Many teams assume that a page ranking well in search will naturally appear in AI generated answers. Sometimes it does. Often the relationship is weaker than expected. Understanding why helps you measure the right things and avoid investing on assumptions.
Rankings and answers are different outputs
A search results page is a list of documents. An AI answer is a composed piece of text. When a system writes an answer, it may draw on retrieved pages, on what the model learned during training, or on both, and it may cite some sources while summarizing others without a link.
That means the question changes from "is our page in the list?" to "is our information used, and is our business named or linked?" A page can rank well and still contribute nothing quotable to an answer, for example when the useful detail is buried beneath long introductions or spread across several pages.
Why the two can diverge
- Different questions. People phrase questions to assistants in longer, more conversational ways. Your ranking keyword may not match the question being answered.
- Different units. Search ranks pages. Answers use passages. A strong page with weak individual passages may be hard to use.
- Different providers. Each assistant has its own retrieval and model behavior, and results vary by model version, settings, and time.
- Entity ambiguity. If a page never clearly states who the business is and what it offers, an answer may describe the information without attributing it to you.
These are reasons divergence can happen, not measurements of how often it does. Test your own questions before drawing conclusions about your site.
What SEO still contributes
SEO remains essential. Google states that its AI features rely on the same fundamentals as search: pages must be crawlable, indexed, and eligible for a snippet, and there are no extra technical requirements or special schema for inclusion. Weak technical SEO can stop a page from being considered at all.
So the practical stance is both, not either. SEO makes content discoverable. GEO makes the content itself easier to understand, summarize, and attribute.
What to measure instead of assuming
- Keep your existing SEO metrics: indexing, impressions, clicks, and rankings for priority pages.
- Add recorded AI tests for real buyer questions. Save the question, provider, model where available, date, full response, and any source links.
- Separate a brand mention from a link to your domain, and both from absence. They mean different things.
- Score the pages that should answer those questions for readiness, so you know what to fix.
- Track referral traffic, qualified leads, and sales separately in analytics and your CRM, and keep unknown attribution unknown.
Over time this gives you a side by side view: where you rank, where you are mentioned, and where content improvements coincided with changes. It will not prove causation by itself, but it is far better than inferring AI visibility from ranking reports.
An illustrative scenario
Imagine a regional accounting firm that ranks well for "small business bookkeeping" in its city. When the team tests the question "Which bookkeeping firms in our city work with restaurants?" in two assistants, one response names three competitors and the other gives general advice with no businesses named. The firm's service page never mentions restaurants and introduces the firm only by its logo.
The likely next step is not more keyword work. It is adding a clear section on restaurant clients, naming the firm and city in the text, and then retesting the same question. This scenario is invented to illustrate the process, not a reported result.