30-Item GEO Checklist — From Question Design to Remeasurement

The difficult part of GEO analysis is not calculating metrics, but maintaining the correct order. If you measure before defining questions, you will not know what you are measuring; if you record before defining judgment criteria, two people will count the same responses differently. This document is a checklist that fixes that order into 30 items.

Detailed explanations are found in other documents in this series. We will not repeat them here but will link to them at the necessary points.

The Role of this Checklist and Detailed Explanations Document

| Coverage | What it covers |

| --- | --- | | This document | In what order is the analysis run — 30 items and why each is needed |

| How is AI Search Visibility Determined | What determines mentions, citations, and recommendations |

| How is AI Citation Rate Measured | Measurement environment and response collection procedures by platform |

| 5 types of GEO Performance Metrics | Numerator and denominator of metrics and sample limits |

| 25 types of Website Audit for AI Search Visibility | Technical checklist items such as robots.txt, canonical, sitemap, and JSON-LD |

| What is GEO | The concept and its relationship to SEO |

We have divided them into seven groups. For each item, we have included the reason why it is necessary. If you cannot understand the reason for an item, remove it to fit the project, but please record the reason for the removal.

① Analysis Goal (Before Measurement)

• 1. Have you written the business goal to be verified in this analysis in a single sentence? Without a goal, there is no standard to judge which questions to include or exclude.

• 2. Have you determined a single core conversion among Inquiry, Comparison, Recommendation, and Brand Understanding? If the conversion is different, the linked page and evaluation metrics will differ.

② Question Design (Before Measurement)

• 3. Have you included definition, comparison, recommendation, and problem-solving questions separately? If questions are concentrated in one type, only the performance of a specific stage will be visible. - 4. Did you attach versions to the question set and keep a change history? It is impossible to distinguish whether an increase in the month the questions were changed is due to performance or because the questions became easier.

• 5. Did you include questions that do not contain brand names? Measuring only branded questions reveals only the perspective of users who are already familiar with the brand.

③ Measurement Conditions (Before Measurement)

• 6. Did you record the platform and model/service version? Even for the same question, the method of citing sources changes if the platform changes.

• 7. Did you record whether web search or search mode was used? Responses with search turned off lack sources, distorting the denominator of the citation rate.

• 8. Was it measured in an independent conversation or session? If previous conversations remain, their context gets mixed into the answers.

- 9. Did you keep error/abruption responses in a separate state instead of counting them as unexposed?** If you include errors in the denominator, the value will be lower than the actual result.

④ Judgment (Measurement in Progress)

• 10. Did you record brand mentions and source citations in different columns? Including a name and having our document used as the basis are different outcomes.

• 11. Did you distinguish between recommendation inclusion and simple mentions? Only questions with a recommendation intent should be included in the denominator of the recommendation inclusion rate.

• 12. Did you save the cited URL exactly as is? If you do not know which page is being cited, you cannot determine what to reinforce next.

• 13. Did you verify that the cited URL actually opens? Non-existent links generated by the model may be mixed in.

• 14. Did you record unmentioned and uncited results in the same format? If you only collect successes, the basis for determining reinforcement priorities disappears.

⑤ Competitors/Sources (Under Measurement)

• 15. Did you record the order of brands that appeared together in the same response? The position in which they are introduced, rather than just whether they appear, reveals the competitive landscape.

• 16. Did you classify the domains and source types that were cited repeatedly? The response varies depending on whether official sites, media, or communities are used as the basis.

• 17. Did you list the informational differences between competitor documents and your own documents as items? Observing that "competitors appear frequently" alone does not yield results.

⑥ Sites/Content (After Measurement)

• 18. Is the official brand name and service name consistent across pages? If the notation is inconsistent, it is difficult to recognize them as belonging to the same entity.

• 19. Does an official page actually exist that connects to the core question? If there is no page, new content comes first, not technical enhancements.

• 20. Does that page respond to 200 and is it indexable? Inaccessible documents do not serve as candidates for sources.

• 21. Do the title, description, and canonical points to the representative URL? Duplicate URLs disperse citations.

• 22. Does the structured data match what is visible on the screen? Google explicitly states that you should not mark up content that is not visible on the screen.

• 23. Are the author, publication date, modification date, and source visible on the screen? If it is not confirmed who wrote the document and when, it is difficult to use it as evidence.

