GEO and AEO: a practical audit for AI search visibility
Neaptide · July 28, 2026 · 5 min read · Updated: October 3, 2026
Audit crawler access, answer quality, evidence, translations and measurement. A practical GEO/AEO checklist with an example and clear acceptance criteria.
On this page
GEO usually means working on how visible a site is in generative answers. AEO is the broader label and takes in answer-oriented search and assistants as well. The two terms overlap, and neither is a technical standard. In practice the job is the same: the relevant information has to be findable and understood correctly.
Choose a page and a customer question
Take a real question such as “Can I move my customer database without losing its history?” Open the migration page and try to answer using only the published text. Which file formats and fields are supported? What cannot be transferred? Who checks the result? Write down anything you would still need to ask a salesperson. Those gaps are your first editing task. Markup will not supply facts the page does not contain.
Check technical access
- Open the public URL without login. Check errors, redirects and incorrect canonicals.
- Inspect indexing and display restrictions in the relevant search engine’s tools.
- Check the main text in the fetched response and, for JavaScript-heavy sites, in the rendered page.
- Review robots.txt, CDN and server blocks. Allow verified crawlers as documented rather than disabling protection.
OAI-SearchBot and PerplexityBot serve search. GPTBot collects training data, and whether to allow it is a separate decision. A blanket rule for all AI bots ignores that difference, so follow each provider’s documentation.
Make the answer complete
Lead with the answer, then explain conditions, examples and exceptions. How long that takes depends on the question, and nobody needs a complex explanation squeezed into 40–60 words to please an assumed algorithm.
Example: importing customer data
‘Fast migration without data loss’ gives a reader little to assess. Say which formats you accept, which fields transfer, what is not supported, how duplicates are handled and who checks the result, using your product’s actual limits. An anonymised sample file and a few validation steps help, as do links to the relevant pricing and terms.
Check facts, responsibility and translations
- Identify the author or responsible team and cite technical sources.
- Give case-study figures a period, sample size and measurement method. Separate observation from explanation.
- Change the update date when the content materially changes, not merely to look recent.
- Review each translation for meaning, currency and regional service conditions. hreflang tells systems that support it which pages are equivalent language or regional versions. It does not determine which language version an AI service will cite.
Markup, FAQs and llms.txt
Structured data gives systems a machine-readable description of what is on the page, such as an article’s author and dates. It must match the visible content. Adding more schema types does not make the information more useful, and Google does not require special GEO markup or llms.txt for its generative search features.
Use FAQs when readers still have questions after the main text. According to Google’s changelog, FAQ rich results stopped appearing in May 2026. A good Q&A section still helps readers. It just comes with no promise of a special search display or an AI citation.
Accept the work and measure changes
- Save the original question, page state, service answer and cited URLs.
- After publication, check access, new links, mobile layout and agreement between markup and text.
- Repeat the same questions under comparable settings on different days. Keep answers without citations too.
- Track visibility, visits and enquiries separately, noting other changes that happened during the same period.
The audit should identify missing information and access problems, and implementing its recommendations should resolve them. If mentions go up later, that is an observation. A single before-and-after test cannot establish what caused the rise or predict sales.