GEO Claims That Lack Evidence

Branching conduits end at translucent barriers; only those with glowing nodes reach the central chamber, symbolizing filtering information based on evidence
In fields with undisclosed mechanisms, whether a practice is worth doing depends on publicly verifiable evidence, not popularity

IN ONE SENTENCE

In a field where the mechanism is not public, whether a practice is worth doing depends on checkable evidence — not on how many people repeat it.

Because the mechanisms are not public, a lot of claims travel on repetition rather than evidence. Checking them against public counter-evidence removes a fair amount of wasted effort.

OUR POSITION

When you meet a claim about a proprietary method, ask three questions: what is the source, how large was the sample, and has the platform said the opposite. A claim that fails all three is not worth funding.

01

Disproven by data

'Publish llms.txt and large models will index you': Ahrefs' study of 137,000 domains found 97% of those files received zero requests in May 2026, and Google has repeatedly said it does not use them and has no plans to. Treat it as cheap agent-readiness infrastructure, not a visibility lever.

'AI traffic converts several times better than organic': studies agree on direction and range from a few times to more than ten, and at least one statistical analysis found no significant overall difference. With that spread, no specific multiple belongs in an argument.

02

Named as unnecessary by the platform

Google's May 2026 guidance lists several practices as unnecessary: llms.txt, content chunking, AI-specific rewrites, structured data for AI search, and inauthentic mention-building.

⚠️ That guidance covers only Google's own AI features and does not apply to ChatGPT, Claude or Perplexity. Treating it as an industry-wide rule is its own misreading.

03

Where the time should go instead

Three actions have evidence: adding statistics, adding cited sources, adding quotations — a 30–40% relative lift across 10,000 queries in the Princeton paper.

Plus two pieces of groundwork that are almost never wrong: confirm retrieval crawlers receive complete HTML, and confirm structured data appears in what the server returns.

Data behind this page

97%

Share of published llms.txt files with zero requests in May 2026

Source：Ahrefs, across 137,000 domains，2026

30–40%

Relative visibility lift from adding statistics / citations / quotations

Source：Princeton GEO paper, KDD 2024, GEO-bench 10,000 queries，2024

Sources

  1. [1]Optimizing your website for generative AI features on Google Search．Google Search Central．2026-05-15
  2. [2]GEO: Generative Engine Optimization．Aggarwal et al., KDD 2024．2024

Updated 2026-08-10