How this site measures citations — and makes content

The citation rate is the share of questions for which an AI search engine displays aeo.io.kr as a source in its answer — as a link, a footnote, or an entry in a source panel — out of the questions for which the engine displayed an AI answer at all.

We chose this unit for a simple reason. The goal of AEO is not a ranking position; it is being selected as a source when an AI composes an answer. Rankings and traffic do not measure that event directly. A citation, by contrast, is an observable on-screen event: it can be captured in a screenshot, and anyone can attempt to reproduce it by asking the same question.

AI answers are generated probabilistically, however. A citation rate is therefore never a yes/no claim — it is only meaningful as a ratio over a fixed question set. That is why the questions, the judging criteria, and the protocol are all published and frozen below.

Citation rate for engine E (%) = (questions where E's answer displays aeo.io.kr as a source — link, footnote, or source panel) ÷ (questions for which E displayed an AI answer) × 100

The denominator — "questions for which E displayed an AI answer" — is the number of questions actually asked minus those judged no-ai-answer. Each question is judged into exactly one of four states:

• cited — aeo.io.kr appears as a source (link, footnote, or source panel). The only state counted in the numerator.
• mentioned — named in the answer body without a link. Tracked separately as a secondary signal; not counted in the citation rate.
• absent — does not appear in the answer at all.
• no-ai-answer — the engine shows no AI answer for that question. Excluded from the denominator and tallied separately.

The sample is a fixed set of 20 questions × 4 engine families = 80 cells. The full question texts are published below: 6 definition, 4 mechanism, 5 how-to, and 5 purchase-decision questions.

Engines measured:

• ChatGPT (search enabled)
• Perplexity (default mode)
• Google AI Overviews (Korean-language Google Search)
• Naver AI Briefing

Protocol:

• Measured once per quarter; the measurement date is recorded.
• Personalization is minimized — fresh sessions wherever possible.
• Each question is asked exactly once, verbatim. No retries, no leading follow-ups.
• Every cell's answer screen is preserved as a screenshot.

Results are published only as a per-engine grid. We do not publish a combined average across the four engines: they differ in when an AI answer appears at all and in how sources are displayed, so an average would distort the picture in either direction.

The Naver AI Briefing column always carries this footnote: "Closed and personalized by design; not directly comparable to the other engines."

This methodology has stated limitations.

• Non-determinism — the same engine can answer the same question differently. A single query is a sample of that moment, nothing more.
• Personalization — even with fresh sessions, the influence of region, language, and account signals cannot be fully removed.
• Variance — one query per quarter is a high-variance design. Trends across quarters matter more than any single cell.

Honesty clause:

• Results are published as measured, even when they are low.
• Unmeasured cells are marked "pending" — never filled in with a number.
• This methodology was not designed to favor us. We apply the same yardstick to aeo.io.kr first.

v1.0 freezes the 20 questions and the judging criteria. Wording-only fixes that cannot affect results bump the minor version (v1.1); changes that can affect results, such as replacing questions, bump the major version (v2.0). Every change is recorded in this changelog, and numbers from different versions are never compared directly.

Changelog

• v1.0 (2026-07-12) — Methodology established. 20 questions frozen; formula, four-state judging, measurement protocol, and publication rules fixed.

Every article on this site passes through this pipeline before publication.

1. Draft generation — drafted against the topic-cluster plan.
2. Review — structure and logic are checked; hype ("instant citations", "guaranteed #1") and unsupported claims are removed.
3. Web fact-check audit — every concrete figure and URL in the body is checked against primary sources: official documentation from engines, platforms, and standards bodies. Anything that cannot be verified against a primary source is deleted or rewritten qualitatively.
4. Publication under the five invariants — an article ships only when all five are in place:
   • a question-shaped H1
   • a self-contained answer in the first paragraph
   • Article + FAQPage JSON-LD
   • cluster internal links
   • the pipeline footer (a permanent link back to this document)

Whether content produced this way actually gets cited is measured — on ourselves first — using the citation-rate methodology above, and published.