SEO in the AI era: is it still worth it?

Every 1,000 Google searches in 2025 result in roughly 360 clicks to websites outside of Google. That number was 440 a year ago.

SEO isn't dead, but a lot of the tactics that worked in 2022 either don't work anymore or actively hurt. The rules shifted enough that most SEO advice from that era is now misleading.

What actually happened

Google's AI Overviews (formerly Search Generative Experience / SGE) began broad rollout in 2024 and expanded significantly through 2025. It's now present on roughly 13% of all queries - heavily concentrated in informational, top-of-funnel searches.

The CTR impact is not subtle. Ahrefs' analysis found a 58% reduction in click-through rate when an AI Overview is present. Seer Interactive's September 2025 data puts it at 61% - from 1.76% CTR down to 0.61%. Paid results fell 68%.

This isn't evenly distributed. Definitional and informational content is getting hit hardest. HubSpot reportedly lost 70-80% of organic traffic. CNN lost 27-38%. Chegg lost 49% of non-subscriber traffic between January 2024 and January 2025. Sites built on "what is X" and "how to Y" articles are facing structural collapse.

Meanwhile, AI-referred traffic from ChatGPT, Perplexity, and similar tools grew 527% in 2025 across 400+ sites tracked by Superprompt. The conversion rates are interesting: visitors from ChatGPT converted at 15.9% vs. Google organic at 1.76%. Small in absolute volume, but not negligible.

Two search paradigms now coexist: the Google era of ranked blue links, and the LLM era of AI-synthesized answers with citations. You need a presence in both.

GEO - the new optimization target

Generative Engine Optimization (GEO) is the practice of optimizing to be cited in AI-generated responses across Google, ChatGPT, Perplexity, and Claude. It's distinct from traditional SEO in one critical way: instead of 10 competing results per query, LLMs cite 2-7 domains on average. The competition is fiercer and less visible.

A Princeton/Columbia/UMass research paper that established much of the GEO framework identified the signals that actually increase citation frequency:

• Direct answers in the first 40-60 words of each section (AI extracts from the opening, not from the middle)
• One statistic or verifiable fact every 150-200 words
• Question-based headings that mirror how users phrase queries
• Content updated within the last 30 days gets cited 3.2x more than stale content
• Sites that themselves cite authoritative sources get cited more frequently

That last point is counterintuitive but consistent: LLMs reward intellectual honesty. A page that links out to sources reads as more trustworthy than one that doesn't.

AI bots are already crawling your site

There are now multiple major crawlers hitting your robots.txt beyond Googlebot. Each has its own user agent:

You can allow or deny each independently. The distinction between Googlebot (search ranking) and Google-Extended (Gemini training) is especially relevant - you can opt out of feeding Gemini's training data while still ranking in Search.

User-agent: *
Allow: /

# AI crawlers
User-agent: GPTBot
Allow: /

User-agent: ClaudeBot
Allow: /

User-agent: PerplexityBot
Allow: /

User-agent: Google-Extended
Disallow: /

User-agent: Applebot-Extended
Allow: /

Sitemap: https://yoursite.com/sitemap.xml

A few things worth knowing about how these bots actually work:

Perplexity doesn't purely crawl. It calls the Bing search API to retrieve SERPs, then programmatically scrapes the top 5-10 results at query time. If your page loads slowly or returns non-200 status codes, it gets skipped.

Most LLMs don't crawl in real time. The "what the model knows" (training data) and "what the model retrieves at query time" (RAG) are separate processes. Optimizing for AI citation affects both, but differently.

SSR and SSG are a genuine advantage here. Bots like ClaudeBot and PerplexityBot handle JavaScript-rendered content poorly. If you're serving a fully client-side SPA, a significant portion of your content may be invisible to AI crawlers. Nuxt's static generation solves this cleanly.

The llms.txt proposal

There's a proposed standard - llms.txt - modeled on robots.txt but designed to give LLMs a structured, plain-text summary of your site's content and purpose. Anthropic has endorsed it in their documentation. Evidence of actual traffic impact is currently thin (Search Engine Land tested 9 sites, 8 saw no measurable change), but the implementation cost is near-zero and the direction is clear.

