Summary: ChatGPT reads, at best, the single page behind a link — and often misses JavaScript-rendered text. A Firecrawl blog post by Hiba Fathima (October 29, 2026) walks through two workarounds: an official ChatGPT plugin that crawls a site from inside the chat, and saving a full crawl as text files you upload to a Project. Both keep each page’s URL attached, which is what makes page-by-page summaries verifiable.
Why a link isn’t enough
Paste a homepage link into ChatGPT and you get that page at most. Two published tests show where this breaks: Peec AI’s logs found ChatGPT Deep Research reads only about 5,000–6,000 characters per page (June 2026), and Glenn Gabe’s October 2026 tests found ChatGPT and Claude picked up just the navigation and footer on JavaScript-rendered pages. So for anything bigger than one static page, you need to bring the site’s text to the model yourself.
Firecrawl’s post compares the options: pasting links (one page), copy-pasting pages (high effort), a CMS export (your own site only), or the Firecrawl plugin / a saved crawl — the last two reach every page, JavaScript rendered, with URLs intact.
Option 1: crawl inside the chat
Install the official Firecrawl plugin, sign in, and prompt something like:
Use Firecrawl to crawl https://docs.python.org/3/tutorial/, up to 20 pages,
only URLs under /3/tutorial/. For every page, give me its URL and a
three-sentence summary. Then write a one-paragraph overview of the whole section.The plugin runs Firecrawl’s /crawl under the hood, so each summary comes from the page’s full text, and it costs 1 Firecrawl credit per page. The free plan includes 1,000 credits a month. Claude users can do the same through the Firecrawl connector. This is the same crawl endpoint I covered in Mastering Firecrawl’s Crawl Endpoint, just triggered from a chat prompt.
Option 2: save the site as files for a Project
For a large site you’ll ask about repeatedly, crawl once and store the result. Firecrawl’s flow:
- Map first. List the site’s URLs (1 credit) to gauge size before crawling.
- Crawl with summaries. Request markdown plus summary formats; summaries add no extra cost. In their test, a 17-page crawl of the Python tutorial produced a full-text file of ~98,500 tokens and a summaries file of ~3,560 tokens — about 28x smaller.
- Upload as .txt. One text file holds up to 2 million tokens (a few hundred typical pages), per OpenAI’s file uploads help page. Projects hold 5–40 files depending on plan. Use .txt — some forum users reported .mdx files attached but weren’t read.
- For hundreds of pages, use Codex. The plugin also works in Codex, which can write full-page text to files and split them before hitting token limits.
Put recurring files in a Project with an instruction like “Answer only from the uploaded files and cite the URL above each fact.” For one-off questions, a chat attachment is fine.
Prompts that surface sloppy reads
Tell ChatGPT what the file contains, then probe it: summarize each page in order with its URL, group pages into topics, or “list every URL in the file.” That last one is a handy integrity check — if the list comes back shorter than the crawl, ChatGPT didn’t read the whole file.
Worth knowing before you start
Crawling costs 1 credit per page (a 100-page site is roughly 101 credits including the map call), some Firecrawl Alexandria providers need a team admin to accept terms first, and uploaded files go stale — rerun the crawl and replace the file when the site changes. If your goal is structured output rather than prose summaries, a dedicated extraction setup may serve you better.