• Claude 3.5 Sonnet launched June 2024 but fails to parse basic headlines on live sites.
• Anthropic claims web search proficiency rivaling Google and Perplexity.
• Tests show Claude mislabeling or skipping H1s, even on Wikipedia and Shopify.

Let’s be brutally clear: Anthropic’s Claude Search, as shipped mid-2024, is a peak nothingburger when it comes to understanding actual site hierarchy. The Claude 3.5 Sonnet demo reels are slick, but paste a real URL—say, a Shopify product or a WordPress blog—and the model will either hallucinate the title, ignore headline tags, or regurgitate whatever junk is highest in the DOM. I ran a public test on June 19: Claude returned the site’s tagline as the ‘headline’ and ignored the article title sitting in a plain-as-day

tag.

Anthropic wants you to believe they’ve cracked AI-powered search, as if Google’s years of HTML parsing horror stories never happened. Sorry, but any “search engine” that can’t reliably extract a page’s most basic semantic element has no business pitching itself to agencies, publishers, or even bored interns. Even after months of feedback and LLM fine-tuning, Claude whiffs headlines on canonical sources like Wikipedia, Forbes, and Shopify. It’s not just a weird edge-case, it’s systemic, and rivals at Perplexity and even Gemini share the same bug.

This is not some esoteric parsing problem. The LinkedIn SEO influencer crowd will try to hand-wave this with “contextual understanding” babble, but here’s a one-line Python BeautifulSoup snippet: soup.find('h1').text. That’s it. If your $5 million foundational model can’t beat 10-year-old Python scraping on headlines, your AI “search engine” is a glorified chatbot with a browser plugin.

Let’s call out the enablers: lazy agencies slinging “AI SEO” audits, plugin pushers like Yoast and Rank Math who still generate nonsense schema, and the cottage industry of “Claude prompt templates” that amount to copy-paste cargo cult. Your client doesn’t care about “advanced summarization”, they want the headline to show up in the damn answer box. If Claude can’t do that, it’s not ready for prime time, period.

Here’s the uncomfortable fix: stop outsourcing basic web parsing to LLMs. Build deterministic headline extraction using real HTML and fallback rules, then feed the results to your AI model. Anything less is peak grift.

Frequently Asked Questions

Why can’t Claude Search extract page headlines reliably?

Claude Search’s current LLM pipeline tries to infer “main ideas” from page content instead of strictly following HTML structure. This leads to missed or hallucinated headlines, especially when the page uses custom themes, nonstandard markup, or has multiple prominent text blocks. It’s a core design flaw, not a fluke.

Are other AI search engines like Perplexity or Gemini any better?

No, similar issues persist across the AI search field. Perplexity and Gemini also often skip H1 tags or misattribute headlines, especially on sites with complex front-ends or semantic noise. LLMs’ “contextual reading” is still no match for traditional parsing on this front.

What’s the fix for headline extraction in AI-powered search?

Integrate deterministic HTML parsing (e.g., BeautifulSoup, Cheerio) before sending page text to the LLM. Extract the H1 and key metadata, then prompt the AI with those, not raw page content. Don’t trust generative models for basic web structure, they simply aren’t reliable yet.