How to Implement Generative Engine Optimization: A Practical Plan

Choose one buyer question and one page

To implement generative engine optimization, improve one useful page around a buyer question, record the change and review comparable data. This plan focuses on Google's AI search features. The deliverable is a better answer and a documented decision, not a guaranteed citation.

Choose a question heard in enquiries or visible in Search Console, then identify the page that should answer it. Keep related wordings together instead of creating competing pages. Set an effort cap before starting.

Illustrative service-business example, not a client case: a Zagreb office-cleaning business could answer 'What does an office cleaning quote include?' on its existing Croatian quote guide. The intended action is a relevant quote request.

• Acceptance check: record one question, market, language, target URL, intended action and effort cap.

Save the baseline and clear eligibility blockers

Export the page's Search Console Web performance before editing. Save dates, property, page, country, device and any query filter alongside impressions, clicks and average position. Keep an unfiltered page view as well as any query subset. Record qualified enquiries separately if tracking exists. If the page is new, label its baseline unavailable; do not substitute sitewide figures as if they belonged to it.

Inspect the indexed URL and canonical. Check the live response, Googlebot access, rendered answer text, indexing directives and snippet controls. In Search Console, check the effective Search generative AI inclusion setting, including inherited choices. Repair blockers first and date those repairs separately. Being indexed and snippet-eligible, with inclusion enabled, does not guarantee selection.

• Acceptance check: baseline export saved; HTTP 200, crawl access, visible answer, canonical, index status and snippet controls checked; inclusion setting recorded.

Replace generic copy with checkable evidence

Draft the direct answer, then explain what changes it. In the cleaning example, distinguish routine cleaning from optional tasks, state what information a quote needs and explain exclusions. Do not invent a price, turnaround time or customer outcome to make the answer sound specific.

Add something the business can substantiate, such as its actual site-visit checklist or an anonymised scope example used with permission. Explain where the material came from and its limits. Identify the real author or reviewer and link to relevant background. Check service areas, inclusions and contact details against current business records.

• Acceptance check: a buyer can identify the answer, verify one original supporting item and see who checked the facts.

Connect the page and describe it accurately

Link to the pilot page from a relevant service page and an existing related guide, where useful. Use descriptive link text and confirm the links reach the canonical destination. Give the answer page a clear route to the intended enquiry action.

Review existing structured data rather than installing an AI schema package. For an article, check that its author and dates match the visible page. Use only markup appropriate to the content. Google requires neither special AI markup nor llms.txt for these features. The downloadable Markdown record is a working document for people, not an eligibility mechanism.

• Acceptance check: contextual links work, the enquiry route works on mobile and any structured data matches visible facts.

Publish a limited change with a release record

Save the previous version and record exactly what changed: answer, evidence, authorship, links and markup. Name one accountable owner, whether the business owner or a specialist such as Koupoli's Karlo Ridan. Separate technical repairs from content improvements in the log.

After publishing, repeat the live-page checks and record the release date. Later, note when URL Inspection shows a crawl after the release. Choose the comparison windows and review date now. A 28-day checkpoint is a planning choice, not Google's promised response time. Avoid unrelated rewrites during the pilot; log unavoidable changes. The implementation can pass its acceptance checks even when visibility data remains inconclusive. Do not make a citation the condition for completing useful page work.

• Acceptance check: the previous version, change log, owner, live checks and scheduled review are saved in the pilot record.

Compare like with like and make one decision

Compare equal complete windows using the same property, page, country, device, search type and query-filter rules. Review overall Web impressions and clicks, plus qualified enquiries separately. Where available, record the page's Generative AI report impressions with matching supported filters. Those impressions already sit within Web totals; do not add them together.

Do not infer AI citations from ordinary query impressions or average position. The dedicated AI report has no query dimension. A missing report may reflect insufficient impressions. Search Console also omits anonymised query rows, so a handful of observations cannot establish success or failure.

Choose one decision: fix a blocker, keep the improvement, extend observation or revise the answer. Note recrawling, demand changes and other releases as possible explanations. A before-and-after difference alone does not prove the pilot caused it.