How Does AI Search Work? I Asked Claude to Show Me

I’d already tested this with Gemini, which uses Google, and ChatGPT, which uses Bing. Both times I wanted the same answer to the same question: is there still a search engine sitting underneath these tools, or has that layer quietly disappeared. So when it came to Claude, Anthropic’s AI assistant, I went in with the same goal. I asked it to walk me through its process live, on a real question, to confirm how AI search actually works this time around.

The question I used was who is the best business credit card provider in Europe. Commercial, current, exactly the kind of query where a search engine would need to earn its keep.

How Does AI Search Work When the Answer Isn’t Fixed?

I wanted to know whether Claude would reach into some internal ranking and hand me a name, the way you might expect from a model relying purely on what it learned during training. It didn’t. For a question like this, where pricing, providers and terms shift constantly, Claude confirmed it doesn’t trust its own memory for anything current. Instead it runs a live web search and reasons over the results in real time, in the same way Gemini pulls from Google and ChatGPT pulls from Bing.

That confirms the pattern I’d already seen twice before. Claude estimated that for stable, general knowledge questions, historical facts, established concepts, it can answer confidently from training data alone most of the time. But the moment a question carries commercial or buyer intent, best provider for X, pricing comparisons, current recommendations, that reliance flips hard towards live search.

By Claude’s own estimate, that’s somewhere between eighty five and ninety five percent of the time.

If you work in marketing, this is the detail worth taking away. Search engines haven’t gone anywhere. They’ve simply moved a layer deeper, sitting underneath the assistant rather than in front of the user.

How AI Search Works Out Which Sources to Trust

What I wanted to check next was whether Claude assessed credibility, or just retrieved information and stitched it together. When the search results came back, it didn’t treat every source as equally trustworthy. It looked at a handful of things:

• Who published it, since an independent financial outlet carries more weight than a card provider’s own blog pushing its own product
• Whether methodology was disclosed, and one source stated it had checked pricing directly on provider websites rather than relying on affiliate content, which Claude flagged as a stronger signal
• How current the information is, because rates and terms go stale fast in financial services
• Obvious bias, where providers ranking themselves as “best” got treated with more scepticism

It also flagged the limits of what it found, rather than smoothing them over. The results skewed UK and fintech heavy rather than genuinely representing the whole of Europe, and they blurred together different product types. Some were true credit cards. Others were prepaid or debit cards branded as business cards. Claude said so plainly instead of presenting a confident answer built on shaky foundations.

Confirming Which Search Engine Sits Underneath

I already knew Gemini runs on Google and ChatGPT runs on Bing, so I wanted the same clarity from Claude. It was upfront that it doesn’t have direct visibility into that at the API level, but pointed to independent analysis suggesting it’s Brave Search rather than Google or Bing. It flagged that as inferred, not confirmed, which felt like the right amount of honesty for something Anthropic hasn’t officially documented.

Each search pulled back roughly ten results, about a page’s worth, which Claude then weighed by relevance rather than treating uniformly. When the first batch fell short, it ran follow up searches with refined queries rather than settling for whatever came back first.

Three assistants, three different search engines underneath. Google for Gemini, Bing for ChatGPT, and what looks like Brave for Claude. The specifics differ, but the pattern holds across all three: none of them are answering current or commercial questions from memory alone.

Why This Matters Beyond Curiosity

If you work in marketing, SEO or content strategy, this confirms something worth building a strategy around rather than guessing at. The old game was ranking on a results page. The new game includes something else entirely: being visible, credible and clearly sourced enough that an AI assistant chooses to surface and trust your content when it’s answering someone’s question directly, not just when someone happens to click through to your site.

Based on what Claude described, a few things follow directly:

1. Disclosed methodology beats vague superlative claims. Saying you checked pricing directly on provider websites carries more weight with AI search than simply declaring yourself “the best.”
2. Independent, current sources outperform self promotional ones. A blog post from a provider ranking its own product is treated with built in scepticism.
3. If you’re not appearing credibly in that first batch of live results, whichever search engine is doing the fetching, you may be effectively invisible to AI generated answers, regardless of how well you rank in traditional search.

None of this was hidden or hard to access. It just took asking directly, across three different assistants, and asking each one to show its working rather than hand over a polished conclusion.

What This Means for How You Publish

The practical takeaway for anyone producing content with commercial intent, comparison pages, “best of” roundups, pricing guides, is that search engines are still very much in play. They just sit one step further back than they used to. Cite your sources. State your methodology. Update your figures. Avoid dressing up self promotion as independent analysis, because that’s precisely the pattern Claude said it discounts, and there’s no reason to think Gemini or ChatGPT behave differently underneath.

This doesn’t replace the fundamentals of search engine optimisation. Technical performance, structured data and genuine topical authority still matter enormously, because none of these assistants can trust or surface content that was never indexed and ranked in the first place. What’s changed is that there’s now a reasoning layer sitting on top, deciding which of the indexed content is worth trusting.

It’s an odd thing to plan content around, admittedly. But then, so was writing for a search engine algorithm back in 2001, and that turned out to be worth taking seriously too.