AI Search Just Split Into Two Architectures and Most Brands Are Optimizing for the Wrong One

AI search is not one system. It is two completely different architectures retrieving your content in completely different ways, and most brands are treating it as a single optimization problem. That assumption is costing you visibility on whichever machine your buyer actually uses.

I have spent the better part of a decade watching brands optimize for search engines. First it was Google. Then it was "AI search." But "AI search" is not a thing. It is a category label stretched across two retrieval systems that share almost nothing under the hood. And a recent interview with Perplexity's Jesse Dwyer on Search Engine Journal finally put the architectural split into plain language that every founder should understand.

The Two Architectures, Named Plainly

Here is the split, stripped to the load-bearing difference.

Whole-document retrieval. ChatGPT, Google AI Overviews, and Google AI Mode work this way. They run a traditional search, pull the top 10 to 50 web pages, and hand those complete documents to a language model. The model reads them, synthesizes an answer, and cites the pages it drew from. Dwyer called ChatGPT's web search "4 Bing searches in a trenchcoat", and the joke is directionally accurate. The retrieval layer is still classic search. The AI sits on top.

Sub-document retrieval. Perplexity works this way. Instead of indexing whole pages, it indexes granular snippets of roughly 5 to 7 tokens each. When you query it, the system does not retrieve 50 documents. It retrieves approximately 130,000 tokens worth of the most relevant snippets, enough to fill the entire context window of the underlying language model. The goal, as Dwyer explained, is to saturate that window so completely that the model has no room to make things up.

That is not a minor technical distinction. It is a completely different philosophy of what "relevant" means. One system judges your page. The other judges your best paragraph.

Why This Matters More Than You Think

If you are optimizing for whole-document retrieval, your job is page-level authority. Domain reputation, backlink profile, title tags, structured data. The things SEO teams have measured for 20 years. ChatGPT commands 92.4% of all trackable AI referral traffic according to Previsible's analysis of 6.77 million sessions. So this path covers the volume.

If you are optimizing for sub-document retrieval, page-level authority is necessary but not sufficient. What matters is whether any individual passage on your site directly answers the decomposed sub-query that Perplexity constructed from the original prompt. Perplexity does not care that your page ranks number 3 on Google. It cares whether your second paragraph contains an extractable, specific, cited claim that matches what the user actually asked. This is why 80% of URLs that AI assistants cite do not appear in Google's top 100 results at all.

Two architectures. Two definitions of relevance. One brand, forced to serve both.

The Numbers That Prove the Fork

The market data confirms the architectural split is also a behavioral one.

ChatGPT crossed 1 billion monthly active users in mid-2026. It processes roughly 2 billion queries per day. It owns the casual discovery layer: "What's a good CRM for a 10-person team?" type queries where the user wants a fast answer and maybe clicks one link.

Google AI Mode hit 75 million daily active users and carries a 93% zero-click rate. AI Overviews appear on roughly 48% of Google searches. Google still holds 90.46% of total search market share. The AI layer is being welded onto the existing infrastructure, not replacing it.

Perplexity sits at 45 million monthly active users. Smaller in raw numbers. But its users run research-depth queries, citation-heavy comparisons, and buying-decision investigations. It averages 21.87 inline citations per response compared to ChatGPT's 7.92. More citations per answer means more chances for your content to appear in the answer, or more chances to be absent from it.

And there is a counterintuitive data point worth sitting with: Claude overtook Perplexity in monthly referral sessions in March 2026 and has maintained the lead since. Perplexity's referral traffic peaked in March 2025 and declined 61% by May 2026. The platform growing its user base while its referral traffic drops means users are getting their answers inside Perplexity without clicking out. That is what a saturated context window does. It makes the answer so complete that the source page becomes unnecessary to visit.

For the brand, that is the worst possible combination: your content is being used, but nobody is arriving at your site to convert.

What to Actually Do About It

Stop treating AI search as one optimization problem. Here are the three moves.

First, audit your retrieval surface on each architecture. Go to ChatGPT and ask it the buying question your ideal customer asks. See if your brand appears. Then do the same on Perplexity. The overlap between the two answer sets will be smaller than you expect. If your brand shows up on one and not the other, you know which architecture your content fails.

Second, build passage-level extractability into every page you publish. This is the move that works on both architectures. Answer the question in the first 100 words. Make every H2 a standalone claim with its own evidence. Use specific numbers, not vague assertions. "We reduced onboarding time by 43% across 200 accounts" is extractable. "We help companies onboard faster" is not. Whole-document systems benefit from this because the LLM finds your answer faster when summarizing. Sub-document systems benefit because your specific paragraph becomes a citable snippet.

Third, treat earned media as the retrieval bridge. 37% of consumers now start their searches with AI rather than Google. Earned media, the kind of coverage that lives on authoritative third-party domains, works on both architectures simultaneously. Whole-document systems weigh domain authority, and a mention in a DA-90 publication inherits that authority. Sub-document systems index the passage, and a specific claim about your brand in a Forbes or TechCrunch article creates a citable snippet that Perplexity can extract. One earned media placement feeds both machines.

The Architectural Split Is the Machine Relations Problem

I built AuthorityTech because I saw that the machines deciding who matters are not one machine. They are several, each running different logic, each learning from different signals. That was true when it was Google versus Bing. It is exponentially more true now that the retrieval architectures themselves have diverged.

The old question was: "Does my brand rank?" The new question is: "Which machine is my buyer asking, and does my content survive that machine's specific retrieval process?"

AI-referred traffic converts at 4.4x to 23x the rate of organic search traffic. That number is not going to shrink. The buyers who come through AI search are deeper in the decision cycle, more informed, and more likely to act. But they will only find you if your content passes the retrieval gate of the specific engine they chose.

Two architectures. Two retrieval logics. One brand.

You are either visible on both, or you are leaving one machine's entire audience to your competitors.

FAQ

What is the difference between whole-document and sub-document AI search?

Whole-document systems like ChatGPT and Google AI retrieve complete web pages and then summarize them. Sub-document systems like Perplexity index individual text snippets of roughly 5 to 7 tokens and retrieve the most relevant fragments to fill the language model's context window. The distinction determines whether your page or your best paragraph is what gets evaluated.

Which AI search engine sends the most referral traffic to websites?

ChatGPT sends 92.4% of all trackable AI referral traffic according to a Previsible study of 6.77 million sessions through May 2026. Perplexity's referral traffic has declined 61% from its March 2025 peak, suggesting users increasingly get complete answers without clicking through to source pages.

How should brands optimize for both AI search architectures?

Build passage-level extractability into every page: answer the query in the first 100 words, make each H2 a standalone claim with specific evidence, and use exact numbers instead of vague assertions. This structure serves whole-document systems because the LLM finds your answer faster, and sub-document systems because specific paragraphs become citable snippets. Earned media placements on authoritative domains feed both architectures simultaneously.