How AI Reads Your Content: Chunks, Vectors, Rerank

A retriever can find the one document that satisfies every part of a buyer's question and still rank an incomplete document above it. In a 2026 long-document benchmark, the strongest system placed a fully qualifying document somewhere in its top 10 for 81.1% of 1,000 queries, yet ranked that complete document ahead of known partial matches only 35.8% of the time. Findable and chosen are separate events, and the gap between them is where most uncited pages live.

Here is the pipeline in one paragraph. Your page is split into passages, each passage becomes a vector in an index, a query becomes a vector and pulls back the nearest passages, a second model often reorders those candidates in the research systems described below, and only the few survivors reach the model that writes the answer. Google states that its AI features use a query fan-out technique and that pages must be indexed and snippet-eligible, and OpenAI describes ChatGPT search issuing one or more targeted queries first. No vendor publishes its retrieval stack, so the specifics below come from published research systems, not a leaked spec.

How does a document become something a model can retrieve?

It becomes a set of short passages with vectors attached, because both the encoder and the answering model have finite input windows. A 2026 enterprise retrieval study running 1,280 experimental conditions describes its own baseline concretely: 512-token chunks with 64-token overlap, 768-dimensional embeddings, cosine similarity, BM25, pgvector, and cross-encoder reranking. That is already a more capable baseline than plain vector search.

Passages are small. In the multi-hop benchmark published in Transactions on Machine Learning Research in May 2026, each evidence passage averaged 188 tokens with a 56.26-token standard deviation, and even a four-hop question supplied only about 750 tokens of oracle evidence. The unit the machine handles is a paragraph or two, not your page.

That reorganizes everything downstream. The model never weighs your page against a competitor's. It weighs the fragments that surfaced, so a strong page whose answer is spread across four paragraphs can lose to a weaker page resolving the question inside one.

Why does splitting a page into passages change what gets cited?

Because the boundaries decide what a retrieved fragment contains, and a fragment arriving without its answer is dead weight. The test is to lift any paragraph out of its section and ask whether it still answers something. If it leans on the three paragraphs above it, the retrieved chunk will not carry that setup.

Structure helps, though less dramatically than the advice implies. The 1,280-condition study added a heading and table-of-contents retrieval channel and measured Cohen's d of +0.41 over the chunk-based baseline on answer completeness, with a 95% bootstrap interval from +0.00 to +0.87. In its ANOVA the same feature explained a small share of score variance, at F = 4.69, p = .031, and partial eta squared = .012. Query and document difficulty did far more of the work.

The channel still carried weight when it fired. Across the study's 192 ToC-on conditions, headings supplied 20.4% of the 1,226 citations in those runs. The grid was eight documents ranging from 5 to 195 pages and the per-document regression was not significant, so read it as support for clean structure rather than proof of a lever.

Placement inside the assembled context matters too. Liu et al. (TACL 2024) documented a U-shaped curve in which models use material at the start and end of a long input more reliably than material in the middle. That is about where your passage lands in the model's context, which you do not control. What you do control is where the answer sits inside the passage, and the chunk-size figures above are the reason to put it first.

How does a query actually find your passage?

By geometry rather than keywords. Your passage and the question both become vectors, and the system returns the passages sitting closest to the question's vector, which is why "ending my plan" can retrieve a passage about cancelling a subscription with no words in common.

Matching improves when the stored text has the shape of a question. A six-dataset experiment comparing naive retrieval, hypothetical document embeddings, and question-matched indexing found the question-matched pipeline reached 71.5% average claim recall against 53.6% for naive retrieval, with average context precision of 63.5% against a 42.3% baseline. Both figures come from the same study, published in IEEE Access and posted as an arXiv preprint. It used one embedding model and one generator on an internal index at retrieval depths of 1, 3, 5, and 10 chunks, tested no public websites, and saw noise sensitivity in relevant context worsen to 21.0% from 13.8%.

Naming yourself consistently helps a human reader, but it is not a retrieval guarantee. An audit of 443 entity-oriented retrieval configurations on a standard TREC newswire test collection found that once entity selection was kept independent of the relevance judgments used for evaluation, the best gain over a BM25 baseline was 0.051 MAP points, and none of the 437 unsupervised configurations beat BM25's 0.292. Across 193 entity-selection strategies, relevant and non-relevant coverage correlated at r = 0.954.

What does reranking do, and can it still drop you?

Reranking reorders the candidate set with a slower, more accurate model, and the gains are real without solving the problem. In the long-document benchmark above, a reranker improved one system's complete-first success by 6.8 points, with a corrected p-value of .005, and another by 7.3 points, with a corrected p-value of .002. Lexical and visual fusion added 8.7 points.

Model size does not rescue it. Scaling one embedding family from 0.6B to 8B parameters changed complete-first success by 0.0 points, with a 95% interval from -2.7 to 2.8, while moving raw discovery by 27.9 points. The benchmark ran on 2,021 documents and 41,441 rendered pages with a median gold length of 67 pages, a deliberately hard setting.

The failure this exposes is specific. A passage that covers the use case while omitting the one constraint deciding the answer gets retrieved, ranked, then quietly outranked by something carrying that constraint. Across all 1,000 queries the best systems delivered gold plus supporting evidence for only 5.1% to 5.3% of queries.

Can the model still drop you after your passage survives retrieval?

