A diagnostic that ends in "it depends" is not finished. This guide shows how to convert a vague AI-search concern into a small test whose result you can read, argue about, and hand off — using only what publishers officially document.
Start from a single answerable question
Before writing any test, write the question it answers in one sentence. "Why aren't we in AI Overviews?" is not testable. "Does our key page pass the technical eligibility Google describes for AI Overviews and AI Mode?" is. The second question has an official answer you can check, which is the whole point.
evidence label · s1 Google states that to be eligible as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown with a snippet, meeting the Search technical requirements — and that there are no additional technical requirements for these features (developers.google.com/search/docs/appearance/ai-features). Google also states that same best practices for SEO remain relevant and no special optimizations are necessary. So a legitimate first test is eligibility, not secret tactics.
Make the test falsifiable
A useful small test has four parts: a stated claim ("our page is indexable and snippet-eligible"), a check that could contradict it (the page is blocked from indexing or snippet display), the evidence you will collect, and the date you will judge it. Write all four down before you look. That ordering is what separates investigation from retrofitting a story to whatever you find.
Keep the test small on purpose. One page, one question, one check. A large multi-variable "AI visibility audit" can produce a score with no clear next action. A single eligibility check produces a yes, a no, or a "can't tell yet" — all three are usable.
Separate access checks from content checks
Access and content are different hypotheses and should be tested separately. On the access side, Google documents that robots.txt directives for Googlebot are the control site owners have for managing how their sites are crawled for Search, and that nosnippet, data-nosnippet, max-snippet, or noindex can limit what is shown from pages (developers.google.com/search/docs/appearance/ai-features).
Google is also explicit about a trap many teams fall into: a robots.txt file tells crawlers which URLs they can access and is used mainly to avoid overloading a site — it "is not a mechanism for keeping a web page out of Google" (developers.google.com/search/docs/crawling-indexing/robots/intro). A disallowed page can still be indexed if other sites link to it, with the URL and potentially anchor text still appearing in results. So "we blocked it in robots.txt, that should fix it" is a hypothesis Google's own documentation contradicts. Test it against that.
Treat AI-crawler settings as independent switches
If the claim is about ChatGPT search visibility rather than Google, the switches are documented separately. OpenAI states it uses OAI-SearchBot and GPTBot robots.txt tags, and that each setting is independent — you can allow OAI-SearchBot for search while disallowing GPTBot for training (platform.openai.com/docs/bots). It states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links, and that adjustment can take about 24 hours from a robots.txt update.
A clean test here checks one user agent at a time. Note that ChatGPT-User is used for user-initiated actions rather than automatic crawling, and OpenAI directs site owners to OAI-SearchBot for managing search opt-outs. Mixing the three into one "we blocked AI" change makes the result uninterpretable — you won't know which setting produced the outcome.
When content is the hypothesis
If access checks come back clean, the next small test is about content quality, not cleverness. Google's helpful-content guidance frames this as three questions — who created it, how, and why — and states that trust is the most important aspect of E-E-A-T, while noting E-E-A-T itself is not a specific ranking factor (developers.google.com/search/docs/fundamentals/creating-helpful-content). It also states that fabricating creator profiles, such as AI-generated headshots, made-up names, or false credentials, is a form of deception and a signal of low quality.
That last point constrains your test design: you cannot improve a content hypothesis by inventing an author. A defensible small content test asks a specific question — for example, does this page clearly answer who created it and why — and records the finding. Inference to flag: whether any given content change affects AI-feature inclusion is not guaranteed by any source here; Google explicitly says indexing and serving are not guaranteed even when requirements are met.
Record a result you can read
At the end, write the outcome in three lines: what you predicted, what you observed, and what the observation rules in or out. If the result is ambiguous, say so — ambiguity is a finding, not a failure. Do not upgrade a single test into a general claim about "what AI search rewards"; the sources describe individual systems from a single point in time, with tentative language throughout.
Use the private diagnostic worksheet to structure the question before you test, then move through framing the diagnostic and separating access from content if your first test points you toward the other hypothesis. When the result is in, make the handoff concrete enough for whoever acts next.
What this guide does not claim
This notebook reports no completed review, no verified visibility result, and no certification. It offers a method for investigation and advice, not a delivery promise. The claims above cite the linked primary documentation; anything beyond that is labelled inference.