Skip to content
Broadcastwell Fix one answer, $490
Menu

Numbers we could not verify, and claims the originals correct

GEO’s reported 41% improvement concerns position-adjusted word-count visibility, not citation probability (Aggarwal et al., 2024, Table 1).

The 41% GEO claim has a real ancestor, but “citation probability” is the wrong outcome. Aggarwal and colleagues' Table 1 describes improvement in a position-adjusted word-count visibility metric, with a separate subjective-impression measure. It is not a measured probability that a business will be cited. Aggarwal et al., 2024, author manuscript v3, Table 1.

This ledger preserves the claim as encountered, then states what survives checking. It includes missing sources, changes of unit, invalid inference and corrections to our own initial objections. A rejected sentence is not proof that its underlying topic is unimportant, or that an author acted dishonestly.

You can use the status column to distinguish an unavailable report from a claim contradicted by an opened original. Those are different findings. A number appearing repeatedly in secondary sources has not acquired independent confirmation merely through repetition.

What do the recovered originals actually measure?

Claim encountered Where it appeared Primary check and status
Statistics boost AI visibility by up to 40%; expert quotations increase citation probability by 41% Vidhate, Jadhav and Gole, IJNRD 2026, PDF p. 3 Aggarwal et al.'s abstract uses “up to 40%”; Table 1 describes a 41% best improvement in position-adjusted word count. The conditions and metrics must remain attached. A universal citation-probability uplift is unsupported. Corrected metric belongs in the quotable ledger.
Voice searches average 29 words Anttila, 2021, p. 9 Backlinko's original 2018 study reports the average answer length in its Google Home results. It is not query length. The corrected historical observation is usable with that sample and date.
96% of businesses are not voice-search ready Anttila, 2021, p. 2 Uberall's 2019 primary release concerns 73,000 business locations and its own listings-readiness definition. It reports only 3.82% with correct information on its three principal directories. That does not establish the share of all businesses technically unable to appear in voice search.
Organic traffic research proves a general marketing tactic works Secondary accounts citing Baye et al., 2016 The publisher's original studies retail organic clicks in August 2012. Its sample, econometric design and outcome are identifiable. It does not supply a general effect on present-day AI-answer citation.

The full Backlinko population is stated as 10,000 Google Home results; the reported average is a historical vendor study result. Its rounded average does not provide the individual-length distribution or a reconstructable uncertainty interval. It should not become a required answer length for every page you publish.

Uberall's release also distinguishes its readiness score from real-world return on investment. A vendor-defined score can identify missing listing information under its rules. It cannot, without another design, establish how often those errors caused an assistant to omit a business.

Which claims remain unsupported?

Claim encountered Source location Why the sentence remains withheld
Half of cited AI-search content is less than 13 weeks old IJNRD 2026, PDF p. 4 No identifiable primary dataset, sampling window or age definition is supplied for this statistic. We have not established a source that supports the stated population.
E-E-A-T has become measurable Entity Confidence Scores IJNRD 2026, PDF p. 3 The paper supplies no primary operator specification for the promoted metric. It is not accepted here as a defined Google score.
A pooled SEO effectiveness effect size of d = 1.049 Usmany et al., 2024, abstract and results The methods refer to 50 included studies while the results describe 10 and label rows P1 to P10 without identifying the included studies. The pooled result cannot be checked from the reported study mapping.
SEO produces “85.4%” brand positioning Ahmad et al., F1000Research versions, Table 1 The table reports Pearson r = 0.854 from a convenience-sampled questionnaire. A correlation is not the percentage of a business outcome attributable to SEO, and the design does not establish causation.
88% of search users never click paid ads Historical HubSpot ebook, p. 5 Its citation chain points to a discussion rather than a checkable primary sampling report. We do not treat it as measured user behavior.
Organic converts three to one over paid Kuehn interview, SEO Bible, 2009 A practitioner's recollection is not a recoverable comparative conversion study. Attribution can preserve testimony; it cannot turn the ratio into a general fact.
Up to 70% of searches happen on destination sites Hoppe interview in SEO Bible, citing IDC The underlying IDC study and its population have not been recovered here. No current or general rate is asserted.
SEO has higher long-run ROI as a measured finding Singh and Kesarwani, IJRPR 2025 The supplied paper discusses the comparison but does not present an auditable outcome dataset establishing that claim. The leftover phrase “next point same instructions” appears on printed p. 4132, PDF p. 4; it is an editorial defect, not proof of research fraud.
A historical traffic share is the share of searchers who never leave page one Secondary retellings of Chitika, 2013 The recovered primary report measures Google referral traffic visible through Chitika ad impressions, not unique users or all searches. Its Figure 2 prints 91.5% for page one, while its prose rounds to 92%. Methods: May 21 to 27, 2013, US and Canada, network coverage only. The blanket user-behavior claim remains withheld.
Voice survey percentages describe the share of all search queries Anttila and related secondary accounts citing Microsoft, 2019 A recovered primary copy reports respondents and selected activities, not query traffic. Its methods describe two surveys; the summary’s 69% digital-assistant usage and usage page’s 72% are not accompanied by a figure-specific base mapping that permits us to combine them. Those historical survey figures do not establish present query shares.

