Classic SEO is a subset of AI search optimization. Both halves of that sentence carry weight. Subset: crawlability, content quality, entity work, and link authority all still pay, and the retrieval layer of AI answers runs on the search indexes you already know how to rank on. Subset, though, because ranking no longer finishes the job: answers are probabilistic, one prompt fans out into many queries you never see, and the model's memory of your brand was set at training time by work that classic SEO never touched. This chapter maps which of your skills transfer as-is, which need adjustment, and which instincts now point the wrong way.
The asymmetry that frames everything
Two numbers from OppAlerts' study across 100 industries set the frame. When an AI product recommends a brand in its top picks, that brand ranks somewhere in Google's organic results for the category 81% of the time. In the other direction, only about 27% of the brands holding Google top-3 positions get named in AI answers at all.From the OppAlerts ranking-factors study: 403,000 prompts, 100 industries, 13 models, with AI recommendations matched against Google organic results for the same categories. Method and tables in The Research Behind This Guide.
Ranking is necessary and no longer sufficient.
The 81% says AI's winners are overwhelmingly drawn from Google's winners, so the ranking work you already do remains the entry ticket. The 27% says the ticket stops guaranteeing admission: roughly three of four top-ranking brands get filtered out on the way from the results page to the answer. What does the filtering is everything this book adds to SEO: model memory, entity presence, extractable content, and the persona and phrasing of the prompt.
What transfers
Google's own documentation for AI features says the quiet part plainly: AI Overviews and AI Mode build on existing Search ranking systems, and no separate or exotic optimization is required for them.Google's "AI features and your website" guidance, which points AI-features optimization back at standard Search best practices. That covers Google's own products; the other platforms retrieve from other indexes, but the same logic holds, because an index is an index. The map from each classic activity to its AI-search counterpart:
| Classic SEO activity | AI-search counterpart | What changes |
|---|---|---|
| Keyword research | Prompt and topic research | Conversational prompts, fan-out queries, and personas replace the keyword list; see Prompt and Topic Research |
| Technical SEO | Crawler access for search bots plus AI bots | Same discipline, more user agents; see Technical Accessibility for AI Crawlers |
| On-page content | Writing for extraction | Passages must stand alone and answer directly; see Writing for AI Search |
| Structured data | Structured data, unchanged in markup | Consumed by more systems; see Schema and Structured Data for AI |
| Link building | Links plus unlinked brand mentions | Still the strongest correlate of AI visibility; see Links: The Signal That Refuses to Die |
| Entity and knowledge-panel work | Entity work, promoted from side task to core | Feeds model memory as well as Google's graph display; see Entities: Becoming a Thing the Machine Knows |
| Rank tracking | Sampled share-of-answer measurement | Repeated runs and distributions replace positions; see Measuring AI Search Visibility |
The academic evidence agrees that content-side tactics move answers: the GEO paper measured visibility gains of up to 40% in generated answers from changes like adding quotations, statistics, and cited sources, with effects varying by domain."GEO: Generative Engine Optimization" (Aggarwal et al.), benchmarked on GEO-bench. The 40% is the paper's best case, from experiments on generated answers rather than live products. Those are recognizably content-quality edits; a good SEO content team has been making them for years under a different justification.
What breaks
Four mechanics have no SEO equivalent, and each one breaks an instinct. Answers are probabilistic. An LLM samples its output token by token with deliberate randomness, so the same prompt on the same model names different brands on different runs. There is no fixed answer to check, only a distribution to sample. Checking once and reporting "we're in / we're out" is the new version of screenshotting one personalized results page.
The query multiplies. Products like Google's AI Mode take one prompt and fan it out into many simultaneous queries against the index, composing the answer from all of them. You never see most of those queries, so you are no longer optimizing for a keyword; you are optimizing for a cloud of related queries the system generates itself. Retrieval and Grounding: How an AI Answer Gets Assembled covers the pipeline.
The asker changes the answer. Who the prompt says the user is, a budget shopper, a luxury buyer, a senior, can rewrite which brands appear, far beyond what personalization did to rankings. In OppAlerts' persona testing, taste-driven categories reshuffled heavily by persona while commodity categories barely moved.The persona-sensitivity analysis within the OppAlerts study; method and per-industry tables in The Research Behind This Guide.
Half the game happened at training time. A model answering from memory draws on nothing you shipped this quarter. Model Memory and Live Retrieval quantifies the split; the operational point is that some of your visibility cannot be moved by publishing anything today, only by the slower reputation work that shapes the next training run.
The unlearning list
Habits to retire, in order of how much damage they do. Drop daily rank-check instincts: a model that has not been updated gives you the same distribution tomorrow, so daily sampling of a chat product mostly measures randomness. Drop single-answer thinking: never conclude anything from one run. Drop CTR reasoning: there is no title-tag bait that wins a click inside an answer that gets no click. And three keyword-era habits now actively hurt: exact-match keyword phrasing reads worse than natural language to systems that model meaning; one-page-per-keyword-variant architectures fragment the topical depth that extraction rewards; and chasing search volume misses the conversational prompts your buyers actually type, which no keyword tool records.
What a veteran can bill tomorrow, and what to build this year
Billable now, because they transfer intact: technical crawl work extended to AI user agents, content quality and information architecture, structured data, digital PR and link earning, and entity cleanup. These are the same deliverables with a second buyer reading the output.
Build this year, because they are new: prompt research as a discipline, sampled visibility measurement through model APIs, memory-versus-retrieval diagnosis, and persona-aware testing. None of them is deep engineering, but each requires accepting the probabilistic frame first, and the veterans who accepted the machinery early have won every previous transition. The rest of this book, starting with How an LLM Turns a Prompt into a Response, is the machinery.