Skip to content

The AI search evidence ledger

Every claim below circulates widely in writing about generative engine optimisation. We went to the primary source for each one and tried to reproduce the number. The grade records what happened. Where we failed to reproduce something, we say so rather than repeat it, and where we got something wrong ourselves, it is in corrections.

19 claims · last checked 2026-07-10 · how we grade

The scale

VERIFIED
We fetched the primary source ourselves and the numbers reproduced.
DIRECTIONAL
A real, named source exists, but it is single-sourced or only partly reproducible.
UNREPRODUCED
A source is cited for the number, but we could not surface that number from it.
UNTRACEABLE
The number circulates widely with no locatable primary source.
MISLEADING
The source is real, but the claim misrepresents what it measured.
REFUTED
A named primary source directly contradicts the claim.

REFUTED

4 claims
REFUTED #answer-instability

The claim, as it circulates

“Brands hold a ranking position inside AI answers, and you can track it.”

Measured and contradicted. There is under a 1-in-100 chance of even the same list twice.

What the source actually says

Same list of brands: under 1 in 100. Same list in the same order: roughly 1 in 1,000. Runners used default, personalised settings rather than pinned temperature, deliberately reflecting real-world conditions. The direct implication is that a single 'AI ranking position' is not a stable quantity.

“there's a <1 in 100 chance that ChatGPT or Google's AI, if asked 100X, will give you the same list of brands in any two responses”

Source
SparkToro (Rand Fishkin) with Gumshoe.ai (Patrick O'Donnell) , AIs are highly inconsistent when recommending brands or products
Published
November to December 2025
Sample
600 volunteers, 2,961 runs, 12 brand-recommendation prompts run 60 to 100 times each, across ChatGPT, Claude and Google AI Overviews / AI Mode
Reproduction
We reproduced this from the primary source. checked 2026-07-10

We fetched the SparkToro post and confirmed both probabilities and the methodology verbatim.

REFUTED #consensus-gap

The claim, as it circulates

“AI visibility is one number, and a brand visible in ChatGPT is visible across AI search.”

Only 2.37% of cited URLs appear in all three major engines. 91.07% appear in exactly one.

What the source actually says

Citations are overwhelmingly fragmented by engine: 2.37% universal overlap, 91.07% single-engine. The commercial consequence is that any tool reporting one blended 'AI Visibility %' across engines is averaging three near-disjoint populations.

“only 2.37% of cited URLs show up across all 3 engines”

Source
Kevin Indig, Growth Memo , The Consensus Gap
Published
2026-05-11
Sample
3,700,000 URL citations from a 20,000-prompt random sample, across ChatGPT, Perplexity and Google AI Overviews, Q3 2025 to Q1 2026, from Omnia's live prompt monitoring pool
Reproduction
We reproduced this from the primary source. checked 2026-07-10

We fetched the article and confirmed both figures and the dataset. One caveat the secondary coverage consistently omits, which we found on the source: the prompt pool is Europe-weighted and Spain-heavy, plus UK, Nordic and EU markets. It is not a global or US-representative sample. Treat the exact percentages as European.

REFUTED #schema-boosts-citations

The claim, as it circulates

“Adding schema markup increases how often AI assistants cite your pages.”

The publisher that sells rank tools ran the experiment and found no uplift on any platform.

What the source actually says

Across the platforms measured, the change in citations after adding schema was statistically indistinguishable from zero: Google AI Mode +2.4%, ChatGPT +2.2%, and Google AI Overviews actually -4.6% (a small but significant decline). The pages studied were already receiving AI citations, so the experiment tests whether schema lifts already-cited pages, not whether it helps undiscovered ones.

“Adding schema produced no major uplift in citations on any platform.”

Source
Ahrefs (Louise Linehan, Xibeijia Guan) , We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved.
Published
2026-05-11
Sample
1,885 URLs that added JSON-LD, against 4,000 matched controls, Aug 2025 to Mar 2026
Reproduction
We reproduced this from the primary source. checked 2026-07-10

We fetched the article and confirmed the conclusion, the per-platform figures, the sample of 1,885 treated pages against 4,000 controls, the dates and the authors. Ahrefs sells rank tracking rather than schema tooling, so if anything the incentive ran towards a tidy positive result; it did not find one.

