How does an AI
chunk your page?
Paste your page content. We run the same deterministic chunker the paid report uses, show every passage the retriever would see, and flag the chunks that lose attribution when they are pulled out of context.
What chunking does to your content
When an answer engine retrieves a source, it does not load your whole page into context. A retrieval pipeline cuts the page into passages — chunks — of a few hundred tokens, indexes them, and fetches only the ones its ranker thinks are relevant to the question. The model then answers from those passages alone. The split is mechanical, but its consequences are not: a chunk that reads perfectly in the context of the page above it can become unattributable noise once it is the only passage the model can see.
This tool makes that split visible. It runs the same deterministic chunker the paid IndexHalo report uses — roughly 400-token passages with a small overlap, never breaking mid-paragraph unless a single paragraph alone exceeds the budget — and shows every chunk in order, with its token count, a preview, and a verdict on whether it survives being retrieved alone.
The attribution test
A chunk is attributable when it still carries the subject — your brand or page name — so a model that retrieves only that passage can still say whose content it is. A chunk is broken when it cannot: usually because it contains a factual claim but no subject name, or because it opens with a pronoun whose antecedent was split into the previous chunk. Both are common, both are silent, and both mean an engine can use your words without naming you.
The simulator flags each broken chunk with the specific reason — a claim without a subject, a pronoun opener that lost its antecedent, or a fragment too small to stand as a passage — so the fix is concrete rather than a general 'write better content'. Often the fix is simply repeating the brand name where a new claim block begins, instead of relying on 'it' or 'they' from the paragraph above.
What this does and does not measure
Chunking is the deterministic half of a RAG simulation, and it is measured — the same split every time, over the exact text you paste, with no model in the loop. What it does not do is estimate how close your content sits to the questions buyers actually ask. That is retrieval drift, the second half of the paid report's RAG section, and it needs an embedding provider — so it is not part of this free tool. The two halves compose: fix the chunking here, then run a full report to measure drift against real queries.
Nothing you paste is stored, and no external API is called. The analysis runs entirely in the request and returns the passages and verdicts. If you want this over a live URL with the full drift step, that is the full evidence-readiness report.
Frequently asked questions
What is a RAG chunk?+
A retrieval-augmented-generation pipeline splits your page into passages of roughly a few hundred tokens so a model can fetch only the relevant ones when it answers a question about you. How that split falls determines whether a passage can still be attributed to you when it is retrieved alone, out of the context of the rest of your page.
Why does chunking matter for AI answers?+
Because an engine does not quote your whole page — it quotes the passage its retriever pulled. If that passage opens with 'it' or 'they' because the subject name lives in an earlier chunk, the model has your words but no way to name you. The chunker is the first place citation readiness is won or lost, and almost nobody checks it.
What does 'attributable' mean here?+
A chunk is attributable when it still carries the subject — your brand or page name — so a model retrieving it alone can say whose content it is. This tool flags chunks that contain factual claims but no subject name, and chunks that open with a pronoun whose antecedent was split off, because both lose attribution in isolation.
Is this the same chunker as the paid report?+
Yes. The paid IndexHalo report runs this exact deterministic chunker over your page as one step in its RAG pipeline simulation. The free tool exposes the chunking half — no embeddings, no external calls — so you can see the split and the per-chunk verdicts without running a full report.
Does it call any external API?+
No. Chunking is a deterministic text split — measured, not modelled. There is no embedding call and no network request; the analysis runs entirely on the text you paste. The paid report adds a retrieval-drift step that does call an embedding provider when one is configured, but that is not part of this free tool.
What should I fix if chunks are broken?+
Lead each chunk with its subject. Repeat the brand or page name where a new section or claim block begins rather than relying on a pronoun whose antecedent sits in an earlier passage, and keep standalone factual claims attached to the name they belong to. The goal is that any single chunk, retrieved alone, still names you.