How AI Search Engines Select Sources for Search Summaries

Watch an AI summary assemble itself and it looks like judgment. Ask a question, and a paragraph appears that seems to have weighed the good sources against the bad ones and picked the three that deserved to be quoted. That impression is the single most misleading thing about the format, and it leads marketing teams to argue about authority and brand strength when the actual reason their page was skipped is that a reranking model scored it 0.31 against a competitor’s 0.44.

Nothing in this process is mysterious. The architecture is published, it has a name, and the individual stages have been studied and measured by academics who had no commercial reason to flatter anyone. Understanding how AI search engines select sources is therefore an engineering question with an editorial answer. This article walks the whole pipeline from crawl to citation, names what each stage is actually judging, and shows where the published data on AI search ranking factors contradicts itself and why.

How AI summaries work, from crawl to citation

Nearly every AI answer engine runs some version of retrieval-augmented generation, an architecture formally described in 2020 research from Facebook AI, University College London, and New York University. The idea is simple enough to state in a sentence: rather than relying on what a language model memorized during training, fetch relevant documents at question time and let the model write its answer from those.

Between fetching and writing, though, sit several filters, and each one eliminates candidates. The stages below are where your page either survives or quietly drops out.

Read the failure column and a useful realization follows: only the first row is a technical problem in the usual sense. The rest are editorial. That is an unusual balance of responsibility, and it is why the work sits awkwardly between departments in most companies, something we account for when scoping the AI visibility engagements collected in our GEO portfolio, where the fix is nearly always split between an engineer and a writer.

Retrieval runs two searches at once

The retrieval stage is where most people’s mental model breaks, because they picture one search. Modern systems typically run two and blend the results, an approach usually called hybrid search.

The lexical half looks for the actual words. Classic scoring functions in this family reward passages containing the query’s terms, weighted so that rare words count for more than common ones and long documents do not win simply by containing more text. This half is why naming the specific thing still matters. A passage that says “commercial roof inspection” can be found by that phrase; a passage that says “we take a look at what’s up top” cannot.

The semantic half looks for the meaning. Passages are converted into embeddings, long lists of numbers positioning each passage in a space where related ideas sit near each other, and the query is converted the same way. Proximity in that space is the score. This half is why a page can be retrieved for a phrasing it never contains, and why writing around the concept helps even when you cannot guess the exact wording.

Neither half is sufficient alone, which is why serious systems combine them. Purely lexical retrieval misses paraphrases entirely, and purely semantic retrieval is notoriously weak on exact identifiers, product codes, and proper nouns that carry little conceptual signal. The practical instruction that falls out of this is to state things both ways: the precise term someone would type, and enough surrounding explanation that the meaning is unmistakable. Vocabulary gaps are the most common version of this problem, and they are usually specific to a sector, which is why the check runs early across the industries we build search programs for.

The reranker is where most candidates actually die

If you remember one thing about how AI search engines select sources, make it this stage. Retrieval is deliberately generous. It returns far more candidates than an answer can use, because it is optimized for speed across an enormous corpus and would rather over-collect than miss something. The narrowing happens next.

A reranking model takes each retrieved passage together with the query and scores how well that specific passage answers that specific question. Unlike the fast first pass, it reads the pair properly, which makes it slow, expensive, and applied only to the shortlist. It is also the stage with the highest elimination rate, and the one that punishes a particular kind of writing.

What gets punished is topical relevance without an actual answer. A page about roof inspections is topically perfect for a question about how long a roof inspection takes, and if it never states a duration, the reranker has nothing to score highly. The competitor’s page saying “a standard commercial roof inspection takes two to four hours for a building under 20,000 square feet” wins the slot with one sentence. Depth loses to directness here, reliably and unsentimentally, and it is the clearest answer to why AI cites some websites and not others when both cover the same ground.

Diagnosing this in a real content library is unglamorous work. It means taking the questions buyers actually ask, finding the passage on the site that should answer each one, and reading that passage cold, as if the rest of the page did not exist. Most fail. That exercise is a fixed early step in the way we work through an audit before recommending anything, because it separates a content problem from a technical one in an afternoon.

Surviving the shortlist is not the same as being used

Passages that clear reranking get assembled into a prompt and handed to the language model. Where a passage lands in that assembly turns out to matter more than anyone expected.

Researchers at Stanford, Berkeley, and Samaya AI tested this directly by moving the document containing the answer to different positions in a model’s input. Performance formed a U-shape: highest when the relevant material sat at the beginning or end, worst in the middle. With twenty documents in context, GPT-3.5-Turbo answered correctly 75.8 percent of the time when the right document came first, 63.2 percent when it came last, and 53.8 percent when it sat in the middle. The middle result fell below the 56.1 percent the same model scored with no documents at all, meaning a correctly retrieved source can be buried so badly that it makes the answer worse.

You cannot control your position in someone else’s prompt. What this finding does explain is the ordinary experience of appearing in an answer one day and vanishing the next without changing anything, and it is a strong argument against treating a single spot check as measurement. It also means the goal is not merely to be retrieved but to be scored highly enough by the reranker to be placed near the top of the pile, since the assembly order generally follows those scores.

Citations are attached to the answer, not chosen before it

The last stage is the one that surprises people most. The model writes the answer first, and the links are matched to sentences afterward. Attribution is a post-processing step, not the reasoning behind the sentence.

