Measuring GEO in Korea: Naver Gets Its Own Line

Measuring GEO in Korea works like measuring GEO anywhere, with one non-negotiable difference: Naver AI Briefing is reported as a separate line, never blended into a global visibility average. The arithmetic is what makes this rule rather than preference. A brand at 45% citation rate across ChatGPT, Gemini, Perplexity and Claude, and 5% inside AI Briefing, reports a blended 37% that looks perfectly healthy. The 5% is the number costing Korean revenue. The blend is what hides it.

Everything else in this article is standard measurement discipline applied to a market with two front doors. That part is portable. The blending rule is the part Korea-specific programmes get wrong, and it is the one that quietly wastes a year of budget.

Start with the prompt panel, because nothing works without it

You cannot measure AI visibility by asking ChatGPT a few questions and screenshotting the good ones. Generative answers vary between sessions, across accounts and after every model update. A single run is an anecdote.

What replaces it is a version-controlled prompt panel: a fixed, documented set of real Korean buyer questions, grouped by funnel stage, re-run on a schedule, with results logged over time.

Practical shape. Somewhere between 100 and 300 prompts is a workable baseline for most categories, written in native Korean rather than translated from an English list, because a cross-engine audit in June 2026 found that identical commercial questions asked in a local language and in English returned vendor lists with zero overlap. Translate your prompt panel and you are measuring a market that does not exist.

Group them the way buying actually happens. Problem-stage questions where nobody has named a vendor yet. Category-comparison questions. Named-vendor questions where your brand or a competitor is already in the prompt. Objection questions, the ones where a buyer is looking for reasons to say no. Most brands over-index the third group because it flatters them, then wonder why pipeline does not move.

Version-control the panel itself. When a model update shifts your numbers, you need to know whether the market moved or your questions did. A panel that quietly gets edited between reporting cycles produces a trend line that means nothing.

The ten metrics, with their formulas

Definitions matter more than dashboards. Two agencies reporting "citation rate" can mean entirely different things, and in a category roughly a year old with no shared standard, they usually do. Write the denominator down.

Metric	Formula	What it is actually for
Answer presence rate	Prompts mentioning the brand ÷ eligible prompts	Baseline existence. The first number to move, and the least meaningful on its own
Citation rate	Prompts citing a controlled or earned source ÷ eligible prompts	Whether your own material is doing the work, or someone else is describing you
Share of answer	Brand mentions ÷ total tracked competitor mentions	Relative position. Requires a fixed competitor set defined before you start
Top-recommendation rate	Prompts ranking the brand first ÷ recommendation prompts	Being listed and being recommended are different outcomes
Fact-accuracy rate	Correct evaluated brand facts ÷ all evaluated brand facts	The metric that matters in regulated and high-consideration categories
Naver AI Briefing presence	Target Naver queries producing a brand citation ÷ tracked Naver queries	Its own line. Always. Never folded into a blended average
Cross-engine consistency	Agreement of key brand facts across tracked engines	Disagreement between engines usually means your entity data disagrees with itself
AI referral conversions	Conversions from sessions with AI-assistant referrers	Behaviour, not opinion. Volumes are small early, so read direction
Cost per citation gain	GEO spend ÷ net additional citations	Efficiency over time. Meaningless in month one, useful by month six
GEO-influenced pipeline	Pipeline value in KRW touched by AI-assisted discovery	The only number the board will remember. Influenced, not attributed, and say so

Two notes on honesty in that table.

"Eligible prompts" is doing real work as a denominator. Not every Naver query triggers an AI Briefing at all, and briefing coverage passed 20% of Naver searches by mid-December 2025 against Naver's public target of 40% by end-2026. Measuring your citation rate against queries that never produce an AI answer manufactures a low number. Measuring only against ones that do manufactures a high one. Define it, publish the definition, keep it fixed.

And "influenced" is not "attributed". Multi-touch attribution for AI-assisted discovery is genuinely unsolved, in Korea and everywhere. Someone reads an AI answer, does not click, searches your brand on Naver two days later, and converts through what looks like direct traffic. Anyone selling you clean AI attribution is selling you a model with assumptions baked in that they have not shown you.

Why Naver cannot be a share of the average

Worth being concrete about the failure mode, because it is not obvious until it costs money.

A blended visibility score treats every engine as interchangeable. In most markets that is a tolerable simplification. In Korea it is not, because the two doors draw from libraries that barely overlap. Roughly 70% of AI Briefing's citations are reported to come from Naver's own blogs and cafes rather than the open web, a figure worth treating as reported rather than audited, though Naver's behaviour corroborates the direction: it pays around 3,000 creators monthly through Naver Mate based on how often AI Briefing cites them, from a fund of roughly KRW 20 billion a year.

Which means your open-web GEO work and your Naver work succeed or fail independently. They are not two measurements of one thing. Averaging them is like averaging your revenue in two countries and calling it market position.

The practical test for any Korean GEO report: can you see, on one page, your AI Briefing number next to your ChatGPT number, without doing arithmetic? If not, the report is decoration.

Reporting in KRW, and what a Korean board actually asks

Report in the currency the budget was approved in. This sounds trivial and is not. A Korean CFO approving KRW 5,000,000 a month wants cost per citation gain in KRW and pipeline in KRW, and every conversion into dollars is a small tax on credibility.