REFUTED #long-form-cited-more

The claim, as it circulates

“Long-form content (2,000+ words) gets cited more often by AI.”

Word count and citation are essentially uncorrelated, and most cited pages are short.

What the source actually says

There is almost no relationship between how long a page is and whether it is cited: Spearman 0.04. In fact 53.4% of pages cited by AI Overviews are under 1,000 words, and only 16.0% are over 2,000. The idea that long-form mega-guides dominate AI citation is contradicted by the distribution.

“The Spearman correlation between word count and citation position is 0.04, essentially zero.”

Source
Ahrefs , Short vs Long Content in AI Overviews
Published
2025-12-03
Sample
560,346 AI Overviews, 1,677,876 cited URLs, 174,048 pages with valid word counts
Reproduction
We reproduced this from the primary source. checked 2026-07-10

We fetched the article and confirmed the 0.04 correlation, the 53.4% under-1,000-words figure, the sample sizes and the date. Scope caveat worth carrying: this dataset is Google AI Overviews only, not ChatGPT or Perplexity.

MISLEADING

5 claims
MISLEADING #forty-percent

The claim, as it circulates

“Generative engine optimisation boosts a brand's visibility in AI answers by up to 40%.”

The paper says this. The paper is measuring a benchmark, not ChatGPT.

What the source actually says

The uplift is measured on GEO-bench, a research benchmark the authors introduce in the same paper, using a 'visibility' metric derived from how much of a generated response is attributable to a given source. It is a controlled lab result. It is not a measurement of citation behaviour on ChatGPT, Perplexity, Claude or Google AI Overviews, and the paper does not claim it is. Almost every marketing page that repeats the 40% omits this.

“GEO can boost visibility by up to 40% in generative engine responses.”

Source
Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande , GEO: Generative Engine Optimization
Published
2023-11-16, last revised 2024-06-28 (v3). Accepted to KDD 2024.
Sample
GEO-bench, a benchmark of diverse user queries across multiple domains
Reproduction
We reproduced this from the primary source. checked 2026-07-10

We fetched the arXiv abstract and confirmed the sentence verbatim, and confirmed GEO-bench is introduced by the same paper. The abstract page does not state GEO-bench's query count, so we do not publish one. Figures of '~10,000 queries' circulate; we could not confirm that number from the abstract and have not read the full PDF.

MISLEADING VENDOR-INTERESTED #ai-overviews-top-10

The claim, as it circulates

“If you rank in Google's organic top 10, you will be cited in AI Overviews.”

It was mostly true a year ago. It is now true of well under half of citations.

What the source actually says

The share of AI Overview citations that also appear in the SERP's first 10 blocks fell from roughly 76% in July 2025 to 37.9% in January 2026. Restricted to organic results only, the figure is 37.10%. The authors attribute the shift to AI Overviews drawing more on the SERPs of fan-out sub-queries than on the original query's SERP.

“Google is selecting far fewer pages straight from the original SERP (~76% in July 2025 vs. ~38% today).”

Source
Ahrefs (Louise Linehan, Xibeijia Guan) , AI Overview citations and the top 10
Published
2026-03-02
Sample
863,000 keyword SERPs; 4,000,000 AI Overview URLs
Reproduction
We reproduced this from the primary source. checked 2026-07-10

We fetched the article and the figures, sample size, authors and date all matched exactly.

MISLEADING #zero-point-one-percent

The claim, as it circulates

“AI assistants return the same brand list less than 0.1% of the time.”

Off by a factor of ten. 0.1% is the figure for the same list in the same ORDER.

What the source actually says

Two distinct figures exist and are routinely conflated. List membership repeats at under 1 in 100 (under 1%). List ordering repeats at roughly 1 in 1,000 (about 0.1%). Any page citing '0.1% chance of the same list' has swapped one for the other.

