Why an AI Answer Cites Six Sites When One Site Has the Best Material

An AI Answer Is a Set, Not a Ranking

Start with what the answers actually do. Across 68,879 searches by 900 US adults, Pew found that of the Google AI summaries it observed, "The vast majority of these summaries (88%) cited three or more sources. Only 1% cited a single source." A one-source answer is a statistical anomaly.

The slot count varies sharply by surface. A study of 7,583 AI Overviews found Google "cited 61,212 reference URLs drawn from 7,479 unique hostnames, with a median of 8 references per AIO (mean 8.1, range 1-32)." Assistants are tighter. Measured across 55,936 queries over six engines, "all LLM-SE source fewer URLs (mean: 4.3) and domains (mean: 3.4) compared to TSE", which reported means of 10.3 URLs from 7.3 domains.

Notice the second and third columns of that middle row. The mean number of URLs and the mean number of domains sit close together, which means an answer takes only a little more than one URL from each domain it uses. Winning a second slot is rare by construction.

The usual misreading of that table is that the engine simply has a quota to fill, so the extra slots go to whoever is next in line. If that were true, the sites beside you would be the ones ranked just below you, and they frequently are not. Something is choosing them on a basis that has nothing to do with being second best.

So the question was never whether other sites appear beside you. They will. The question is which ones, on what basis they were chosen, and why a page you know to be stronger keeps losing to pages you know to be weaker.

Each Next Source Is Chosen for What It Adds

The selection is not a ranking that gets truncated. It is a set that gets built. Each candidate is scored for relevance, then penalised by how much it resembles what has already been chosen, so a passage covering ground the set already holds scores close to nothing.

That is not a theory about engines. It is written into production search code. The OpenSearch engineering specification for this feature states the reason plainly: "Without MMR, results tend to cluster closely in the embedding space, reducing information diversity." The rule subtracts a candidate's highest similarity to anything already selected.

Google says the same thing to site owners. Its AI features documentation describes fan-out as a way to surface "a wider and more diverse set of helpful links" than classic web search. Its optimisation guide separately asks publishers for "a variety of sources", warning against content that simply restates what already exists.

This is the part that catches strong publishers. Your relevance does not fall. Your pages are as good as you think they are. What falls is the value of the next one to a set that already contains you, and that is the number the selector is maximising.

It also explains the asymmetry you feel. A competitor with a mediocre page on an angle nobody else covers is not beating you on quality. It is being scored on a dimension where the set is still empty, which is where the arithmetic is generous.

Why No Engine Will Stop Doing This

The obvious objection is that this sounds like a defect somebody will eventually fix. It is the opposite. Penalising redundancy makes answers measurably better, so every engine has a reason to keep it and none has a reason to drop it.

On questions that need several sources combined, the paper reports accuracy of 25% for nearest-neighbour selection against 41% once redundancy is penalised. Its description of the failure is the clearest in the literature: "K-nearest-neighbors and related methods (such as a cross-encoder reranker) simply return the highest individually relevant passages, without regard to whether the information in the passages is redundant."

A second group found the same effect on standard datasets, reporting a "margin exceeding 10%" in recall of documents containing the correct answer when a diversity term was added to the same retriever. Only the selection rule changed.

The mechanism has a cause worth understanding. A single relevance score cannot represent a question with two parts, so a relevance-only selector piles onto whichever part scores higher. GeoRAG measures the result: "Cosine top-5 covers Peak B only for 41.2% of HotpotQA queries", rising to 78.4% once coverage is the objective.

Across six benchmarks the same change "improves exact match (EM) by +6.5 to +7.5 points over top-k truncation (up to +9.7 on HotpotQA and ASQA)." Nobody gives up seven accuracy points. Treat the cap on your slots as permanent.

There is a second reason the behaviour is permanent, and it is about cost rather than accuracy. Context is finite, so every passage admitted to an answer spends budget that another passage cannot use. A selector that admits two passages making the same point has paid twice for one fact, which is why the penalty exists in production systems rather than only in papers.

Newer systems remove even the manual tuning. One 2025 selector reports gains of "8-15 percentage points" on hard-to-cover questions while calibrating the trade-off per query, which means there is no fixed setting left to game.

The DSF Evidence Variety Map

If slots are won by difference, the work is finding out what difference is available. The DSF Evidence Variety Map is a terrain map, not a scorecard. It plots the sub-questions a query fans into against the kinds of evidence each one wants, then marks who currently holds each cell.

The unit of contest is the sub-question, not the query. Google describes AI Mode as "dividing your question into subtopics and searching for each one simultaneously across multiple data sources", so a site can be dominant on one subtopic while absent from five others in the same answer.

Stage one is to write out those sub-questions. Stage two names the kind of evidence each one wants, and seven kinds cover most cases: measured data, first-hand operating experience, a standard or specification, a worked example, a price, a definition, a comparison.

Stage three fills in the occupant of every cell. Stage four finds the empty ones, and those are the slots available at almost no contest, because the penalty that discounts you on crowded ground works entirely in your favour on ground nobody holds.

