How GEO Times measures and reports

Put the same question to a generative AI and you will not get the same answer every time. A company recommended on ChatGPT may not appear on Gemini, and asking again a few days later can produce a different result.

That is why GEO Times does not conclude from a single answer that a particular company is well represented.

Our measured articles are based on the results of repeating the same question many times. When we cover overseas reporting or company announcements, we check the original and keep what we could not verify separate from what the editorial team measured.

A single appearance is never a ranking

The first thing we settle in an AI search measurement is the question.

We choose a question a real user would type, then repeat it under the same conditions without changing the wording. Each question is measured at least 5 times per AI engine, each run starting a new conversation by default.

During measurement we record the date, the AI engine, the number of runs, and the companies and cited sources that appeared in the answers. The article also publishes the exact question used.

For example, a company that appeared in 3 of 5 ChatGPT runs is recorded as "appeared in 3 of 5 runs". We do not pick out one appearance and describe that company as the most recommended.

Every company that appears is recorded

We do not re-select results to fit the direction of an article.

An unexpected company is recorded, and if a company we are covering did not appear even once, we say so. We do not exclude results because they favor or disfavor a particular company.

When an AI presents web pages as the basis for its answer, we check those sources too. The number of times a company's name appears and the number of times it is actually cited as a source are treated as different values.

Raw answers are archived separately.

Vendor-reported figures and verified figures are kept apart

GEO vendors publish AI visibility rates, citation rates and client cases on their websites and in press releases.

We may use such material in articles, but we never treat it as equivalent to what GEO Times measured. Material published by a company is labeled with its source, such as "company-reported" or "self-tallied".

Vendor comparisons are likewise based on publicly verifiable material and our own measurement records. Where a ranking or recommendation appears, the article states what was compared.

Unverifiable results and figures of unclear origin are not used as grounds for comparison.

Overseas stories start from the original

When we cover announcements from Google, OpenAI, Perplexity and other overseas companies, or reports in overseas media, we check the original.

Rather than translating the original verbatim, we summarize what was announced and add what matters for AI search and the GEO market in Korea. The original source is shown so readers can check it themselves.

For research, we go to the paper or the research institution's material first. We avoid re-reporting another outlet's summary wherever possible.

AI may draft, but people verify

Generative AI may be used for translation or organizing material during writing.

But no AI-generated summary or translation is published without verification. The editorial team compares it with the original, rechecks figures, company names, dates and quotations, and writes the final copy.

Whether or not AI tools were used, responsibility for a published article rests with the GEO Times editorial team.

When results change, earlier records stay

AI search results can change over time.

A company that appeared often in last month's measurement may not appear this month. In that case we do not edit the earlier article to match the current result. The earlier measurement stays as a record of its time, and the new result becomes a separate article.

Factual errors are different. Typos, wrong figures and source errors are corrected after verification, and the article notes what was corrected and when.

Changes to how GEO Times measures or writes are published with the date in the changelog.

Correction requests and questions about our methods go to [email protected].