PAPER

Measurement at scale — where the citations in AI answers actually come from

Where do the sources cited by AI answers actually live, and how far does that vary with brand size?

Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines

Authors: Pratyush Kumar
Affiliation: Ranqo (a commercial GEO tracking platform)
Venue: arXiv preprint (single author)
Submitted: 2026-06-18
arXiv: arXiv:2606.20065
We verified: 2026-08-22

WHAT THE PAPER SAYS

Mostly outside the brand's own site. Across 149,912 citations, the brand's own domain accounted for just 2.9%, while company and third-party brand pages made up 75.2%. Appearance rates diverged sharply by brand size. Note that this data comes from a company selling GEO tracking tools, produced with its own platform.

Read the original on arXiv ↗

⚠️ THIS PAPER HAS A DISCLOSED CONFLICT OF INTEREST

The author states in the paper that he is a co-founder of, and holds equity in, Ranqo — the platform analysed — and writes that “Readers should weigh the findings as a vendor-produced measurement study.”

METHOD

How it was measured

Read the conditions before the numbers. The same figure means something different under a different sample or environment.

Sample: 102 brands · 3,508 tracking runs · 102,025 prompt responses
Totals: 15,815 brand mentions · 149,912 source citations
Period: March–May 2026
Engines: ChatGPT (GPT-5) · Gemini-3 · Perplexity (Sonar) · Claude Sonnet · Grok-3
Prompts: Generated with Claude Sonnet (temperature 0.7); queries generally do not contain the brand name

FINDINGS

What came out

Distribution of cited sources: Company and third-party brand pages 75.2% · YouTube/video 4.2% · tech and business media 3.8% · communities such as Reddit 3.3% · academic, government, developer and social 3.1% · the brand's own domain 2.9% · Wikipedia 2.6% · G2/Capterra 1.1%
Content format: 59% of citations were content-type, and among those ranked listicles were 35.7% (21.0% of all citations) — the largest single format
Appearance rate by brand tier: On unbranded queries: Tier 1 72.9% (n=11) · Tier 2 43.6% (n=36) · Tier 3 11.4% (n=55)

LIMITATIONS

Limitations the authors state themselves

Not our criticism — this is what the authors wrote in the paper.

Not a causal claim — this is a baseline at the moment of observation, not a treatment effect.
A convenience sample skewed toward SaaS, retail, fintech and Indian DTC; the paper does not claim industry representativeness.
Tier classification is manual, done by hand from public signals of brand prominence (Wikipedia, press, funding).
Platform opacity — every result is inferred from API output rather than observed inside the engines.
One model variant per engine was fixed for measurement.
Sentiment classification is heuristic, so the reported 45.5% sentiment-shift rate carries classifier noise.

NAVIRANG'S READING — NOT THE PAPER'S CONCLUSION

The direction is more useful than the exact numbers. That polishing your own pages alone will not get you into answers, and that ranked listicles take a large share of citations, both match what practitioners see. But it must not be cited without noting that a vendor produced it from its own platform data — the author asked to be read that way. The sample skews to SaaS and fintech, so it does not transfer to Korean hospitals, law firms or education businesses. You have to measure your own vertical.

IN COMPARISON

Where it diverges from other papers

The claims-table entries this paper appears in.

Is fixing your own site enough?

This paper's position — The share held by brands' own domains was small

Can visibility be guaranteed?

This paper's position — You cannot see inside the engines

If you want your own brand's numbers rather than a paper's

Every figure here came from someone else's sample. Send us a URL and we ask all 7 answer engines directly and measure yours.

Request a free AEO audit 070-4015-6848

We reply within one business day.