We initially could not retrieve the legacy Chitika and Microsoft routes. The primary reports have since been opened in archive mirrors, so those availability objections are withdrawn. Their narrower populations and units remain essential. Chitika gives no exact study impression count from which we could reconstruct a Wilson interval; Microsoft does not map each headline to a recoverable response base. Neither report licenses a present-day population rate.

Which deposited calculations failed to reproduce?

Rovira’s one-term title-restricted Google Scholar workbook takes medians over 23 query columns while the described design has 25. Including the omitted theory and variable columns changes 456 of 1,000 position medians and the rounded correlation from 0.990 to 0.991. The other eight reported configurations reproduce at three decimals. Rovira data, 2019, Full1 B2:B1001 and C:AA.

The university paper prints a negative SEO-rank correlation and p = 0.333; its deposited 100 rows produce rho = +0.04382 and a conventional two-sided approximation p = 0.6651. Its website-quality regression gives F(1,98) = 7.4898 from the data, rather than the printed 7.04. These are reproducibility discrepancies, not evidence of intent; the weak SEO association remains nonsignificant. Giannakoulopoulos data, 2019, Results Sheet1; paper Table 9 and regression section.

Which proposed rules lack evidence?

Proposed rule Source family What the record permits
Keyword density should be 5 to 7%, 7 to 10%, 2 to 4%, 1 to 2%, or 15 to 20% in a Yahoo title Old practitioner book, LPU course text and Naila R. guide Conflicting prescriptions without a transferable experimental basis. No density target is recommended from this collection.
Bounce rate, time on site or social signals are proven ranking factors Ziakis article, LPU text and HubSpot ebook Assertion and observational correlation do not identify a causal ranking input. This is not evidence that those inputs never matter; it is an evidentiary limit on these sources.
W3C validation boosts rank, or educational and government domain endings rank higher Ziakis 2019 and derivative teaching The cited exploratory work does not establish the proposed causal rules. A domain ending is not a substitute for checking who published a claim.
The strongest factor in the Ziakis table is a proven optimization prescription Ziakis et al., 2019, Table 2 The small exploratory table's average coefficients do not meet the paper's stated absolute-correlation bar. We use it to examine the evidence problem, not as a list of established effects.
A practitioner textbook is current operator documentation LPU course text and historical SEO books Named tools and policies require current primary verification. A historical quotation remains historical.
Microsoft Academic is a currently available service A present-tense retelling of Rovira's 2019 comparison The study is a dated comparison. Its historical system names must not become an unverified present-day availability claim.

Which initial objections did the originals disprove?

Earlier interpretation What the original actually supports
The 48.4%, 27% and 31% prose shares contradict the optimization figure They describe news-content shares. We withdraw the proposed contradiction; they are not alternative “probably optimized” values. Lewandowski et al., §4 and Figure 1.
Wikipedia held 27% of every first-position result The stated base is Wikipedia's 1,974 results; 27% of those were at position 1 and 79.3% were in the top ten. The reversed denominator is unsupported. Lewandowski et al., §5.
Google results were simply 88% optimized The figure separates 59.0% definitely and 29.0% probably optimized, with 12.0% in the combined negative group. Retain the uncertainty labels, source year and sample.
Wikipedia's rounded prose share and its stated result count can be combined without qualification The prose's 0.7% does not directly reconcile with 1,974 / 256,853. We retain the explicit count for the position statement and do not invent an exact numerator from the rounded share.

Corrections belong in the record even when the initial error came from our own reading material. There is no useful reason to preserve an allegation after its source has disproved it. The unresolved rows remain separate from the withdrawn ones, so a reader can see which evidence would actually change each status.

What our own data shows

We have no measurement on this point: the number of unsupported claims in this selected reading collection does not estimate a prevalence of misinformation across SEO or GEO publishing. Our evidence standard governs individual uses. Only corrected, bounded findings are carried into the quotable-numbers ledger.

Sources

Cite this page

Broadcastwell research team. (2026, October 30). Numbers we could not verify, and claims the originals correct. Broadcastwell. https://docs.broadcastwell.com/research/evidence/unverified-numbers/ Published 30 October 2026; updated 30 October 2026.

@online{broadcastwell2026research23,
  author = {{Broadcastwell research team}},
  title = {Numbers we could not verify, and claims the originals correct},
  year = {2026},
  date = {2026-10-30},
  url = {https://docs.broadcastwell.com/research/evidence/unverified-numbers/},
  urldate = {2026-10-30},
  note = {Published 2026-10-30; updated 2026-10-30}
}

Prepared by the Broadcastwell research team. Lead: Vikram Patel, AI architect and specialist.

Sources checked 30 October 2026. Independent check: 30 October 2026. This page has not been peer reviewed.

Last updated: 30 October 2026. Next review: 30 January 2027.

Continue

Broadcastwell research. Sources and dates are stated on each page. Prose and derived tables CC BY 4.0, site code MIT. External sources retain their own licenses. Not peer reviewed. Evidence standard.