LEARN

How to Judge Whether a Vendor's Claim Holds

Four inclined perforated plates with unique holes; dense particles passing all form a luminous cluster below, representing credible information filtered through four questions
Judging a claim's credibility requires filtering through four questions: evidence source, sample size, official platform counterstatements, and verifiable consistency

IN ONE SENTENCE

Four questions decide whether a claim holds: the source, the sample size, whether a platform said the opposite, and whether the definition is checkable.

Outside Google, mechanisms in this field are undisclosed, which sets a low bar for claims. Whether a claim is worth believing comes down to four questions: the source, the sample size, whether a platform has said the opposite, and whether the definition is checkable.

OUR POSITION

Use the four as a filter. The point is not demanding proof of everything — nobody in this field can provide it — but whether they can separate what was measured from what was inferred. Anyone who cannot is the higher risk.

01

The four questions

What is the source: a named study, publisher and date. 'The industry generally believes' is not a source.

How large was the sample: how many questions, platforms and days. n=1 self-reported cases are everywhere here.

Has a platform said the opposite: take llms.txt — Ahrefs found 97% of files receiving zero requests across 137,000 domains in May 2026, and Google says it does not use them. Anyone still selling it as a core lever has a problem.

Is the definition checkable: does the number carry its tool, date, region and judgement rule.

02

Three phrasings to treat with suspicion

'Proprietary algorithm' or 'private indexing channel': no platform has opened a dedicated channel to any vendor.

'Guaranteed rank' or 'guaranteed mention rate of X': AI answers have no stable position structure, and mention rate depends on the question set — a number without a definition means nothing.

A specific conversion multiple: published ranges run from 4.4× to no significant difference (p=0.794). With that spread, committing to a multiple has no basis.

03

What counts as a good signal instead

Being able to separate measurement from inference, and willing to label uncertainty.

Tiered commitments: mentions only during a pilot, traffic only in a full engagement, with every committed metric appearing as a same-named report column.

Willing to produce a baseline with its definition before producing a target number.

Data behind this page

97%

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

Source：Ahrefs, across 137,000 domains，2026

4.4× / no significant difference

AI visitor value versus organic search (direction agrees, magnitude does not)

Source：Semrush (Jun 2025) versus Amsive's analysis (p=0.794)，2025

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