AI Search Optimization: What Moves Citations and What Is Folklore

AI search optimization is the work of making a page eligible to be retrieved, cited and accurately summarized by AI answer features such as Google AI Overviews and AI Mode, ChatGPT search, Perplexity and Microsoft Copilot. Google's own guidance says doing so is still SEO. What differs is mostly crawler access and measurement.

Google's guide to optimizing for generative AI features says its AI experiences rest on the same core ranking and quality systems as Search, and that from its perspective, optimizing for generative AI search is optimizing for the search experience. It names answer engine optimization and generative engine optimization as terms in circulation and points readers to its guidance on evaluating third-party SEO advice. The acronyms have no agreed boundaries, which SEO, GEO, AEO and LLMO: What Actually Changes? covers in detail.

Our position, stated up front: most of what is sold as AI search optimization is ordinary SEO, plus crawler hygiene, plus measurement. The part that is new is small, testable, and prone to decay once competitors copy it. The rest of this article grades what the evidence supports and what it does not, so you can tell which claims deserve budget.

We use four grades. Platform statement means an engine's own documentation says it. Study-backed means a paper or measurement study supports it, with the setting named. Industry consensus means broad agreement without a controlled study. Single-source means one vendor or analyst said it and we could not corroborate it. Several of the studies below are arXiv preprints, and we did not check whether each has since been peer reviewed.

Yes, in the sense that AI Overviews have their own behavior worth understanding, and no, in the sense that you optimize them with a separate toolkit. Google says AI Overviews and AI Mode may use query fan-out: issuing several related searches across subtopics and data sources to build a response. The practical reading is that a page can be pulled in for a sub-question the searcher never typed.

That should change how thorough a page is, and it should not push you toward publishing a separate page for every variant. Google's guide says creating separate content for every possible variation, including fan-out queries, primarily to manipulate rankings or generated answers violates its scaled content abuse policy, and calls it ineffective over time. Cover the neighboring questions a buyer really asks, on the page where they would ask them.

Also keep the denominator in view. Google says AI Overviews appear only when its systems judge them additive to classic Search, and that they often do not trigger. The measurement study above found activation of 13.7 percent overall on its sample of trending queries, far higher for questions. If your buyers type short commercial queries, many of your impressions may never see an AI Overview.

No engine publishes a recipe for being cited. What each does document is a gate (what must be true for eligibility), a control (how you allow or limit access) and, in some cases, a report. The table records what we could read in October 2026. A blank cell means the page we read did not describe one, not that none exists.

Crawler decisions are where AI search optimization most clearly departs from classic SEO, because the assistants separate search access from training access. OpenAI says a site can allow OAI-SearchBot to appear in ChatGPT search while disallowing GPTBot so its content is not used to train generative AI foundation models, and that the settings are independent. Perplexity describes PerplexityBot as a crawler that surfaces and links sites in its search results and says it is not used to crawl content for foundation models. Google's AI features follow Googlebot rules, with Google-Extended as a separate control over AI training and grounding in some other Google systems.

Decide in two steps. Do you want to appear in this assistant's answers? If yes, allow its search bot. Do you object to training use? If yes, set the training bot separately. Write down the choice and the date, because both OpenAI and Perplexity say a robots.txt change can take about a day to take effect. Blocking everything by default carries a cost you should name: a site that opts out of OAI-SearchBot is not shown in ChatGPT search answers, though it can still appear as a navigational link.

Folklore here means a claim repeated more confidently than its evidence allows. Several of the most common ones are addressed directly in Google's guide, which has a section of things it says site owners can ignore for Google Search. Keep in mind that Google speaks only for itself, so a tactic can be useless for Google and still matter somewhere else. Where we say “no evidence,” we mean none in the sources we read.

The list runs from best-supported to least. Effort is a separate question. The items near the top are the ones where a platform has said so or a study found it; the ones near the bottom are reasoned from how these systems work, and we mark them.

Start by distrusting a single screenshot. The 2026 survey notes that commercial audits show low source overlap between runs and substantial run-to-run variability, so an answer that cites you once proves little, and one that does not cite you proves just as little. The survey's recommended protocol uses repeated measurements, paraphrases of each prompt, controls, human validation and attention to competitor interference. That is a sound frame for a business team.

Build a fixed prompt set of the questions real buyers ask, say twenty-five to fifty drawn from sales calls and search data. Run each prompt on each engine on a schedule, at least several times per round, with paraphrases. Record the cited URLs, whether your brand appears, and whether what the answer says about you is accurate. Hold back a few comparable pages you do not touch as controls. A GEO Test You Can Defend in the Monday Meeting walks through the design.

Then add the platform reports, each of which sees only part of the picture. Google tells you to use the Generative AI performance report in Search Console for visibility in its AI features, and warns against third-party tools that claim access to internal Google metrics, since no third party has it. LLM Visibility Tools: What They Measure, and What They Can Only Estimate explains the limits of those tools.

Because the space is full of confident numbers, a short interrogation saves money. The questions below apply to a pitch deck, a case study or a tool's dashboard. If the answers are vague, file the claim under single-source and do not repeat it.

In week one, audit eligibility: indexation, snippet controls, robots.txt and CDN rules, and a deliberate policy per assistant crawler. In week two, pick ten pages and ask of each what only you can say about the topic, then improve those, with real data and named sources. In week three, build the prompt set and record a baseline. In week four, review, and decide what to do before touching anything else. The AI Search Visibility Measurement Workbook provides templates.

Skip rewrites aimed only at the acronym of the month, bought mentions, tiny chunking and separate pages per variant. If a vendor quotes a guaranteed citation lift, ask four questions: which study, on which engine, on which dates, and what happened to the control group. A claim that cannot answer all four belongs in the single-source grade.

This area moves quickly. The documentation read here was current on 7 October 2026, the studies are a snapshot of products that change, and the arXiv papers are preprints. Re-read the Google, OpenAI, Perplexity and Bing pages each quarter, and treat any number in this article, including the AI Overviews activation rate, as a dated measurement rather than a constant.