When an AI answers using your site, it never sees your page. A retrieval system slices it into chunks and hands the model one fragment: without your heading hierarchy, without the paragraph before it, without the context that made it make sense. This tool performs that exact slicing and shows every chunk alone, flagging the ones that stop meaning anything the moment they're separated from the rest.
The Retrieval Chunk Preview splits a page the way a RAG system does, by heading section then ~300-token windows with overlap, and shows every chunk in isolation exactly as a model receives it. It flags fragments that collapse alone: pronoun openers with no antecedent, 'as mentioned above' back-references, chunks that never name their subject, and link-list boilerplate. The verdict is one number: how many of your chunks fail.
This is the single most useful thing you can look at to understand Generative Engine Optimization, and no other free tool shows it. Retrieval-augmented generation splits documents before it stores them: first by heading section, then (when a section exceeds the embedding window) into token windows of roughly 300 tokens with a small overlap. The model receiving your content sees one of those windows at a time. A paragraph that opens 'It also supports transparency, which is why designers prefer it' is perfectly clear on your page and completely meaningless as a retrieved chunk: there is no 'it', and the sentence names nothing. The preview reproduces the standard chunking (heading-first, then ~300-token windows with ~45-token overlap), renders each chunk in its own bordered card exactly as a model receives it, and flags four specific failure modes: opening with a pronoun that has no antecedent in the chunk, referencing text that isn't there ('as mentioned above', 'see below', 'as we discussed'), never naming a concrete subject at all, and being mostly navigation or link boilerplate. It also flags the chunks that lost their heading to window splitting (the ones arriving with zero context) which is why long sections are riskier than most writers realize. The output is a single blunt sentence: how many of your chunks would be incomprehensible if retrieved alone.
Last updated
Markdown headings and full HTML both work. The tool uses your real heading structure to split the way a retrieval system would.
How many chunks the page produced, and how many of them fail when read alone.
Read each one as if it were all you had. Failing chunks are outlined in red with their flags and the exact triggering phrase listed underneath.
Name the subject again instead of 'it'. Delete 'as mentioned above'. Split long sections so each window keeps a heading. Restate context at the start of each section.
Paste the revision. The failing count should drop. Aim for zero on the pages you most want cited.
Heading sections first, then ~300-token windows with ~45-token overlap when a section runs long: the defaults mainstream retrieval pipelines actually use.
Each fragment gets its own card with nothing around it, so you read it the way the model does. This is the part that changes how people write.
Pronoun openers with no antecedent, back-references to absent text, chunks that never name a subject, and link-heavy boilerplate. Each flagged with the exact words that triggered it.
Only the first window of a long section keeps its heading. Later windows arrive with zero context. The tool counts them, which is the strongest argument for shorter sections you'll ever see.
"4 of your 11 chunks would be incomprehensible if retrieved alone." No jargon, no dial to interpret: a number you can act on immediately.
Chunking and judging run entirely in your browser. Paste unpublished content freely. Nothing is uploaded or stored.
Reading your own paragraph stripped of its context does more to change your writing than any amount of GEO advice. That's the whole product.
Every flagged chunk reads perfectly well in place: that's precisely why these problems survive editing and only appear when the page is fragmented.
A page can be accurate, well-ranked and thoroughly ignored by AI because none of its fragments stand up alone. This shows that directly.
Instead of an arbitrary 'keep sections short' rule, you see exactly which of your sections got split and which fragments lost their heading.
From quick one-off fixes to daily workflows, see how people put this tool to use.
Chunk your strongest article. The paragraphs you're proudest of are often the ones that lean hardest on the sentence before them.
Docs are written assuming linear reading and retrieved non-linearly. Every 'as described above' in your docs is a broken chunk waiting to be served to a user.
A red-flagged chunk card showing their own copy stripped of context communicates more than any deck. It's the most shareable artifact in the studio.
Building a RAG pipeline over a knowledge base? Preview how the source content chunks before you discover the problem in answer quality.
Boundaries stated plainly, with the right tool for each neighbouring job.
Accepts URL, TXT, MD and HTML, and produces Report and Share link, all processed locally in your browser.
See how the retrieval chunk preview fits into a step-by-step journey with related tools.
Score any page or draft on the factors a published study (KDD 2024) measured as increasing AI citation: fact density, source citations, quotability, answer-first openings, question headings and self-containment. Every sub-score shows its formula and what it counted.
Check whether every section of your page opens with a self-contained answer in 40 words or fewer. Flags throat-clearing intros, buried answers and openings that depend on missing context, with a rewrite target for each failure.
One URL in, one AI-visibility report out: crawler access with Cloudflare detection, JS rendering, citability scored with published methodology, retrieval chunks, answer snippets and schema. Each finding linked to the free tool that fixes it.
See your page the way AI crawlers see it: without JavaScript. Fetch the raw HTML, count the words that survive, detect client-rendering framework markers, and diff against the rendered DOM to list exactly which content is invisible to GPTBot, ClaudeBot and PerplexityBot.
Paste your description as it appears on your site, GitHub, LinkedIn, Product Hunt, Crunchbase and G2, and see where your own profiles contradict each other: name-spelling drift, conflicting numbers, conflicting years, profiles too thin to corroborate anything.
Check whether GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, Bingbot and every other AI crawler can read your site. Real RFC 9309 robots.txt matching with the deciding rule quoted, plus Cloudflare default-blocking detection. Free, instant, no signup.
Hit a snag? Here are quick fixes for the issues people run into most.
Read the flags on the first two or three cards: the pattern is usually a single habit repeated (pronoun openers, or a section that never restates its subject). Fixing that one habit typically clears most of the chunks at once.
Short is fine for a human skimming and weak for retrieval: under about 25 words there isn't enough to answer anything, so it will rarely be selected. Either merge it into a neighboring section or expand it into a complete standalone statement.
Chunking parameters vary by system (window size, overlap, whether headings are prepended). This uses common defaults (~300 tokens with ~45 overlap, heading-aware) so treat it as representative rather than byte-identical to any one vendor. The failure modes it flags are parameter-independent.
Chunks that keep their heading are judged with it. If you're seeing that flag, the chunk is a later window from a long section. It lost the heading in splitting, which is exactly the risk the tool is surfacing. Shorten the section.
Write every section as if it will be read alone, because it will be. That single habit fixes most chunk failures before they exist.
Restate the subject at the start of each section: 'WebP compresses…' beats 'It compresses…' even when the heading says WebP.
Keep sections under roughly 300 tokens (about 220 words) so each one stays a single chunk and keeps its heading.
Search your content for 'as mentioned', 'above', 'below', 'previously' and 'as we discussed'. Each hit is a chunk that will fail.
Link lists and 'related posts' blocks become boilerplate chunks that compete with your real content for retrieval. Keep them out of the main content area.
Recent updates and improvements to the retrieval chunk preview.
Initial release: heading-aware chunking with ~300-token windows and overlap, per-chunk isolation cards, four standalone-failure flags with quoted evidence, lost-heading detection for split sections, Markdown/HTML/URL input.
Chunking and judging happen entirely in your browser: pasted content is never uploaded, logged or stored. URL fetches use a stateless proxy that fetches, returns and discards.
Free and instant, Retrieval Chunk Preview processes your file securely and removes it right after.
