Generative engine optimization: a practical website audit.

Choose a question that affects a real buying or implementation decision. “How can I stop an AI workflow acting without approval?” needs an explanation, a tested example and clear limitations. A page that only defines artificial intelligence does not resolve that decision.

Use customer conversations and your own search queries as starting evidence. Group variants that require the same answer on one page. Create a separate page only when the reader needs different instructions, evidence or a next step.

For Marsen, the human approval workflow guide serves the design question, while a service discussion establishes which systems and permissions an actual implementation needs. Keep the educational answer useful even for someone who does not contact the business.

A CMS owner can usually improve explanations, sources and links without changing the application. Start with one existing page rather than creating a new collection of pages.

1. State the answer. Use a descriptive title and opening paragraph that match the question. Explain who the answer applies to.
2. Make the business identifiable. Keep the trading name, services, contact information and genuine profiles consistent. Do not invent an office, client, review or certification.
3. Add evidence. Publish an approved example, reproducible calculation or first-hand test. State the date, method and limits. Link factual platform claims to the primary documentation.
4. Connect the page. Add relevant links from a product page and a related guide. Use text that explains the destination rather than “click here.”
5. Preview and check. Open the page on a phone, follow its links and verify that its promised download works.
6. Record the change. Save the URL, actual update date, question and evidence added. Review equivalent reporting periods later.

Ask an engineer to investigate an unexpected redirect, a blocked page, conflicting canonicals, missing server-rendered text or a broken form. Changing visible copy cannot repair those problems. A plugin's green indicator is not confirmation of Google indexing.

On 28 September 2026 we fetched every URL declared in Marsen's production sitemap and inspected the returned HTML. This was a public HTTP and HTML check, not a Googlebot crawl, ranking experiment or answer-engine citation test.

The practical finding was to distinguish technical access from content usefulness. The enquiry-receipt pilot had only one inbound page in the crawl. Adding relevant links from CRM and workflow content would help readers find the example. The original GEO guide explained principles but needed an inspectable worked example.

Download the dated public observations. Counts describe that snapshot and may differ from the current site. No private Search Console queries, customer details or conversion records are included.

First fetch the URL and confirm it resolves to the intended page. Then inspect robots controls, the canonical URL and the useful text returned in the HTML. Finally use Search Console's URL Inspection on your own property to see Google's recorded state. These are separate checks.

A sitemap supports discovery; it does not force indexing. A successful HTTP response says the server answered that request. It does not prove Google selected the same canonical, processed the newest copy or chose to show the page.

Google's AI optimization guidance ties generative search to ordinary SEO foundations and original useful content. It also describes Search Console inclusion for Google's generative features. Check the current documentation and your property's report rather than assuming a crawler permission creates eligibility everywhere.

There is no special AI schema to add. Keep structured data accurate and aligned with visible content. Google says it does not use llms.txt to improve Search rankings; maintaining it for another system is a separate choice.

Give an important claim a method and a boundary. “Automation saves time” is difficult to verify. “One synthetic request and one replay produced one CRM receipt in this dated test” is a smaller but inspectable claim. It establishes the tested behaviour, not customer revenue or reduced staff time.

The Marsen enquiry-receipt pilot shows that distinction. A separate downloadable approval fixture lets a reader inspect approve, reject, expiry and replay handling locally. Neither should be described as a customer deployment.

For a client case study, record the baseline, implementation, dates, measurement method, outcome and remaining limitations. Obtain permission before publishing client information. If no result was measured, explain the delivered work without inventing an improvement.

Use a stable set of questions and record the engine, model or search mode when visible, date, language and locale. Preserve the answer and exact cited URLs. Different interfaces can retrieve different sources; an API response does not establish what a consumer interface shows.

Download the observation worksheet. A blank outcome means unmeasured, not zero. If you report a citation rate, divide prompts citing the owned domain by completed prompts in the stated sample. Report failed or unavailable runs separately so a changing denominator does not create false progress.

Repeat the same questions before drawing a trend. A new citation after an edit does not prove the edit caused it. Compare relevant page/query pairs and received enquiries alongside the answer sample, using the measurement guide.

1. Select a useful existing page with observed demand or a documented customer question.
2. Fix retrieval problems and confirm its intended canonical.
3. Add the missing example, limitation or source.
4. Link it from the pages where readers need that answer.
5. Verify the live page and inspect indexing after publication.
6. Review comparable periods and repeat the fixed answer sample.

Use the AI visibility audit guide for the starting review and Marsen's GEO service when you need help carrying out an agreed scope. Prioritize a useful answer and evidence over page count.