A Stanford audit of four generative search engines, Bing Chat, NeevaAI, perplexity.ai, and YouChat, measured how well that step actually holds up. Only 51.5 percent of generated sentences were fully supported by their citations, and only 74.5 percent of citations supported the sentence they were attached to. The authors also observed the uncomfortable pattern that the responses which read best were not the ones best supported.

“Responses from existing generative search engines are fluent and appear informative, but frequently contain unsupported statements and inaccurate citations.”

Two consequences follow for anyone trying to be cited. The first is that contributing to an answer and being credited for it are separate events, so a brand can shape what an engine says while receiving no visible attribution at all. The second is more actionable: because attribution is a matching step, content whose claims are stated as discrete, self-contained, checkable sentences is easier to match than content that develops an argument across a paragraph. A sentence carrying its own subject, its own number, and its own qualifier is a sentence an attribution step can bind to confidently.

Why the published citation studies disagree with each other

Most public claims about how AI search engines select sources trace back to a handful of vendor studies, and those studies do not agree. Ask what share of AI citations come from pages that already rank, and you will find confident, contradictory answers. Both camps have real data. The disagreement is instructive rather than embarrassing.

BrightEdge, tracking AI Overviews across nine industries from May 2024 to September 2025, reported the opposite direction of travel: overlap between citations and organic rankings rising from 32.3 percent to 54.5 percent, with the steepest growth in early 2025 before flattening. One dataset shows convergence with organic search. The other shows divergence.

The reconciliation is mostly methodological, and Ahrefs said so plainly. Its parsing improved between studies, so the later run detected citations the earlier one missed, many of them from deeper pages. It also pointed to query fan-out, where a single question is split into multiple related searches, as a growing factor: each sub-query has its own results, and a page that ranks nowhere for the original phrasing can rank first for one of the fragments. Different sampling windows, different industries, and different definitions of a citation produce the rest of the gap.

Two things are true across every dataset. Ranking helps, since Ahrefs found the first, second, and third citations sat at median organic positions two, four, and five. And ranking is not a gate, since a substantial share of citations come from pages that rank poorly or not at all. Regional variation compounds this further, since the results feeding a fan-out differ by market, and that inconsistency is a standing complication in the global search programs we manage across several countries.

What this changes about how to get cited in AI search

Once you know how AI search engines select sources, reading the pipeline backwards produces a short list of things that actually move the odds, and it looks nothing like a conventional ranking checklist.

Write the answer, then the context, not the reverse. The reranker scores a passage in isolation, so the first sentence beneath a heading has to contain the answer, including the subject, the number, and any qualifier. Everything that would normally build up to that conclusion belongs after it.

Let headings mark real boundaries. Chunking follows document structure, so a section that covers three subjects produces a chunk that answers none of them cleanly. One question per section, answered under its own heading, is a retrieval decision as much as a readability one.

Repeat the entity instead of relying on pronouns. A passage extracted from your page arrives without the paragraphs above it. “It usually takes two weeks” is unusable in isolation. Naming the subject in the sentence that carries the fact costs a few words and makes the passage portable, which is the single cheapest change on this list.

Cover both vocabularies. Hybrid retrieval means the technical term and the plain-language description are each doing separate work, and a page that only speaks one of the two languages is invisible to half the search. This matters most where regional or colloquial phrasing diverges from industry terminology, a pattern that shows up constantly in local search work built around how people in a specific market actually talk.

Those four changes are all editorial, which means a competent writer can make them without waiting on a release cycle. The fifth is the exception, and it belongs to whoever owns the codebase.

Make the technical layer boring. Server-rendered HTML, real heading elements rather than styled divs, and no bot rules that quietly exclude an engine you meant to reach. This is the only stage where a single mistake removes you from every subsequent stage at once, and it is the part we lock down first in the builds and rebuilds documented in our development portfolio.

Measuring something this noisy without fooling yourself

Knowing how AI search engines select sources does not by itself tell you whether your own pages are surviving the process, and that measurement is harder than it looks. Every stage above introduces variance. Retrieval depends on a query the engine wrote, reranking scores shift as models update, and position bias means the same shortlist can produce different answers on different runs. A single check proves nothing.

Useful measurement therefore looks like sampling rather than auditing. Fix a set of buyer questions, run each in several phrasings, repeat on a schedule, and record three states: cited with a link, mentioned without one, or absent. The pattern across a month is the signal. Any individual answer is a coin flip dressed up as a verdict, and treating one screenshot as evidence is how teams end up chasing changes that were never real.

The reporting standard that follows is simple and rarely met: say what was measured, how many times, and with what variance, and label the rest as inference. Holding that line is part of what clients tell us separates working with our team from the alternatives, because the alternative is a confident chart built on three lucky prompts.

A practical sampling plan is smaller than people expect. Fifteen questions, three phrasings each, run twice a month, gives forty-five data points per cycle and enough repetition to tell a real change from noise. Anything less is anecdote, and anything much more tends to get abandoned by the third month.

One last point worth keeping in view. Because attribution is imperfect and a large share of answers are assembled without visible credit, absence from citations is not proof of absence from the answer. Brand mentions inside AI responses that carry no link still shape what a buyer believes before they ever reach a website, and campaigns spanning many markets feel that first, which is why measurement plans built for national search campaigns covering a wide footprint should track mentions and citations as two separate numbers rather than one.

Frequently Asked Questions

Sources