How to Adapt Your Website for AI SEO or GEO

Google search referrals to publishers fell 33% globally in the year to November 2025, per the Reuters Institute’s January 2026 report. Publishers expect another 43% drop over the next three years. Position 1 click-through rates drop 18% on average when an AI Overview appears above them, and 34.5% on purely informational queries. Half of all Google searches now end without a click.

Those numbers sound like the end of SEO. They’re not. Brands cited inside AI Overviews earn 35% more clicks per impression than uncited brands on the same queries, and AI-referred traffic converts 4.4x better than standard organic. The same pages that win classic search still win AI search because 76% of AI Overview citations come from pages already in the top 10 organic. What changed is which pages get extracted and that comes down to specific structural choices.

The SEO Basics Still Apply

The temptation reading those numbers is to rip everything up and start over with a new “AI SEO” stack. That’s the wrong call. AI search systems crawl the same pages Google indexes and pull from the same organic top results in most cases. They use the same E-E-A-T signals to evaluate sources. The work you’ve already done on crawlability, helpful content, page speed, and trust signals is the entry ticket to AI search, not legacy weight to throw out.

In practice, if a page isn’t ranking in the top 20 for its target query, structural changes for AI search won’t save it. Classic SEO fundamentals still gate everything else, and Google’s official guidance hasn’t changed in any material way for AI features. The new work sits on top of those fundamentals.

How AI Search Picks Sources

Each major AI search system has slightly different mechanics, but they share a pattern. They issue queries, retrieve documents, rank them against quality signals familiar to anyone who’s done classic SEO, then extract specific passages to quote or summarize. SEO writers have started calling this practice “generative engine optimization”, so if you’ve wondered what is GEO, that’s it. The page-level changes that get you cited are similar across systems. The crawling and ranking details vary.

Google AI Overviews and AI Mode

AI Overviews are the AI-generated summary boxes that now appear on 13-16% of Google queries globally, and as high as 70-80% for B2B Tech and education queries. They cite 2-3 sources typically, drawing from a candidate set that’s heavily weighted toward the top 10 organic results. Ahrefs research from March 2026 found 38% of AIO citations come from the top 10 organic, and a separate CXL analysis put it at 76%. The number varies by query type, but the pattern holds. Classic ranking is the entry filter.

AI Mode is the separate tab Google rolled out across 120+ markets through 2025. It replaces the traditional SERP entirely with a conversational interface. The mechanic underneath is “query fan-out”. Google issues multiple sub-queries in parallel, then synthesizes a single answer with a small set of citations. To be cited in AI Mode, a page needs to satisfy at least one of those sub-queries unambiguously.

Bing Copilot

Bing Copilot runs across Bing, Edge, Windows, and Microsoft 365. Citation behavior mirrors AI Overviews. Synthesized answer plus a small set of source links. The foundation is Bing’s own index, so strong Bing rankings still matter for Copilot visibility. Microsoft’s Bing Webmaster Tools now surfaces Copilot impression data alongside traditional ranking data.

ChatGPT, Claude, Perplexity, and Gemini

Standalone LLM search assistants behave differently from in-SERP AI. They have their own retrieval systems and don’t necessarily rely on Google or Bing rankings. These systems run their own real-time search crawlers (covered in detail in the robots.txt section below). Citations appear inline with clickable links back to the source.

The traffic these systems send is still small (ChatGPT accounts for 0.02% of publisher referrals and Perplexity 0.002% as of Q4 2025) but growing. The visibility from being cited matters for brand recognition even when the click isn’t there. The structural optimizations that work for AI Overviews also work for LLM search.

The pattern across all these systems is that GEO isn’t a separate discipline. It’s classic SEO plus a small set of additional structural choices that make your content easier to extract.

Page Structure for AI Extraction

Three structural choices account for most of the variation between cited and uncited pages on the same topic. Get these right and pages that are already ranking organically start showing up inside AI Overviews and LLM responses. Skip them and your pages stay in the top 10 organic but get bypassed when the AI synthesizes its answer.

Lead With the Answer

AI extraction engines look for self-contained, declarative statements that answer the query without requiring page context. The single highest-impact change is putting the direct answer in the first 1-2 sentences after the H1. A 40-80 word answer block hits the sweet spot. Under 30 words is too thin to carry context. Over 100 words gets split across retrieval chunks, which damages extraction coherence.

The opposite pattern kills citation rates. Pages that open with “We’re a leading hosting provider trusted by businesses worldwide” or “There are many factors to consider when picking a CMS” don’t answer any actual question, so the AI looks elsewhere for an extractable answer. A page that opens with “WordPress is open-source software that runs about 43% of all websites in 2026” gets cited every time someone asks what WordPress is, because the first sentence is the answer. The two examples could sit on the same domain with similar authority and content depth, and structure is what separates them.