Yes, and a large share of wrong answers is made at this stage. The multi-hop benchmark covering 1,186 chemistry questions across 11 models found mean accuracy moved from 37.16% without context to 69.14% with complete oracle evidence, reaching 80.89% mean accuracy when a controller retrieved in steps and decided what to search next.

Coverage was the hinge. With every required hop present, accuracy reached 77.9%; with at least one hop missing it fell to 49.2%, a 28.7-point penalty. An incomplete passage set does not degrade the answer gently. It changes the answer.

Composition failed even where retrieval succeeded. 87.3% of incorrect answers had already retrieved the final-answer evidence and used it wrongly, and after restricting the analysis to runs with every oracle hop retrieved, 58.6% of errors were composition failures. One distracting passage was enough to drag accuracy from 84.8% without a distractor latch to 30.9% with one.

Which stage dropped you, and why does rewriting the page rarely fix it?

Match the symptom to the stage before spending anything. If a raw fetch returns an empty shell, you failed before chunking: Vercel and MERJ's crawler research found that OpenAI's and Anthropic's crawlers fetch raw HTML and showed no evidence of JavaScript execution, while Google's AI features reuse the same Web Rendering Service capabilities as Search. Disable JavaScript, load the page, and read what is left.

If the page is fetchable but never appears among cited sources, the failure sits in indexing or similarity. Seer found 87% of SearchGPT citations matched Bing's top results across 500 citations. The overlap is large enough that absence from conventional indexes is worth ruling out first, though the study measures agreement between result sets, not a dependency between them. If you appear and a thinner competitor gets quoted instead, you are losing at rerank or composition, where the fix is the missing constraint rather than the tone.

This is where most published advice goes wrong. C-SEO Bench tested conversational-SEO rewrite methods across 1,900 queries and 16,000 documents spanning 2 tasks and 6 domains, with 8 inherited content transformations. Only 3 of 54 unilateral conditions produced statistically significant citation-rank gains, and outside those 3 conditions the best content-side methods produced close to 0 boost in citation ranking. The benchmark paper points to traditional retrieval-side improvements as what moved citation position.

That corrective does not make evidence worthless. The earlier GEO study (Aggarwal et al., KDD 2024) tested nine content transformations against GEO-bench and found adding statistics, quotations, and cited sources improved on baseline by 41% on its visibility metrics, while keyword stuffing did nothing. The instruction is narrow: put checkable facts where the passage boundary will keep them.

How does Trovance show you which stage dropped your content?

Trovance works backward from the answer instead of forward from the page. You define the buyer questions that matter, and it runs them repeatedly across AI engines as tracked questions, preserving each answer run as a snapshot with the mentions, the citations, the recommendations, and the sources that carried the answer. That record makes the stages distinguishable, because a stage failure is only visible as a pattern across runs.

Each stage leaves a different trace in the record. A brand absent from every run while its pages resist a raw fetch is a retrieval failure. A brand cited but never quoted for the deciding detail has a passage-boundary or rerank problem. Answers built from third-party pages that describe you inaccurately point at entity resolution, and answers that quote you and still choose someone else usually name the missing constraint in their own citation list.

The diagnosis then has somewhere to go. Your Brand Core holds the claims you are entitled to make and the proof behind each one, so a recommended action names the asset the evidence record says is absent, once the fetch and rerank explanations have been ruled out: the constraint your comparison page never states, the number a rival's cited source already carries. An absence in the answers is not by itself a content brief. Drafts are produced from approved claims, a person reviews everything before it publishes, and the next analysis cycle reruns the same questions against what you observed before.

What Trovance will not do is promise the citation appears. No system can, because retrieval is probabilistic, vendors change their stacks without notice, and the research above shows a complete document losing to a partial one inside a controlled benchmark. Trovance is building toward tighter stage-level attribution. Today it observes answers, preserves the evidence, and helps you decide what to produce, which is a smaller and more honest claim than controlling the answer.

What should you do this week?

Start at the bottom of the stack, because nothing above it matters if the fetch fails. Pull your five most important pages with JavaScript disabled and read what comes back. Then run your ten buyer questions several times each across two engines, recording who is cited and which sources carried each answer.

Then do the passage work on the pages that lost. Give each section one question and answer it in the first sentence, name the subject instead of pointing back with a pronoun, and put the deciding constraint inside the same paragraph as the claim it qualifies. Use a table for anything comparative.

Be honest about the timeline and the ceiling. Fetchability fixes can surface within weeks, passage work follows recrawling, and none of it guarantees a citation on any schedule. If you would rather read the observed answers and their sources than assemble them by hand, start a free Trovance analysis and work from the record.

How retrieval actually behaves

• Do headings help AI retrieve long documents? - the full read of the 1,280-condition heading study.

• A relevant page can still miss the proof - why complete documents lose to partial matches.

• A named entity is not a retrieval signal - what the 443-configuration entity audit found.

• Why more evidence can still produce worse AI answers - composition failures after retrieval succeeds.

Fix the stage that dropped you

• AI crawlers don't run your JavaScript - the failure that happens before chunking.

• Can likely buyer questions improve AI retrieval? - the question-matched indexing experiment in detail.

• The most-quoted study in AI search, quoted correctly - what the GEO benchmark actually tested.

• How to show up in ChatGPT without chasing hacks - what to do once the rewrite tactics are off the table.

FAQs