“when it comes to ordering, AI tool responses are so random that it's more like 1 in 1,000 runs before you'd see two lists in the same order”

Source
SparkToro (Rand Fishkin) with Gumshoe.ai , AIs are highly inconsistent when recommending brands or products
Published
November to December 2025
Sample
as above
Reproduction
We reproduced this from the primary source. checked 2026-07-10

We caught this because we made the error ourselves. During the research behind this site we were handed the conflated version and repeated it before checking the primary source. The correction is logged at /corrections/.

MISLEADING VENDOR-INTERESTED #reddit-most-cited

The claim, as it circulates

“Reddit is the most-cited source in AI answers.”

True on some engines, false on the biggest one. It flips by platform.

What the source actually says

The most-cited domain depends entirely on which engine you mean. On ChatGPT the top source is Wikipedia at 7.8%, with Reddit second at 1.8%. On Google AI Overviews and Perplexity, Reddit is first (2.2% and 6.6%). A blanket 'Reddit is most-cited' is true for two engines and false for ChatGPT, the largest.

“Reddit emerges as the leading source for both Google AI Overviews (2.2%) and Perplexity (6.6%).”

Source
Profound , AI platform citation patterns
Published
August 2024 to June 2025
Sample
680,000,000 citations across ChatGPT, Google AI Overviews and Perplexity
Reproduction
We reproduced this from the primary source. checked 2026-07-10

We fetched the Profound analysis and confirmed the per-platform leaders and their percentages, and the 680M-citation dataset. The claim is graded MISLEADING because it states as a universal fact something that is platform-specific and reverses across platforms. Profound sells AI visibility tracking.

MISLEADING VENDOR-INTERESTED #adobe-58-percent-ai-shopping

The claim, as it circulates

“Adobe (or Salesforce) found that 58% of consumers now use AI instead of search engines for product recommendations.”

The 58% is real, but it is Capgemini's, not Adobe's, and it is a survey of what people say they do.

What the source actually says

The 58% figure comes from Capgemini, not Adobe or Salesforce, and it is a self-reported survey of stated behaviour (up from 25% the year before), not a measurement of actual traffic. Adobe did publish a 58% figure in the same period, but it is about consumers' willingness to share personal information for convenience, an unrelated question, which is the likely root of the misattribution.

“More than half (58%) have replaced traditional search engines with Gen AI tools as their go-to for product/service recommendations.”

Source
Capgemini Research Institute , What Matters to Today's Consumer (4th edition)
Published
2025-01-09
Sample
12,000 consumers across 12 countries, surveyed October to November 2024
Reproduction
We reproduced this from the primary source. checked 2026-07-10

We fetched the Capgemini release and confirmed the 58% wording, the up-from-25% comparison, the 12,000-consumer sample, the countries and the dates. Graded MISLEADING because the claim as circulated attributes the number to the wrong company and presents a stated-preference survey as measured behaviour. The underlying Capgemini number is real.

UNTRACEABLE

4 claims
UNTRACEABLE #llms-txt-helps

The claim, as it circulates

“Adding an llms.txt file improves your chances of being cited by AI assistants.”

No study we can find demonstrates it. Google's John Mueller says no AI service has confirmed using the file.

What the source actually says

Nothing traceable supports the positive claim. No controlled study, no server-log analysis and no vendor dataset that we could locate shows llms.txt increasing citation. The evidence that exists points the other way, though see the companion claim on this page, because that counter-evidence has its own problems.

Source
unattributed; circulates across vendor blogs and SEO guides
Published
unknown
Sample
none disclosed
Reproduction
No primary source exists to reproduce. checked 2026-07-10

We looked for a study supporting the claim and found none. We publish an llms.txt file ourselves, to describe the site rather than because we believe it changes anything. If you can produce a controlled study either way, we will grade it.

UNTRACEABLE #youtube-correlation

The claim, as it circulates

“YouTube mentions correlate with AI visibility at r = 0.737.”

We cannot find the study. The Semrush study that does exist says it did not test this.

What the source actually says

Nothing, as far as we can establish. The Semrush study we could locate (11,882 prompts, 337,785 URLs, published 14 January 2026) explicitly states that it did not test YouTube or backlink correlations. The 0.737 figure appears across multiple vendor blogs with no primary link.