FAQ Blocks

FAQ sections at the end of a page are disproportionately cited because they’re already formatted as question-answer pairs, exactly how AI constructs responses. Pages with FAQPage schema get cited at rates 20%+ higher than equivalent pages without it, per multiple 2026 citation studies.

The pattern that gets cited: 4-8 questions per page (6 is a fair default), each answer 40-80 words. Mark them up using JSON-LD FAQPage schema. Pull from Google’s People Also Ask box and your Search Console query data. Customer questions from support tickets are also useful sources. Avoid invented questions that nobody actually asks. The schema must match the visible page content exactly because Google enforces that rule and violations can trigger penalties.

Schema Markup

FAQPage schema is the highest-impact type for AI citation, but it doesn’t work alone. The pages that get cited most consistently combine it with Article or BlogPosting schema for the page itself plus BreadcrumbList for site structure. Organization or Person schema layers on for entity signals. Three to four complementary schema types per page works better than maxing out on a single type.

Specific numbers help inside the content itself. AI extractors prefer “WordPress usage grew from 38% to 43% between 2020 and 2025 according to W3Techs” over “WordPress has been growing in popularity.” Replace vague modifiers like “many” and “often” with specific percentages and dates where the data supports it.

Robots.txt and AI Crawlers

Most website owners haven’t thought about AI crawlers since they started showing up in server logs. The default behavior is either ignoring them or blocking them all with a blanket Disallow rule, and both leave value on the table. Some AI crawlers send referrals when they cite you, while others train models on your content without sending a user back. Treating them as the same thing is the most common preventable mistake on this topic in 2026.

Search Bots Versus Training Bots

The distinction every site owner should make is between AI crawlers that return citations with clickable links and crawlers that exist purely to feed model training. The split is roughly:

• Allow these (they send referrals): The active AI search crawlers are OAI-SearchBot, ChatGPT-User, Claude-User, Claude-SearchBot, and PerplexityBot. These power AI search products that link back to your site when they cite you. AI-referred traffic converts at roughly 4.4x the rate of standard organic search per industry data.
• Decide carefully on these (training only, no referrals): GPTBot, ClaudeBot, Google-Extended, Meta-ExternalAgent, CCBot, and Bytespider. Anthropic’s ClaudeBot is particularly aggressive, crawling roughly 20,583 pages for every single referral Claude sends back. OpenAI’s GPTBot sits at about 1,255 pages crawled per referral. Meta-ExternalAgent returns zero.

The most common mistake here is treating “ClaudeBot” and “Claude-SearchBot” as the same crawler, which they aren’t. Blocking ClaudeBot protects your content from being used for training. Blocking Claude-SearchBot, on the other hand, makes you invisible to Claude’s search feature, which is the one that actually sends real users to your site.

CDN Blocking

Roughly 27% of B2B SaaS and ecommerce sites are accidentally blocking major AI search crawlers at the CDN layer, per data from ziptie.dev. Cloudflare’s “Block AI training bots” toggle blocks search crawlers as well by default. Check your CDN dashboard before assuming your robots.txt is what’s actually being served.

llms.txt is Optional but Cheap

llms.txt is a Markdown file at the root of your site that gives AI crawlers a curated entry point to your most important content. Adoption sits at around 10% of domains per SE Ranking analysis of 300,000 sites. Anthropic has officially endorsed it, and OpenAI’s crawler has been observed checking for it. There’s no formal standard yet, and effectiveness is hard to measure, but creating one takes five minutes and has no downside.

Hosting Speed Matters More Now

AI crawlers have retrieval timeouts. A slow page gets skipped, regardless of how well-structured it is. About 69% of AI crawlers can’t execute JavaScript, which means heavy client-side rendering hides your content from them entirely. Server-side rendering or static HTML plus a fast host is foundational. More on hosting performance in our related guides.

Checking If It Works

Measurement is the part nobody has fully figured out yet, since traditional ranking and click-through metrics underweight what matters most now. The most reliable place to start is Google Search Console, which added AI Overview impression data in mid-2025 under the Performance report. It isn’t separated cleanly from regular impressions in most accounts, but the “Web” search type already captures it, and the gap between impressions and clicks tells you a lot. When impressions stay flat or rise while clicks drop, your content is probably showing up inside AI Overviews without earning the citation back to your site.

Beyond Search Console, the most useful exercise is manual checking, which still beats every automated tool currently on the market. Once a month, run your top 10-20 target queries through Google AI Mode, ChatGPT, Claude, and Perplexity, then note whether your site gets cited and which competing sources are pulled in instead. Branded search volume is the other proxy worth watching, because users who see your name mentioned in an AI answer without a clickable citation will often google your brand afterward, which shows up as a rising branded search trend even when direct traffic stays flat.

The number that should anchor everything in this article is that 76% of AI Overview citations come from pages already ranking in the top 10 organic. Classic SEO is still the foundation that AI search builds on. The structural work in the sections above adds a layer on top of those fundamentals rather than replacing the underlying SEO work that gets pages ranking organically in the first place.