Source Concentration: Why AI Recommendations Are Winner-Take-Most

Here is the uncomfortable truth about how AI answers actually get built: for any given question, the model doesn't consult ten sources, or fifty. It pulls a handful — often two to five — into the context it uses to write its answer, and everything else it could have used simply doesn't participate. I call this source concentration, and once you see it, the whole game of getting recommended by AI stops looking like SEO and starts looking like something much more brutal: winner-take-most.

This is my read from auditing how businesses show up across ChatGPT, Gemini, Claude, and Perplexity — not a claim about any single engine's internals, which none of us can see directly. But the pattern is consistent enough that I'd bet my practice on it. In classic search, being the eleventh-best result still put you on page two, where a determined searcher might find you. In an AI answer, there is no page two. There is the answer, and there is the void. You are in the small set of sources the model retrieved and trusted enough to use, or you contributed nothing to what the customer heard.

What source concentration actually means

When you ask an answer engine to recommend a business, a product, or an approach, a lot happens between your question and the reply. The engine interprets what you meant, often fans your one question out into several sub-queries, retrieves candidate material for each, and then — here's the part that matters — compresses all of that down to the small amount of evidence it will actually reason over. Context windows are large, but the model's attention is not infinite, and its instinct is to answer confidently and concisely. Concise answers are built from few sources.

So the retrieval funnel narrows hard at the end. Thousands of pages could be relevant. Dozens might get retrieved. But the answer the customer reads is assembled from the few the engine judged most retrievable, most corroborated, and most directly responsive. Everyone else — including businesses that are genuinely excellent — falls out at the compression step and never gets named.

That's the concentration. Visibility in AI isn't distributed across a long tail the way ten blue links were. It piles up on a short head of sources, per question, per engine.

Why "winner-take-most" is the right frame

Marketers spent two decades internalizing a distribution: rank one gets the most clicks, rank two a bit less, and so on down a gentle curve. There was room to be mediocre and still get scraps. Source concentration kills the curve. When an answer is built from three sources, the difference between being source number three and being source number four isn't a smaller slice — it's the difference between showing up and not existing. There's no consolation traffic for near-misses.

I say winner-take-most rather than winner-take-all deliberately, because it's not a single winner. An answer usually names a few businesses or cites a few sources, so there's a small club, not a throne. But the club is small and the door is narrow, and almost all the recommendation value in a category flows to the members. That's a fundamentally different competitive shape than the one your SEO habits were built for.

The concentration is per-prompt, not per-category

This is the nuance that trips people up. Source concentration doesn't mean one business dominates your whole industry across every AI conversation. It means each specific prompt has its own tiny winners' circle. "Best custom engagement ring designer in Denver" and "affordable engagement rings near me" are different questions that fan out differently, retrieve different evidence, and can name completely different businesses — even though a human would call them the same category. You can own one prompt and be invisible on the one next to it.

That's actually good news, because it means the winners' circle is contestable one prompt at a time. You don't have to beat the whole market. You have to become one of the few corroborated, retrievable sources for the specific questions your customers actually type.

How the few get chosen

If concentration decides that only a handful make it, the obvious question is: on what basis? From everything I've observed, three properties do most of the sorting.

Retrievability. The evidence has to exist somewhere a model can reach and parse — fast, structured, on the open web, not buried in a slow homepage, a PDF nobody links to, or a walled platform. If a model can't retrieve it, it doesn't matter how true it is. Untouchable excellence loses to retrievable adequacy every time.

Corroboration. Engines are conservative about repeating a claim that only you make about yourself. When multiple independent sources say the same thing — you exist, you do this, you're credible at it — the model's confidence clears the bar it needs to name you. A claim that lives only on your own site is a claim the engine is reluctant to stake an answer on. This is why third-party mentions, directories, and genuine press do disproportionate work.

Direct responsiveness. The source that answers the actual question cleanly beats the source that's merely about the topic. A page that directly addresses "how much does a custom engagement ring cost and what drives the price" is more useful to the engine's compression step than a beautiful brand page that gestures at craftsmanship. Answer the question the way the customer asked it, in the words they used.

What this changes about your strategy

Once you accept source concentration as the operating reality, several conventional instincts flip.

Stop chasing coverage. Start choosing prompts.

