Text Chunking Simulator: How AI Splits Your Page into Chunks

Text chunking means splitting a text into pieces. Before an AI system reads a web page, it cuts the page into text blocks, and each block is found or missed on its own. This text chunking simulator shows where the cuts fall on your page: once after a plain conversion, once in the version publicly described for ChatGPT in September 2026. Below, you find what my benchmark across 1,000 domains shows and where the simulation reaches its limits.

Go to the simulator

Text chunking simulator: paste HTML, see the chunks

What is text chunking in AI?

Text chunking breaks a long text into smaller units that a language model processes one at a time. The term comes from memory research: people remember information more easily when it is grouped. In AI systems it serves another purpose. A search compares the question with individual blocks, not with the whole page.

For ChatGPT, Metehan Yeşilyurt (Peec AI) described on LinkedIn in September 2026 what fetched pages look like: the text resembles the output of the html2text tool, navigation and similar containers are gone, and the page arrives in blocks of roughly 170 words. The simulator rebuilds this description. It splits at paragraph boundaries and starts a new chunk as soon as the next paragraph would push it past 170 words.

Two versions of the same page

The simulator shows your page twice. The plain conversion writes links as [text](address), keeps image addresses and drops form fields. That is how many tools work that output “the text of a page”.

The ChatGPT version follows the eight features from the description. Links appear as text only, in the search result as 【1†text】. Images become “Image: description”, input fields and buttons become [Input: …] and [Button: …]. Emphasis uses different characters.

Both texts differ in length. Splitting both into blocks of 170 words therefore almost always cuts in different places. The simulator marks every cut of the ChatGPT version that does not occur in the plain conversion.

What my benchmark across 1,000 domains shows

I fetched the home pages of the first 1,000 domains on the Tranco list once and split them in both versions. 502 pages could be fetched, 475 returned text. The results are in my study I cut 475 homepages twice ↗, the data and code in the repository chatgpt-render-check ↗ (MIT).

• On 365 of the 475 home pages with text, the cuts landed elsewhere. Among pages with more than one chunk, it was 363 of 376.
• Pages without links never shifted (0 of 50). Link rendering is the main difference between the two versions.
• 27 of the 502 fetched home pages contained no text at all after conversion, among them TikTok, Spotify and five Amazon country shops. Their content is built by JavaScript or sits behind a bot check.
• 22.4% of home pages with text fit into a single chunk.

Reading the result: what to look for

If little or no text is left after conversion, your content is not in the HTML. A fetch without a browser then sees an empty page. Check this first; everything else depends on it.

Read the first chunk of the ChatGPT version. If it consists mainly of link text, menu items or teasers, your actual message only comes further down. The simulator flags it when more than 40% of the words in the first chunk come from link text.

A paragraph longer than 170 words forms an oversized chunk of its own, because the split only happens at paragraph boundaries. Short paragraphs with one idea each produce chunks that make sense on their own.

Moved cuts are not a flaw of your page. They show that a tool with a plain conversion forms different blocks than the described ChatGPT version. If you plan content chunks with such a tool, you plan them in different places than the description.

Limits of the simulation

The simulator reconstructs a second-hand description. Nobody here has insight into OpenAI’s technology. According to Malte Landwehr (Peec AI), OpenAI closed the gap through which this data was visible in October 2026. The described processing can no longer be observed since then and may have changed.

The result describes the structure of your text. It says nothing about whether ChatGPT finds, cites or recommends your page. It is a look at structure, not a lever for rankings.

I wrote the simulator from scratch for this page and measured it against my benchmark’s renderer on Oct 29, 2026. The cuts matched on 13 of 13 own pages and on 36 of 40 third-party home pages. Whether the cuts shift between the two versions, it judged the same way on 53 of 53 pages.

Frequently asked questions

Is the text chunking simulator free?

Yes, without sign-up. The tool runs entirely in your browser, sends nothing to a server and sets no cookies. Whatever you paste disappears when you close the page.

Why do I paste HTML instead of a URL?

Because a browser may not simply fetch third-party websites, for security reasons. A detour through my server would transmit your address. With the pasted source, everything stays with you. If you want your domain checked by address, use the free QuickScan ↗; it measures other signals than the simulator.

How big is a chunk in ChatGPT?

About 170 words, according to the description from September 2026. A fixed upper limit is not documented. The simulator uses 170 words as a target and splits only at paragraph boundaries, like the renderer of my benchmark.

How is this different from chunking strategies for RAG?

RAG systems (retrieval augmented generation) split their own documents, and their developers choose the strategy, such as fixed lengths, overlap or semantic chunking. The simulator shows how a third-party system splits your page, which you have no control over.

Chunking, structure and AI visibility in one review

I check what arrives from your website without JavaScript, how AI search splits your pages and whether ChatGPT and Perplexity name you for your customers’ questions. That is part of GEO and Agentic SEO. Related tools: the llms.txt generator and the robots.txt generator. The initial call takes thirty minutes and is free and without obligation. You receive all findings in writing afterwards.

Book an initial call ↗ Prefer to write