Stage five is the discipline that keeps this honest. Take only the empty cells you can fill from something you actually have, such as operational data you can measure or results from having run the thing yourself. A cell you would have to fabricate is not an opportunity.

Reading a finished map is quick, because the verdicts fall into three groups. A cell held by nobody is worth taking. A cell held by you is worth defending rather than improving, since a better version of it adds nothing. A cell where a definition is already stated identically everywhere is worth skipping entirely, whatever its search volume looks like.

Stage six is declaring the kind, which the next section covers. The order matters: most content plans start at stage six, publishing more of what already exists in a slightly better form, which is precisely the move the selector ignores.

Concentration and Spread Are Different Problems

A naive reading of everything above is that difference alone wins, so any obscure angle buys a slot. The aggregate data says otherwise, and the correction matters because it decides where the work goes.

Citation share across many answers is severely unequal. One study reports that "OpenAI models have the highest level of inequality (G=0.83) in their citation patterns, followed by Perplexity models (G=0.77) and Google models (G=0.69)." That is more concentrated than income in most countries.

So entry into the eligible pool is genuinely hard, and being merely different does not buy it. But the pool is not simply the organic ranking with a new coat of paint. The same AI Overview study found nearly 30% of cited domains do not appear in the organic results shown beside the answer.

Read those two findings together and the split is clean. Concentration decides who is eligible to be considered, which is a matter of authority and corroboration. The redundancy penalty decides who gets picked from among the eligible, which is a matter of difference. They need different work, and confusing them wastes budget on the wrong one.

This is also why original content earns more than the consensus does at both stages. Originality is what gets a site into a concentrated pool, then the same originality is what distinguishes it once inside.

Declaring the Kind, Not Just Publishing It

The remedy is not more markup, and it is worth closing that door before opening the right one. Schema.org reports that "over 45 million web domains markup their web pages with over 450 billion Schema.org objects." Having structured data separates nobody.

Google has already demonstrated what happens when a format stops discriminating. Its documentation changelog records that FAQ rich results became limited to "well-known, authoritative government and health websites". The same changelog records How-to documentation removed outright because the result is no longer shown. Everyone had adopted both identically, so neither told an engine anything.

What still discriminates is the KIND of thing declared and the evidence behind it. Google documents seven artifact forms that qualify as a dataset, and pairs them with properties that say how a number was produced: "[measurementTechnique] The technique, technology, or methodology used in a dataset" alongside variableMeasured for what it measures.

The vocabulary even names this article's distinction directly. A Claim carries both appearance and firstAppearance, defined as "the first known occurrence of a Claim in some CreativeWork." Two pages can assert the same fact. Only one of them originated it, and the standard provides a way to say so.

Google's own publisher guidance is written in the same comparative terms. Its self-assessment asks "Does the content provide substantial value when compared to other pages in search results?" That is not a quality bar. It is a difference bar, and it is the editorial statement of the same rule.

What This Changes About What You Commission

The practical consequence lands on the brief. The instruction "make our page on this better than theirs" describes the one move the selector is built to disregard, so a content plan made entirely of improving existing coverage is buying discounted slots at full price.

The alternative usually means commissioning fewer pages rather than more. A plan organised by evidence kind tends to produce a short list of things only your business can supply, because the empty cells are empty precisely where somebody had to operate something to fill them.

It also changes what counts as done. A page that restates the consensus well is finished the day it reads well. A page carrying a measurement is finished when the measurement is repeatable, dated, and declared as the kind of evidence it is.

Measurement has to change with it. Rank is the wrong instrument when the contest is set membership, which is why Bing now reports citation share by topic rather than position. Share of a topic's citations answers the question rank cannot: how much of the set you hold.

None of this replaces the older work. Being retrievable at all still comes first, which is what the gap between being cited and being quoted covers. How many hidden queries your answer is really competing in is covered by query fan-out. This article is what happens after both.

One practical warning about sequencing. Evidence of the kind that wins empty cells usually has to be produced before it can be written, because a measurement takes a quarter to gather and an operating result takes longer. A plan built this way has a longer lead time than a plan built on rewriting, which is worth saying out loud before anybody commits to a publishing calendar.

A second passage that says the same thing better adds nothing to the set. The engine is not asking who is best. It is asking what is still missing.

If you want the map drawn against your own questions before anything is written, that is what a Website Health Audit produces: the sub-questions, the evidence kinds, the current occupant of every cell, and the empty ones you can actually fill.

FAQ — Why AI Cites Six Sites

Next Steps — Draw the Map

▶ Take your five highest-value questions and write out the sub-questions each one fans into, because the sub-question is the unit that wins a slot.

▶ For each sub-question, name the kind of evidence it wants: measured data, first-hand experience, a standard, a worked example, a price, a definition, or a comparison.

▶ Fill in who occupies every cell today, marking the ones held by nobody.

▶ Cross out every empty cell you cannot fill from something you already hold, then keep what survives.

▶ Publish the survivors as the kind of thing they are, declared machine-readably, then track citation share by topic rather than rank.

When the map shows which cells you can own, filling them is the engagement. Answer Engine Optimization (AEO) is where that work happens.