Schema markup is the connective tissue

Structured data matters more now, not less. Pages with rich results earn 82% higher CTR than unformatted pages. A DataWorld benchmark found LLMs grounded in structured knowledge achieve 300% higher factual accuracy. SearchVIU testing confirmed that ChatGPT, Claude, Perplexity, and Gemini all actively process Schema Markup when accessing content directly.

The most impactful schema types in 2026:

• Article / BlogPosting - establishes authorship, publication date, and topic. Required.
• FAQPage - feeds directly into People Also Ask and AI-extracted Q&A. High ROI.
• Person / Organization - entity recognition for E-E-A-T signals
• BreadcrumbList - navigation context for crawlers

Use JSON-LD. Google recommends it, and it's cleanest for AI parsing. Here's the minimum viable Article schema:

{
  "@context": "https://schema.org",
  "@type": "BlogPosting",
  "headline": "Your article title",
  "author": {
    "@type": "Person",
    "name": "Your Name",
    "url": "https://yoursite.com/about"
  },
  "datePublished": "2026-02-26",
  "dateModified": "2026-02-26",
  "publisher": {
    "@type": "Organization",
    "name": "Site Name"
  }
}

If you have FAQ content, add FAQPage schema to that section.

E-E-A-T has become load-bearing

Google's E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) framework predates AI Overviews, but the AI era has made it the primary differentiator. The flood of low-quality AI-generated content has pushed Google to double down on signals of real human expertise.

The practical requirements:

• Named author bylines with verifiable credentials (LinkedIn, GitHub, publication history)
• Transparent "last updated" dates on all content pages
• Your own content should cite authoritative external sources inline - not just claim things
• For technical content: demonstrate the work. Screenshots, real results, code that actually runs.

The connection to AI citation is direct: Google doesn't just evaluate your page in isolation. It cross-references author credibility across the web. If your author has no presence outside your own domain, that absence is itself a signal.

What to stop doing, what to start

The clearest strategic shift is this: stop optimizing for informational queries you can't win anymore. Thin definitional articles - "what is X", "how does Y work" - are being absorbed by AI Overviews entirely. If your content can be fully answered in one paragraph, AI will answer it and not link to you.

The content types that still drive clicks:

• Original research with unique data - AI engines cite you because nobody else has this
• Comparison and "versus" content - transactional intent that AI Overviews don't fully resolve
• In-depth tutorials with real code and screenshots - can't be synthesized from scratch
• Interactive tools - pure functionality that AI cannot replicate

Update your best existing content. Content refreshed within 30 days gets cited 3.2x more frequently. This is not expensive to do and it compounds.

Add AI referral tracking to your analytics now. Filter for chatgpt.com, perplexity.ai, claude.ai, and gemini.google.com as referral sources. This traffic is small today and growing at 527% year-over-year. Getting visibility into it early matters.

The new KPIs

The metrics that matter have changed:

Tools like Profound and Brandwatch now track how often your brand appears in AI responses. This is the equivalent of rank tracking for the LLM era.

The foundation still matters

Traditional technical SEO - crawlability, Core Web Vitals, clean HTML, sitemaps - is still the foundation. AI crawlers are less tolerant of broken infrastructure than Googlebot, not more. But if that's all you're doing, you're competing for a shrinking slice of a shrinking pie.

Getting cited in AI answers requires being genuinely useful at a level AI can't synthesize on its own - original data, real demos, actual expertise. Honestly, that's a higher bar than keyword density, which is probably a good thing.

Sources:

• Ahrefs: AI Overviews reduce clicks by 58%
• Seer Interactive: AIO impact on CTR, September 2025
• Superprompt: AI traffic up 527% study (400+ sites)
• GEO research paper - Princeton/Columbia/UMass (arXiv)
• Search Engine Land: Mastering GEO in 2026
• Writesonic: Structured data in AI search
• BrightEdge: Structured data in the AI search era
• The Digital Bloom: 2025 organic traffic crisis report
• SE Ranking: AI traffic research study 2025
• Dataslayer: Zero-click searches and CTR