Generative engine optimization and llms.txt without the hype

Generative engine optimization, often shortened to GEO, is the current label for making a site work in AI search. Most of it is ordinary technical SEO plus honest evidence about what you can and cannot prove. The useful reports remove technical blockers, improve the page for people, or make evidence easier to verify. They turn dangerous when an optional file or markup pattern gets sold as a citation score.

The reports keep those two jobs separate.

Reuse fresh page evidence instead of crawling twice

seo crawl --project example --save
seo ai-readiness --project example
seo entity-readiness --project example
seo llms audit --project example

These reports reuse a saved crawl. Run a fresh crawl when the site has changed or the saved pages no longer represent what is live.

Start with ai-readiness for technical eligibility and page observations. Use entity-readiness for naming, authorship, structured data, and linked identity evidence. Run llms audit only when you care about the optional file for a system that consumes it.

Fix ordinary search blockers before AI-specific experiments

Failed responses, blocked crawling, noindex, a conflicting canonical, missing main content, and restrictive snippet controls can remove or limit the content a search feature can use.

Google’s current guidance for generative AI features says a page must be indexed and eligible for a Search snippet to be eligible for AI Overviews and AI Mode. Google also says meeting the requirements does not guarantee crawling, indexing, serving, or selection.

That gives the technical report a clear job. It can find evidence that blocks eligibility or deserves verification. It cannot predict a mention.

Use page evidence to improve the source itself

The crawler records visible content structure, semantic HTML, authorship, dates, question headings, lists, tables, entity links, structured data, and media evidence. Those observations can help an agent inspect whether a page:

• answers the user’s actual question with specific, checkable information;
• makes the responsible person or organisation clear where that matters;
• keeps names, dates, facts, and linked identity signals consistent;
• uses structure that helps people scan and assistive technology navigate;
• contains original evidence rather than a rewrite of the same generic advice.

No single item in that list is a universal ranking requirement. Google advises site owners to create helpful, reliable, people-first content and bring first-hand knowledge or original analysis where it fits. Its people-first content guidance is a better editorial reference than a made-up “AI-ready” word count.

Treat structured data as eligibility evidence

Structured data can describe page entities and make supported content eligible for rich results. The markup must match the visible page and include the required properties for the chosen search feature.

Google’s structured data introduction does not promise a rich result after valid markup. Google also says there is no special schema required for its generative AI search features.

The reports therefore record detected types, missing or invalid fields, and page mismatches as evidence to inspect. They do not convert schema coverage into an AI visibility score.

Use llms.txt only for systems that consume it

seo llms generate --project example --output llms.txt
seo okf export --project example --output ./okf
seo okf validate ./okf
seo export knowledge --project example --format markdown --output knowledge.md

These exports can help an agent or retrieval system that explicitly reads them. They can also give you a concise inventory of the pages and concepts in a saved crawl.

Google’s generative AI search guidance says Google Search does not use llms.txt or special AI text files. Publishing one neither helps nor harms Google Search visibility according to that guidance. Maintain the file only when it serves a real consumer. Otherwise it is another stale index waiting to happen.

Respect snippet controls as intentional site policy

Google applies controls such as nosnippet, max-snippet, and data-nosnippet to its search appearances, including AI Overviews and AI Mode. Its robots meta and snippet specification defines the current behavior.

The crawler should report those controls. It should not tell you to remove them without understanding why they exist. A publisher may have legal, licensing, privacy, or product reasons to limit snippets.

Measure referrals without pretending they cover every AI visit

seo ai-referrals --project example

The GA4 report groups measurable visits from known AI referral sources. It can show a real landing page, source, and session count when GA4 received them.

Referrer stripping, apps, redirects, privacy controls, and changing source domains make that evidence incomplete. “No retained AI referrals” means the query found none in the available GA4 rows. It does not prove that no person found the site through an AI product.

Google now directs site owners to its own Search Console reporting for visibility in Google generative AI features. Other providers expose different or no first-party reporting, so keep source-specific evidence separate.

Give an agent a claim it can defend

Ask for findings in this shape:

1. The exact page evidence or provider row observed.
2. The eligibility issue or optional observation derived from it.
3. The limit on what that evidence can establish.
4. A bounded change and a way to verify it.

The agent workflow guide covers that evidence discipline. Use the crawler guide when a finding depends on robots, canonicals, rendered HTML, or structured data.