How to evaluate GEO results
Use a three-part review table and sample records to separate completed work, platform performance and genuine enquiries. Learn what evidence you can verify and what a single AI answer cannot prove.
To evaluate GEO, first check whether the agreed content and technical work is complete, then review platform performance and business feedback. Record these three areas separately so you can decide whether to investigate implementation, platform performance or enquiry handling next.
A screenshot of a brand mention shows what appeared in one answer. On its own, it cannot establish consistent recommendations, actual visits or new customers. Agreeing how to keep records before work starts is more useful than arguing afterwards about whether GEO worked.
Agree on three levels of review
Define the target market, language, products and pages first. Then document the buyer questions to address, the planned changes, who is responsible and what will be delivered. Save the original pages and existing data as a baseline for later comparison. Mark any missing records explicitly.
Swipe sideways to see the full table
| Review area | What to check | Evidence to keep |
|---|---|---|
| Completed work | Whether the agreed page, content and technical changes are complete, and whether the information and contact options are usable. | Task list, pages before and after changes, source materials, approvals and records of actual checks. |
| Platform performance | Whether the brand is mentioned, cited or recommended for the agreed questions and conditions, and whether descriptions are accurate. | Full questions, answers and source links, sampling conditions, total attempts and records of unusual results. |
| Business feedback | Whether relevant visits and genuine requests have occurred, and whether they have led to meaningful conversations. | Available traffic sources, actual enquiries and follow-up records. Label unknown sources separately. |
Content verification, page changes and supporting checks can be reviewed against the agreed tasks. Whether platforms use the content, buyers enquire or sales follow also depends on other factors. Google explicitly states that meeting its requirements does not guarantee crawling, indexing or serving. This guidance applies to Google and should not be treated as a rule for other AI platforms.
Separate mentions, citations, recommendations and enquiries
Define how each metric is counted and use that definition consistently. An answer may include both a brand name and a link, or cite a source without recommending a purchase. Record each signal separately:
- Mention: does the answer explicitly name the brand? Record whether the description is accurate so that a negative or incorrect description is not counted as a positive recommendation.
- Citation: does the answer identify a source? Note whether it cites your own page, a third-party page or other material.
- Recommendation: does the answer present the brand or product as an option for a particular need and explain when it is suitable?
- Visit: can your existing analytics identify a visit from a relevant source? Keep the source marked as unknown when it is unavailable.
- Genuine enquiry: has someone actually contacted you with a relevant product or service need that can be discussed further? Copying an email address, opening a QR code or clicking a contact button is not the same as receiving an enquiry.
Group brand-name searches separately from buying questions that do not name a brand. When reporting a percentage, keep the numerator, total sample size and counting unit, such as individual answers or distinct questions. Do not select only answers that mention the brand, or add together platform metrics that use different definitions.
If you use Google Search Console, its impressions, clicks and average position are Google Search reporting metrics. Traffic from AI Overviews and AI Mode is included in overall Web search data. This is not a separate measure of all AI platforms, and it does not directly represent sales.
Keep records that can be compared and checked
You can start by manually recording a small set of important questions in a spreadsheet without a paid tool. Keep the original evidence and continue observing a question set chosen in advance. The table below provides a starting point for the fields to record; it does not require a separate system.
Swipe sideways to see the full table
| Field | What to record | What to verify |
|---|---|---|
| Business scope | Market, language, product, target page and changes made in this round. | Keep the original state and review different markets or pages separately. |
| Question set | Question ID, exact wording and branded or non-branded group. | Reuse a fixed set. Record the version when questions are added or rewritten. |
| Sampling conditions | Platform, visible model information, date, time zone and number of tests. | Record login status, region settings, whether the conversation is new and any preceding context. |
| Results and denominator | Mentions, citations, recommendations and accuracy for each answer; total attempts and valid answers. | Keep answers without mentions, failures and unusual results, not just successful screenshots. |
| Original evidence | Full answers, screenshots or accessible records, plus source links. | Make each sample traceable. Mark information the platform does not display as unknown. |
| Business feedback | Identifiable visits, genuine enquiries, sources and follow-up status. | Distinguish analytics data, customer-reported sources and unknown sources. Avoid double counting. |
To calculate a mention rate by answer, divide the number of valid answers that name the brand by the total number of valid answers. Report total attempts and failures separately. Do not automatically count timeouts or errors as answers without a mention. Define inclusion rules in advance and apply them consistently across comparison periods.
Consistent conditions help comparison, but they cannot remove answer variability, platform updates or the effects of other marketing activity. Explain changes to the question set, sample size or test conditions separately. With small samples, report what you observed rather than drawing conclusions about the entire market or claiming that a particular change caused growth.
Example: match each conclusion to its evidence
This fictional product-page scenario shows how to report incomplete information. Clear labels such as “checked”, “needs further observation” and “source unknown” make it easier to choose the next step than combining everything into one performance score.
Illustrative record, not client results
Suppose the English product page for a Model A folding camping table has been updated with storage guidance, conditions of use and enquiry information. All statuses below are illustrative. In a real record, include the actual check time, platform, question set and original evidence.
A conclusion without enough evidence
The product page has been optimised, the brand appeared in an AI answer, and GEO is working.
Ask what the evidence actually covers
- What changed on the page, and who approved the product information?
- Which question and answer included the brand, and what proportion of all samples does that represent?
- Were any relevant enquiries actually received, and can their sources be confirmed?
Check the implementation evidence
Match each task to the updated page, supporting materials and check results. Mark supported items as checked. Keep items awaiting approval or checks on the outstanding task list.
Review platform and business signals separately
Place a single brand mention in the context of the full sample record. Without baseline data or consistent conditions, leave trends undecided. Do not automatically attribute enquiries to GEO when their sources are unclear.
A clearer way to record the result
Task: the English product page for Model A. Implementation: storage guidance, conditions of use, source materials and the enquiry option have been checked (illustrative status). Platform performance: one brand mention is recorded. There is no comparable baseline, so consistent performance needs further observation and no improvement percentage is calculated. Business feedback: whether any genuine enquiries were received remains unconfirmed, and their sources are unknown. This does not establish new sales opportunities. Next step: complete the sampling conditions and original records, continue checking the fixed questions, and record actual enquiries separately.
Is the record clear enough?
- Each conclusion has supporting material. Missing information is not described as completed work or a zero result.
- Individual observations, trends and business feedback are kept separate. Correlated changes are not presented as proven cause and effect.
- The next step identifies a missing record or a specific check, rather than simply calling for more articles.
This example demonstrates report writing, not actual measurement. Real records should retain only the information needed for review, without collecting unrelated contact details or private customer information.
Use the review to decide what to do next
If work is incomplete, address the identified gaps first. If the work is complete but platform records are insufficient, gather observations that can be compared. When relevant visits occur but enquiry outcomes are unclear, check whether the page matches buyer needs and whether contact options work smoothly. Do not immediately conclude that the traffic has no value.
When considering further investment, review completed work, current evidence and remaining uncertainties together. Bring the target market, page URLs and records you can share to the discussion. Identify the next problem to solve before adding more tasks.