Trying to be everywhere in a winner-take-most channel is how you end up nowhere. You don't have the corroboration budget to make the winners' circle on every conceivable query. Pick the prompts that matter most — the ones with real buying intent in your category and geography — and concentrate your evidence-building on those. Depth on the prompts that convert beats breadth on the prompts that don't.

Treat "almost recommended" as zero.

The most dangerous place to be is fourth in a three-source answer, because from the outside it looks like you're doing fine. Your site is good, your reviews are strong, you feel competitive. But the recommendation went to the three the engine could assemble the answer from, and you're not one of them. Measure whether you're actually named in the answers your customers get — not whether you could plausibly have been. The gap between "should be recommended" and "is recommended" is exactly the gap source concentration creates, and it's invisible unless you go looking.

Build for the compression step, not the index.

Getting indexed or retrieved is necessary but not sufficient — plenty of retrieved sources still get compressed out. The sources that survive compression are the ones that are easy to lift a clean, corroborated, directly responsive statement from. Write so that a model can extract a quotable, verifiable claim about you in one pass. Make the evidence do the work in the fewest tokens.

A concrete example

I worked with a service business — excellent operators, genuinely the best in their metro at what they do, glowing customer relationships. When I asked the major engines to recommend a provider in their category and city, they named the same two or three competitors over and over. Not because the competitors were better. Because the competitors had left retrievable, corroborated evidence in the places the engines could reach, and my client hadn't. The answer had room for three names. My client's excellence lived in the heads of happy customers, where no model can retrieve it.

We didn't try to win the whole category. We picked the handful of highest-intent prompts, built out directly responsive content that answered those exact questions, and worked to get the business corroborated in third-party places that mattered. The goal was never to appear everywhere. It was to make the winners' circle on the questions that put money on the table. That's what source concentration demands: not more presence, but concentrated, retrievable, corroborated presence exactly where the answers get built.

Does concentration differ by engine?

Yes, and it's worth understanding the variation without over-indexing on it. Perplexity tends to show its work — it cites the sources it used, so you can literally see the concentrated set for a given question. That transparency is a gift: it turns "am I in the answer" from a guess into an observation. ChatGPT and Gemini are less explicit about their sourcing in the answer itself, and their behavior shifts depending on whether they're drawing on live retrieval or on what the model already absorbed during training. Claude, similarly, blends retrieved material with learned knowledge.

The practical implication is that there are really two concentrations happening at once. There's the concentration in live retrieval — the few sources pulled in to answer a fresh query — and there's the concentration in trained knowledge, where the model's default associations about your category were shaped by whatever was most prominent and corroborated when it learned. You want to be in both: retrievable now, and prominent enough over time that the model's baseline instinct already leans your way. The businesses that dominate a category in AI answers usually aren't winning one of these. They're winning both.

The realistic ceiling

Source concentration also sets an honest expectation about how much of a category any one business can own. Because each prompt has a small winners' circle and there are many prompts, no single business appears in every answer — nor should you plan as if it could. Your realistic ceiling is to be a reliable member of the winners' circle across the cluster of prompts that actually drive your business, and to be the first name on the subset where you're genuinely strongest. That's a winnable, defensible position. Chasing total domination of a category in AI answers is a good way to spread your corroboration too thin to make any winners' circle at all.

This is why I push clients toward ruthless prioritization. List the prompts your customers really type. Rank them by intent and value. Then concentrate — the word matters — your evidence-building on the top of that list until you're consistently named there, before you widen out. Concentration is the channel's nature; it should be your strategy's nature too.

The bottom line

Search rewarded breadth and tolerated mediocrity because the curve had a long tail. AI answers have almost no tail. Each question is settled by a handful of sources, and nearly all the value flows to them. That's harsh if you're used to collecting scraps from page two, and it's an enormous opportunity if you're willing to concentrate. The businesses that win the AI era won't be the ones that are technically present in the most places. They'll be the ones that made the short list on the questions that matter — retrievable, corroborated, and impossible for the model to leave out.

About the author

Scott Tischler is the Founder & Chairman of AIrecommend.ai and a practitioner-authority on AI search and Answer Engine Optimization. With 20+ years in marketing technology — including American Express, MetLife, and UBS — and executive and professional study at Wharton, Harvard, and Oxford, he helps businesses become the ones AI recommends.