Source
attributed to a 'Semrush AI Visibility Index' of 126 million prompts
Published
unknown
Sample
claimed 126,000,000 prompts
Reproduction
No primary source exists to reproduce. checked 2026-07-10

No primary source located. Until someone produces the underlying report, treat this number as unsupported. If you have a primary link, send it and we will regrade in public.

UNTRACEABLE #reddit-12x

The claim, as it circulates

“A 100-word Reddit comment gets cited by AI 12 times more often than a 2,000-word guide.”

A suspiciously precise number with no methodology anywhere behind it.

What the source actually says

Nothing traceable. That Reddit and other user-generated content carry meaningful AI citation share is separately supported, notably by Profound's citation dataset. The specific '12x versus a 2,000-word guide' comparison is not, and reads as marketing copy rather than a measurement.

Source
unattributed; circulates across vendor blogs
Published
unknown
Sample
none disclosed
Reproduction
No primary source exists to reproduce. checked 2026-07-10

No primary source located. Note the shape of this claim: an exact multiplier, two arbitrary word counts, no sample, no engine named, no date. That shape is the tell.

UNTRACEABLE #market-size

The claim, as it circulates

“The GEO market grows from $848M in 2025 to $33.7B by 2034, a 50.5% CAGR.”

Repeated across dozens of blogs. We could not trace it to any named research firm.

What the source actually says

Nothing we could verify. Separate analyst projections for this category do exist and are also estimates rather than measurements. Capital flows are checkable and are a better signal: Profound raised a $96M Series C at a $1B valuation in February 2026.

Source
unattributed; circulates across secondary blogs
Published
unknown
Sample
none disclosed
Reproduction
No primary source exists to reproduce. checked 2026-07-10

No primary source located. Market-size projections in emerging categories are, by construction, estimates. Do not use one as evidence that a tactic works.

UNREPRODUCED

1 claim
UNREPRODUCED VENDOR-INTERESTED #llms-txt-ignored

The claim, as it circulates

“AI crawlers requested /llms.txt in only 0.1% of hits: 84 out of more than 62,100.”

We believe the direction. We could not get these numbers off the page they are attributed to.

What the source actually says

The article argues that marketers misunderstand llms.txt and that AI bots largely ignore it. When we fetched the page, it did not yield the sample size, the date range, or the 84-of-62,100 figures that are widely attributed to it across other blogs.

Source
Otterly.ai , Llms.txt Experiment: What Marketers Get Wrong about llms.txt
Published
2026-02-05
Sample
reported as a single experiment site, server logs, about 90 days
Reproduction
We tried to reproduce this and could not. checked 2026-07-10

This is a claim we would like to be true, because it is convenient for our position on llms.txt. That is precisely why it gets UNREPRODUCED rather than a pass. The qualitative direction is corroborated independently by John Mueller's public comments; the specific numbers are not currently reproducible by us and we will not repeat them. Otterly sells AI search monitoring, so the finding is vendor-interested in the direction of 'a static file is not enough, buy monitoring'. That does not make it wrong.

DIRECTIONAL

2 claims
DIRECTIONAL #geo-term-decline

The claim, as it circulates

“Interest in generative engine optimisation is growing rapidly.”

Interest in the TERM peaked in August 2025. Underlying AI usage is a different question, and is still growing.

What the source actually says

Both terms were noise in 2023, measurable in 2024, scaled through 2025, then peaked in late summer 2025 and plateaued. GEO carries roughly twice AEO's search volume, yet the report observes practitioners converging on 'AEO' as the standard label.

“AEO sits ~17% below its September high; GEO ~45% below its August high”

Source
Rankability , State of AI Search
Published
2026
Sample
Google Keyword Planner export, 3,751 keywords x 48 monthly data points, June 2022 to May 2026
Reproduction
We reproduced this from the primary source. checked 2026-07-10

We fetched the report and confirmed the wording and methodology. Graded DIRECTIONAL rather than VERIFIED because it is single-sourced: we found no second independent dataset measuring term-level trend, and the report gives no absolute volumes. Do not treat the exact percentages as settled.

