Google began rolling out AI Overviews to United States search results in May 2024 and has expanded them since, adding an AI Mode that answers questions conversationally. When an AI Overview appears, Google’s system reads several pages from its index, composes a summary answer, and places it at the top of the page with citation links to the sources it drew from. The traditional organic listings still exist, but they start further down the screen, often below the fold on a phone.

The same pattern runs through the rest of the market. Microsoft’s Copilot answers inside Bing. Perplexity built its entire product as an answer engine: a question in, a cited answer out. ChatGPT browses the web and links its sources. Each of these systems performs the same basic act: it selects a small number of pages it trusts, extracts what they say, and presents the synthesis. A website is no longer competing for a click on a list. It is competing to be one of the handful of sources an answer is built from.

That is a narrower funnel, and it rewards different qualities than the old one did.

Early evidence points in one direction: when an AI answer appears, fewer people click through to websites. A 2025 Pew Research Center analysis of real user browsing found that visitors clicked a traditional result link on roughly 8 percent of searches where an AI summary appeared, compared with 15 percent of searches without one. Users who saw a summary were also more likely to end their session entirely, meaning the answer on the page was the end of the journey.

Several SEO platforms have reported similar patterns from their own datasets, with the steepest click-through declines on informational queries where the AI answer fully satisfies the question. The picture is not uniform, and the research is young, so the numbers deserve conditional reading. But no published study so far shows clicks increasing when a generated answer sits above the links.

There is a second side to the same data. The sites that do get cited inside AI answers still receive traffic, and that traffic tends to arrive better informed, because the visitor has already read a summary and chose to go deeper. Fewer clicks, but warmer ones. For a business measuring leads rather than raw sessions, that trade can be survivable, and in some cases favorable, provided the site is among the cited few rather than the invisible many.

AI Overviews concentrate on informational questions: how something works, what something costs on average, how to evaluate a service, what a term means. Queries phrased as questions, and longer conversational queries, trigger them most often.

Other query types still return mostly conventional Google AI search results pages or plain listings. Navigational searches (someone typing a business name to reach its site) rarely get a generated answer. Local searches with buying intent, such as a service plus a city, still lean heavily on the map pack and standard listings, because the searcher wants a provider, not an explanation. Transactional queries where the next step is a purchase or a call also see fewer AI answers, though coverage keeps expanding and the boundaries move every few months.

For a business website, this split matters for planning. The service pages that convert local buyers face less immediate disruption. The educational articles that attract early-stage researchers face the most, because those are exactly the questions AI systems now answer directly. That early-stage content still has a job, but the job description changed: it now needs to be the material the answer is assembled from, with the citation carrying your name in front of the researcher.

Watching what generative systems cite reveals a consistent taste. They favor pages that state facts plainly, define terms, answer a specific question near a clearly worded heading, and support claims with identifiable evidence. They struggle with pages that bury the answer in marketing copy, hide content behind scripts, or spread one idea across a thousand words of preamble.

The practical qualities that correlate with being cited look like an editor’s checklist. Headings that say what the section actually covers. A direct answer in the first sentence or two under the heading, with the detail following. Specific, verifiable statements rather than superlatives, because a language model assembling an answer quotes facts, not slogans. Author and business information that lets the system tie the content to a real organization.

Underneath the prose sits the machine-readable layer. Structured data (the schema.org vocabulary embedded in a page’s code) labels content explicitly: this is an article, published on this date, by this organization, at this address, answering these questions. Search engines have used this markup for years to build rich results, and the same labels help generative systems identify what a page contains and whether the publisher is a real, locatable business. This work has a name now, answer engine optimization, and an AEO effort mostly consists of doing these things deliberately instead of accidentally.

An underappreciated fact about machine readability: the work overlaps heavily with web accessibility. An accessible page is built with semantic HTML, a logical heading hierarchy, descriptive text alternatives for images, and content that reads correctly without visual styling. Those are the exact properties that make a page easy for a screen reader to present, and they are the same properties that make it easy for a crawler or a language model to parse.

A site that has been through genuine WCAG remediation typically has clean heading trees, real text instead of text baked into images, labeled forms, and link text that says where the link goes. Every one of those repairs also removes an obstacle between the content and the systems now deciding what to cite. The reverse holds too: a page built as a stack of styled boxes with headings chosen for font size rather than meaning is hard for assistive technology and hard for an answer engine, for the same underlying reason.

Businesses that invested in accessibility for compliance and usability reasons have, in effect, already paid for part of their AI visibility work. That does not make the two efforts identical: accessibility serves people first, and its legal and ethical case stands on its own. But when both projects are on the table, they share a foundation, and doing the structural work once serves both.

Exposure to this shift can be checked with tools most businesses already have.

Search Console shows which queries a site appears for and how click-through rates trend. A pattern of stable impressions with sliding clicks on informational queries often indicates an AI answer has been inserted above the listings. Searching your own important queries and noting whether an AI Overview appears, and whose sites it cites, takes an hour and shows exactly where you stand. Analytics platforms now record referral visits from chatgpt.com, perplexity.ai, and copilot.microsoft.com; the volumes are usually small today, but the trend line indicates whether answer engines are beginning to send your business its warmer clicks. Asking the tools themselves is also legitimate research: prompting ChatGPT or Perplexity with the questions your customers ask reveals whether your business is part of the answer or absent from it.

Together these checks produce a plain baseline: where AI answers already stand between you and your customers, where you are being cited, and where competitors are. From that baseline, the structural work has a defined target instead of a vague worry.

If you want that baseline built for you, ask us how your site reads to answer engines, not just to Google’s crawler. We will show you what the machines see, what they cite, and what would change that.