How to Optimize Content for Google AI Mode Search

What Is Google AI Mode and Why Does It Change How We Write?

AI Mode replaces search results with a conversational answer built from many sources at once. Google's own report shows it has passed 1 billion monthly users, queries have more than doubled every quarter since launch, and the average query now runs about three times longer than a typical search.

AI Mode is Google's conversational search experience. Instead of ranking ten pages, it reads across many sources and writes a synthesized answer, then lets the person keep asking questions inside the same thread. According to Google's own usage report, published on its corporate blog in May 2026, AI Mode has surpassed 1 billion monthly active users globally, and query volume has more than doubled every quarter since it launched. The average AI Mode query also runs about three times longer than a typical text search, because people ask it questions the way they would ask a person instead of typing keywords.

That query shape shows up directly in Google's word-frequency data. The most common opening words in AI Mode queries are What, How, I, Is, and Can, and the most common keywords anywhere in the query are Find, Information, Identify, Explain, and Summarize. These read like sentences a person would say out loud, and nothing like the clipped keyword phrases SEO has trained us to target for a decade.

Why Doesn't Ranking First Guarantee a Click Anymore?

Ranking first no longer guarantees a click because AI Overviews cut organic click-through on the top result by roughly 58% as of December 2025, and Semrush clickstream data puts the zero-click rate inside AI Mode itself at 93%. A page now has to be cited inside the answer or it is not read at all.

Most SEO advice skips this drop in clicks, and it is the reason the other four signals in this report matter. An Ahrefs study of 300,000 keywords, run by Ryan Law and Xibeijia Guan and updated in February 2026, found that AI Overviews now cut the organic click-through rate on the number one result by 58% compared to searches with no AI Overview present, up sharply from a 34.5% reduction the same team measured in April 2025. Lower positions lost clicks too: position two saw a 50.8% drop and position three a 46.4% drop in the same dataset.

The numbers inside AI Mode itself are more extreme. Semrush analyzed roughly 69 million U.S. desktop search sessions between May and July 2025 and found that 93% of AI Mode sessions ended without a single click to an external website. Separate SparkToro research using Similarweb clickstream data, reported by Search Engine Land, put overall U.S. zero-click search behavior at 68.01% in early 2026, up from 60.45% two years earlier. We trust the trend because three independent measurement methods agree on it.

Why Do Full-Question Conversational Headings Perform Better?

Full-question headings match the shape of the queries AI Mode is trying to answer, which fragment-style headings written for keyword-matching search do not. Google's data shows those queries opening with What, How, I, Is, and Can far more than any other words.

For fifteen years, SEO headings were written to match a keyword fragment, so a page used "AI Mode SEO tips" rather than "How do you optimize for AI Mode?" That style matched how people typed into a search box, but it does not match how people talk to AI Mode. Google's own word-frequency data (What, How, I, Is, Can as top openers) says the query itself has become a full sentence, so the heading that answers it should be a full sentence too. The same thinking goes into our own SEO content system, and every H2 on this page is phrased as the question a reader would type.

When a heading is phrased as the literal question a user asks, the paragraph beneath it becomes a much cleaner unit for a model to lift and cite verbatim. Under a fragment heading the model has to infer the question, while a full-question heading states it outright.

What Does It Mean to Write for the "Do" Lane, Not Just the "Explore" Lane?

Writing for the Do lane means finishing the task on the page instead of only introducing it. Google groups AI Mode use into five themes: Explore, Decide, Learn, Create, and Do. The report gives growth rates but no size ranking between them, and Do/planning content is growing the fastest at 80% quarter over quarter.

Google's report sorts AI Mode activity into five use-case themes: Explore, Decide, Learn, Create, and Do. A popular explainer video summarizing this report claimed "Doing is the biggest bucket," and the report does not support that, because it contains no share-of-volume data at all and never ranks the five themes by size. What it does report is growth: Explore and brainstorming queries are resolving 30% faster, Decide queries (the "which one" comparisons) are resolving 40% faster, Create/image-generation queries have more than tripled since the start of the year, and Do/planning queries are the fastest-growing named theme at roughly 80% faster resolution. Those are growth figures, and nothing in the report measures which theme is biggest.