• 24. Is the core information in the HTML text rather than in the image? Information that does not exist as text cannot be cited.

• 25. Is the same manuscript duplicated across multiple domains? If the authenticity of the original text is compromised, it is impossible to control which URL will be cited.

⑦ Execution & Re-measurement (After Measurement)

• 26. Was the work order determined by impact and execution difficulty? The criterion is connection to important questions, not the number of errors.

• 27. Was the person responsible for the correction and the completion date recorded? Without a specific time of change, it cannot be linked to changes in the results.

• 28. Was re-measurement performed after confirming that the corrections were reflected in the index? Re-measurement before indexing measures the time lag, not the effectiveness of the work.

• 29. Was it measured again using the same question and conditions? If the conditions change, a before-and-after comparison is not valid.

• 30. Were newly cited URLs, disappeared results, and questions that are still not displayed listed separately? Each of these leads to a different subsequent step.

Items 10 through 11 require establishing judgment criteria in writing first. If two people judge the same response differently, the difference between judges is reflected more significantly than changes over time.

Minimum definitions of the three judgments

The reason the three judgments are not combined into a single 'exposure rate' is that they lead to different tasks. If there is a mention but no citation, it means there is no document in a format suitable for citation; if there is a citation but no mention, it means the connection between the document and the brand is weak. The definitions of the numerator and denominator of the metric are covered in GEO Performance Indicators 5.

This is a form that combines items 6 through 14 into a single line. It defines what to put in each column, rather than the values. Copy and paste this exactly as it appears into the spreadsheet columns.

Response 1 = 1 Row Recording Format

Also verify the access conditions of the platform being measured. OpenAI separates search display (OAI-SearchBot) and training (GPTBot), and states that robots.txt rules may not apply to user-initiated queries (ChatGPT-User). Perplexity also operates PerplexityBot for indexing and Perplexity-User for user request queries separately. When measurement results are low, you must first determine whether the issue is a content problem or an access problem. (6, 7)

Once the table is filled, do not look at the values ​​one by one, but read them as combinations. Combinations do not provide a definitive answer to the cause, but rather determine what to check first.

Observation Combinations and Next Verification Order

The first thing to fix is ​​not the items with the most errors, but the gaps identified in questions connected to business objectives. Address high-impact and easy-to-implement items first (standardizing brand notation, meta information, internal links, indicating author and modification date), and set building source data or expanding external channels as long-term tasks by assigning specific responsibilities and deadlines.

This is the flow for items 26 through 30. It follows the sequence: Baseline Measurement → Reinforcement → Check Index → ​​Remeasure Under Same Conditions → Change Log.

Records to Keep by Step

Google impressions can also be checked via the Generative AI performance report in Search Console. However, this report only provides the number of impressions where site links are displayed in AI Overviews and AI Mode, and does not include clicks, click-through rates, or average ranking. The cited context can still only be known by saving the original response text. (5)

What checks 18·22·23 ultimately require is to make it possible to verify on the screen 'who wrote it, when, and on what basis.' However, if you treat E-E-A-T as if it were a goal to be achieved, you lose direction.

Google presents this document as a collection of questions to help you evaluate your own content. It is not a system where you raise your score by adding tags. Therefore, what you leave behind is a record, not a mark.

Record Items Based on Trust

There are no separate requirements for AI feature exposure. Google states, "There are no additional requirements for appearing in AI Overviews or AI Mode, and no special optimization is needed." Check 22 requires not adding more dedicated schemas, but verifying that what has already been added matches the screen. (3)

The purpose of the checklist is not to pass all 30 items. It is to identify which questions do not show the brand, which documents are repeatedly used as evidence, and what is missing from the company, and then determine the next steps.

Most errors disappear simply by following the correct order. First, fix the questions and conditions (1–9), record the judgments without combining them (10–17), compare the results with the site status (18–25), and then verify again under the same conditions (26–30).

Even if numbers are present, display the numerator, denominator, and conditions together, and separate observations from inferences. The observation is "professional media are repeatedly cited," and the inference is "there is a need to reinforce the basis of the company definition document." If you write both in the same column, the next person cannot verify them.

SUMMITFEED operates question design, answer measurement, competitor and source analysis, site structuring, content publishing, and remeasurement as a single workflow. The items to check when reviewing an agency are listed separately in Eight Criteria for Choosing a GEO Agency.