DIRECTIONAL #geo-is-a-category

The claim, as it circulates

“Standalone GEO tracking is a durable product category in its own right.”

Contested by someone who tried it. A GEO tracking startup shut down and its founder concluded the opposite.

What the source actually says

Founder Benjamin Houy shut Lorelight down after roughly seven months. His stated reasons: 'Customers were churning because the product didn't change what they needed to do', and 'There's no secret GEO strategy. AI models reward the same fundamentals that already drive SEO and PR.' Consolidation supports the reading: Sitecore acquired Scrunch AI in June 2026, and Adobe agreed to acquire Semrush for about $1.9 billion.

“GEO makes more sense as a feature within existing SEO platforms, not as a standalone category.”

Source
Search Engine Land , GEO startup Lorelight shuts down
Published
2025-11-04
Sample
one company, April to October 2025
Reproduction
We reproduced this from the primary source. checked 2026-07-10

We fetched the Search Engine Land report and confirmed the quotes, the founder's name and the timeline. Graded DIRECTIONAL, not VERIFIED, because it is one founder's experience with one product, which is evidence but not proof. It is nonetheless the single most useful data point we found, because it is a first-hand account against the speaker's own commercial interest.

VERIFIED

3 claims
VERIFIED VENDOR-INTERESTED #citations-outside-top-10

The claim, as it circulates

“Most AI Overview citations now come from pages that do not rank in Google's organic top 10.”

True as of January 2026, and it was not true six months earlier.

What the source actually says

Ahrefs reports that 37.10% of AI Overview citations come from pages in the organic top 10. The complement, 62.90%, therefore come from outside it. The study attributes the shift to AI Overviews drawing on the SERPs of fan-out sub-queries rather than the original query's SERP. The practical consequence is that optimising only the pages that already rank for your target query addresses a shrinking minority of citations.

“Google is selecting far fewer pages straight from the original SERP (~76% in July 2025 vs. ~38% today).”

Source
Ahrefs (Louise Linehan, Xibeijia Guan) , AI Overview citations and the top 10
Published
2026-03-02
Sample
863,000 keyword SERPs; 4,000,000 AI Overview URLs
Reproduction
We reproduced this from the primary source. checked 2026-07-10

We fetched the article and confirmed the 37.10% organic-only figure, the sample, the authors and the date. Note that the 'most come from outside the top 10' phrasing is our arithmetic complement of Ahrefs' number, not a sentence Ahrefs prints. We flag it because deriving a claim from a source is exactly the step where numbers usually go wrong. Ahrefs sells Brand Radar, so the finding is vendor-interested.

VERIFIED VENDOR-INTERESTED #ghost-citations

The claim, as it circulates

“Most pages an AI cites as a source never get their brand named in the answer.”

True in the study that coined the term: 61.7% were cited without the brand appearing. Small sample, though.

What the source actually says

In 61.7% of cases the AI used a page as a linked source but did not mention the brand in the text of the answer, which undercuts the idea that a citation reliably delivers a brand mention. Note the modest sample: 115 prompts. Perplexity was not included.

“Almost 62% (61.7%) were ghost citations. AI platforms used the page as a source link, but the brand name never appeared in the actual answer.”

Source
Semrush , The Ghost Citations Study
Published
2026-06-09
Sample
3,981 domain appearances across 115 prompts, 14 countries, on ChatGPT, Gemini, Google AI Overviews and Google AI Mode
Reproduction
We reproduced this from the primary source. checked 2026-07-10

We fetched the article and confirmed the 61.7% figure, the 3,981 domain appearances, the 115 prompts and the date. We grade it VERIFIED because the number reproduces exactly, while flagging that 115 prompts is a thin base for how broadly this stat is now generalised. Semrush sells an AI visibility toolkit.

Found something we got wrong?

Send the primary source. If it holds up, we regrade the claim in public and log the change in corrections, with the date and what changed. We do not silently edit pages.

Send us a source →