Anchor the spend against something real. Korean GEO retainers run roughly KRW 300,000 to 3,000,000 monthly at entry, from around KRW 5,000,000 for professional full-service, and KRW 20,000,000 and up at enterprise, with most of those figures self-published by agencies rather than audited. Korea's online advertising spend was projected at approximately KRW 10.7204 trillion for 2025, growing 6.1%. Those two numbers together tell a board the discipline is small, new and attached to something large.

What they will ask, in roughly this order: what did we get for it, how do we know, what would we have got anyway, and what happens if we stop. The first three are answerable with the metrics above. The fourth one is not, honestly, because nobody has run a controlled Korean GEO holdout that has been published. Say that rather than inventing a counterfactual.

The thresholds, and why quarterly is the floor

Two numbers should be written into any Korean GEO plan as automatic triggers, so that rebalancing is a calculation rather than a negotiation.

Naver AI Briefing crossing 40% of queries. That is Naver's own stated target for end-2026, up from the 20% it passed in mid-December 2025. If it lands, the AI layer is answering two in five Korean searches and the Naver-side weighting has to rise with it.

Standalone Gemini or ChatGPT search use crossing roughly 60% in Korea. On OpenSurvey's tracking, ChatGPT reached 58.1% and Gemini 52.2% by July 2026, with Gemini having moved 23.3 percentage points in seven months and closed the gap from 25.6 points to 5.9. Both are within reach of that line inside a single planning cycle.

Which is the argument for quarterly review rather than annual. A budget split set in January against a market that moved 23 points by July is not a strategy, it is a memory. And note the shape of the Gemini move: it is simultaneously the fastest-growing engine, the highest-satisfaction engine at 77.3% against ChatGPT's 70.6%, and only fourth at 9.1% on most-frequent use. All three are true. A plan built on any one of them is built on a fragment.

What good looks like, and what it does not promise

A Korean GEO report that survives scrutiny has a few unglamorous properties. Fixed prompt panel, versioned. Published denominators. AI Briefing on its own row at every layer. KRW throughout. Fact-accuracy tracked alongside presence, not after it. Influenced pipeline labelled as influenced. And a section that says what did not work.

What it does not contain is a guarantee. Naver does not disclose AI Briefing's full ranking, extraction and citation logic, so nobody outside Naver can promise inclusion, and generative outputs shift with model updates in ways no vendor controls. What can be committed to is process: the panel, the cadence, the definitions, the entity work, and honest reporting of what moved.

That is the measurement architecture we run in our GEO practice for South Korea, and the reason we separate the Naver line by default rather than on request. Tessar Napitupulu works through the full multi-engine measurement design, including how to build a prompt panel that survives model updates, in Cited or Silent, available as a free gated edition, with retailer editions on Amazon, Google Play and Apple Books. Our own engine-by-engine tracking is published in the AI Citation Rate Report 2026.

---

Frequently Asked Questions

Why can't Naver AI Briefing be included in a blended visibility score?

Because the arithmetic hides the thing you need to see. A brand at 45% citation rate across ChatGPT, Gemini, Perplexity and Claude and 5% inside AI Briefing reports a blended 37%, which looks healthy while the Korean gap costs revenue. The two doors also draw from different libraries, with roughly 70% of AI Briefing's citations reported to come from Naver's own blogs and cafes, so open-web and Naver work succeed or fail independently. Averaging them is like averaging revenue across two countries and calling it market position.

How many prompts do we need in a Korean panel?

Between 100 and 300 is a workable baseline for most categories. What matters more than the count is that they are written in native Korean rather than translated from an English list, grouped by funnel stage including objection questions, and version-controlled so you can tell whether a change in your numbers came from the market or from your own edits.

Can we just translate our English prompt panel?

No, and this is one of the clearer findings available. A cross-engine audit on 26 June 2026 asked identical commercial questions in a local language and in English across four AI engines and three markets, and the vendor lists returned with zero overlap. Translating your panel means measuring a market that does not exist.

What is an "eligible prompt" and why does it matter?

It is your denominator, and it decides whether your citation rate looks good or bad. Not every Naver query triggers an AI Briefing at all, since briefing coverage passed 20% of Naver searches by mid-December 2025 against a 40% target for end-2026. Measuring against queries that never produce an AI answer manufactures a low number; measuring only against ones that do manufactures a high one. Define it, publish the definition, and keep it fixed across reporting cycles.

Can GEO results be attributed to revenue?

Influenced, yes. Attributed cleanly, no, and not in Korea or anywhere else. Someone reads an AI answer, does not click, searches your brand on Naver two days later and converts through what registers as direct traffic. Report GEO-influenced pipeline in KRW and label it as influenced rather than attributed. Any vendor offering clean AI attribution is offering a model with unshown assumptions.

When should we change our Korean budget split?

On two triggers, written into the plan in advance. If AI Briefing passes Naver's stated 40% of queries, the Naver-side weighting rises. If standalone Gemini or ChatGPT search use crosses roughly 60% in Korea, the global side rises. Review quarterly rather than annually, because Gemini moved 23.3 percentage points in seven months on OpenSurvey's tracking, closing the gap to ChatGPT from 25.6 points to 5.9.

What should a GEO vendor refuse to promise?

Guaranteed citations, guaranteed AI Briefing inclusion, and clean attribution. Naver does not disclose the full ranking, extraction and citation logic behind AI Briefing, generative outputs shift with model updates, and multi-touch attribution for AI-assisted discovery remains unsolved. Process can be committed to: a fixed prompt panel, published denominators, a set cadence, entity consistency work and honest reporting including what did not move.

---