What we can act on is the shape of the Do theme itself. It covers people trying to get something done, such as planning a trip, building a checklist, or completing a form, rather than only researching a topic. Content built for Do puts the finished thing on the page, whether that is the checklist, the comparison table, or the step sequence, instead of a paragraph that points to where the reader could go find it. We build this into every page in our business automation work, because a page that hands over a working answer is also the page a model is most willing to cite.

How Should You Build a Follow-Up Question Block?

Build a follow-up block from the two or three questions a reader would naturally ask next, answered directly near the end of the page, so the model has material to keep citing across the thread. Follow-up queries inside AI Mode threads are growing more than 40% a month, so a single search is now a conversation.

Google's report measured follow-up queries growing more than 40% per month, meaning a search that used to end at one answer now often continues for several turns inside the same thread. A page written to answer exactly one question stops being useful after the first turn, while a page that anticipates the next two or three questions and answers them directly stays relevant for the rest of the conversation. We build this as a dedicated follow-up block near the end of every article, distinct from an FAQ section and aimed at the questions a reader would ask right after reading the first answer. It is the same discipline we apply when we design what an AI agent sees, described in our report on context engineering, where the work is anticipating the next question before it gets asked.

Why Does Multimodal Content Matter More Now?

Multimodal content matters more because more than 1 in 6 U.S. AI Mode searches are already non-text, and image-input queries are growing over 40% a month. Descriptive alt text, real captions, and clear visual hierarchy now serve as the data a multimodal model reads, on top of their accessibility value.

Google's report found that more than 1 in 6 AI Mode searches in the U.S. now involve voice, an uploaded image, video, or Live camera input rather than typed text, and image-input queries specifically are growing more than 40% month over month since launch. A page built only for a text-reading crawler is invisible to a growing share of that traffic. Descriptive alt text, real figure captions, and a visual hierarchy a model can parse (clear headings, labeled tables, images that illustrate the claim beside them) all become direct inputs, where before they were nice-to-haves.

The commerce data supports this from a different angle: the report notes Electronics as the top shopping topic and Price as the single most-referenced product attribute in AI Mode shopping queries. Those queries pull on specific detail that is structured and easy to compare, and the same specificity makes any page quotable in an AI Mode answer, shopping or not.

Is llms.txt a Google Ranking Signal?

No. llms.txt is a community convention that Answer.AI's Jeremy Howard proposed in September 2024, and Google has not adopted it as a standard. Google's John Mueller has said publicly that Google does not use it, so treat it as a technical experiment rather than an AI Mode ranking factor.

A popular explainer video about this same Google report also recommended adding an llms.txt file, calling it "what Google wants," but that claim is not in Google's report. llms.txt is a proposed convention published by Jeremy Howard of Answer.AI at llmstxt.org in September 2024, meant to give an AI crawler a short index of a site's key pages. It has never been adopted as an official standard by any search engine, and Google's own search advocate John Mueller has said publicly that Google does not use it and that server logs show major AI crawlers do not check for it either. Experiment with it if you want, but track it as a separate technical-SEO line item outside the four signals in this report.

How Do You Know If Your AI Mode Optimization Is Working?

Track citations inside AI Mode answers and branded search lift alongside rank position, since a page can rank well and still earn zero visits. Then check the four signals page by page as a checklist, the same verification we apply to any AI system before calling it done.

Because a page can now rank perfectly and still get zero clicks, the old scoreboard (rank position, then traffic) is not enough on its own. Track whether your content gets cited or quoted inside AI Mode answers for its target questions, and watch for branded search lift, meaning people searching your company name after encountering it inside an AI Mode answer they never clicked through. Then run the four signals in this report as a literal page-by-page checklist: full-question heading, a Do-lane answer that finishes the task, a follow-up block, and real multimodal structure. This is the evidence-over-impression standard we hold every AI system to in our report on harness engineering: you check that a page gets cited instead of assuming that a page which "looks optimized" will be.