Google AI search — what changed and what to do about it

Google's AI Overviews assemble answers from pages it trusts enough to cite. The question is whether your clients' pages are in that set.

The change is structural, not cosmetic. The answer box at the top of a search result is not a richer snippet. It is an assembled recommendation — built from pages Google's models read, attributed to specific sources, and surfaced when they match the query's intent. Sites that are not in the citation pool do not appear in the answer. They appear in the links underneath, when users scroll past the answer. When users scroll at all.

What Google AI search changed — not what Google says it changed

Google AI Overviews launched as a user experience upgrade. The framing was: a faster way to get answers without clicking through. That framing is accurate. What it understates is the structural implication for the sites below the answer.

Before AI Overviews, a page ranked in position three got traffic. Visibility and ranking were close enough to the same thing that the distinction rarely mattered. After AI Overviews, a page can rank in the top five and receive near-zero traffic if the answer box resolves the query without a click.

The thing that matters now is whether a page gets cited in the answer. Citation happens when the model reads the page, attributes it to a brand, and trusts the structure enough to quote from it. Ranking is one input into that process. It is not the whole process.

Three structural properties drive citation.

Entity clarity. The model has to know whose page it is reading. Pages that say "we" and "our services" without naming the organization explicitly leave attribution ambiguous. The model cites sources it can attribute. Ambiguous pages get skipped.

Schema coverage. Pages that emit structured data give the model a machine-readable description of the content type, the entity it belongs to, and the relationship between the two. Pages that do not emit structured data ask the model to infer. Inference is less reliable. Citation follows reliability.

Answer shape. AI Overviews extract passages — the answer to the implicit question. Pages structured around answers, with FAQ schema and clear heading hierarchy, provide extractable material. Pages structured as marketing prose do not extract cleanly.

These three properties determine citation-readiness. Most sites were built without them because they were not required before AI search existed as a primary discovery surface.

Why most sites don't appear in AI Overviews

Most sites were built when Google was the only entry point and ranking was the primary goal. The platforms they were built on reflect that goal.

WordPress with schema plugins covers some of the structural surface. The problem is integration. A schema plugin, a FAQ plugin, an AEO plugin, and an entity plugin each do a piece of the work. They do not coordinate. The FAQ plugin does not know about the entity plugin's organization schema. The schema plugin does not check whether its output is structurally valid when the FAQ plugin is also active. The AEO work the agency does depends on whichever combination is installed this month, and the combinations change.

Closed builders handle the template layer well. They do not give the agency access to the schema layer. The agency cannot control what structured data emits. Entity schema is whatever the platform generates, not what the citation work requires. The platform is optimized for publishing; the agency's citation work is not a priority input into the platform's roadmap.

AI-generated site builders generate a site quickly. The site is then structurally invisible to the same AI systems doing discovery. The category exists in a gap it has not yet named.

The common thread: these platforms were not designed for citation-readiness. They were designed for the search environment that existed when they were built. Retrofitting citation work onto a platform not designed for it means re-fighting the same structural gaps on every engagement.

What the pages that get cited have in common

Pages that appear consistently in AI Overviews are not necessarily the longest or highest-ranking. They share structural properties.

Entity markup is unambiguous. The organization name, the product name, the relationship between them — all present in JSON-LD, not inferred from page text. The model knows who the page belongs to without reading the copy.

Schema types are explicit. The page declares what it is — a BlogPosting, a FAQPage, a Product — in machine-readable form. The declaration matches the content. A page that declares itself a BlogPosting and contains blog post structure emits consistent signals. A page with no schema declaration provides none.

FAQ blocks emit FAQ schema. Questions on the page that would earn a citation in a conversational AI answer are marked up as Question entities. The FAQ schema shape is how LLMs source answers — the question-and-answer structure maps directly to how AI search extracts passages for citations.

The content is extractable. Answers are in the first 60 words of each section. Headings state what the section answers, not just what it is about. There is no padding before the answer.

These properties are not the output of a content strategy. They are properties of the foundation the content publishes on. If the foundation emits them automatically, every page is citation-ready by default. If the foundation requires manual intervention, every page is a project.

The foundation question

The question is not whether to optimize for AI search. The question is whether the platform the agency publishes on makes citation-readiness the default or an exception.

AI Launch Kit is the open-source website foundation built for this environment. The typed schema registry covers 44 page types — every page declares its type, its entity relationship, and its structured data at publish time. The AEO scoring CLI (ailk audit) runs before launch and in CI, catching the structural gaps that would have cost citations before they reach a live environment.

The MCP server exposes the site to agent-mediated research. An agent working on behalf of a buyer can call the MCP surface directly — reading page content, structured data, and entity graph without the friction of browser-based crawl. The site is readable by both audiences because the surface was designed for both.

The agency's citation work compounds on a foundation like this. The structural prerequisites are installed. The audit runs in the pipeline. The team's time goes to the work that differentiates the client, not to fighting the platform on every engagement.

AILK is Apache 2.0. The OSS tier ships the full schema registry, the AEO scoring CLI, and the MCP server. Paid tiers add Pro collaboration features, managed deployments, and the Launch Service. Nothing in the citation-readiness stack is gated.

Frequently asked

Is Google AI search the same as Search Generative Experience (SGE)?

Google's AI search features have been named differently at different stages. Search Generative Experience (SGE) was the experimental name during the testing period. AI Overviews is the name Google uses for the feature as deployed in Search. The underlying mechanism — AI-assembled answers cited from indexed pages — is the same. The terminology in this article refers to the deployed feature: AI Overviews and the citation behavior it produces.

Does ranking still matter for AI Overview citation?

Ranking is one input. A page that does not rank will not be crawled frequently enough to be considered for citation. But ranking alone does not produce citation. Pages that rank in positions three through ten and have strong entity markup and FAQ schema often appear in AI Overviews while higher-ranking pages without that structure do not. Citation-readiness requires both presence in the index and the structural properties the model uses to trust and attribute the source.

Does AILK guarantee AI Overview citations?

No. Google does not expose the citation mechanism as a configurable system. What AILK does is ensure the structural prerequisites are in place — entity markup, schema type declarations, FAQ schema, JSON-LD validity, and AEO scoring in CI. Whether a specific page gets cited depends on the query, the competition in the index, and Google's model behavior — factors outside the foundation's control. AILK eliminates the structural gaps that prevent citation. It cannot manufacture citation where relevance is absent.

What is the difference between AEO and SEO?

SEO (search engine optimization) targets ranking — the goal is to appear in the list of results. AEO (answer engine optimization) targets citation — the goal is to appear in the answer, not the list. The tactics overlap: both reward well-structured pages with clear relevance signals. The distinction matters because a page can rank without getting cited, and a page can get cited without ranking in the top position. Building for both requires addressing the ranking signals SEO has always optimized and the structural properties AEO adds.

Can I use the AILK AEO scoring CLI on a site not built with AILK?

No. The audit rules reference the AILK page-type registry. A non-AILK site will fail the schema coverage rules because the registry types are not present. The CLI is built for the AILK substrate.

Two routes from here

For agencies whose clients are asking about AI search. Or for developers who want to read the substrate directly.

Explore the foundation or get started: pnpm create ailk@latest