How to scrape forum and community threads with Playwright

To scrape forum and community threads with Playwright, treat each thread as a tree: walk the reply nesting depth-first, expand every collapsed “load more replies” branch over XHR before reading the DOM, and strip quoted blockquotes so a copied paragraph is not counted twice. Run the whole multi-thread sweep under one seeded identity and pace it to a per-identity request budget. The rest of this page shows each step in stock Playwright.

A product listing is a flat sequence. A forum thread is not. Replies nest under parents to an arbitrary depth, whole branches hide behind a “load more replies” control, and a single post can quote another post word for word. Scrape it the way you would scrape a list of rows and you get a flattened mess: quoted text counted twice, deep branches missing entirely, and vote scores attached to the wrong body.

This page treats the thread as what it is, a tree, and walks it in the order that keeps the structure intact. The extraction is stock Playwright. The one part that is not is keeping a broad, multi-thread crawl reading as a single member rather than a new device per thread.

Why a forum thread is a tree, not a list

Every reply has a parent. The top-level posts hang off the thread, replies hang off those, replies to replies hang off those, and there is no fixed limit to how deep it goes. In the DOM this shows up as nesting: a reply’s children live inside its own container, not as siblings after it.

Three consequences fall out of that shape, and each one breaks a naive scraper:

  • Depth is unbounded. You cannot write a fixed number of nested loops. The walk has to recurse or it will stop at whatever depth you happened to hard-code.
  • Metadata sits beside the body, not inside it. The vote score, the timestamp and the user flair are usually sibling nodes to the post body within the same reply container. Grab the body text alone and you have thrown away everything that tells you who said it and how the community rated it.
  • A post can contain a copy of another post. Quoted text is embedded, usually inside a blockquote, and it is a verbatim copy of a body that already exists elsewhere in the tree. If you extract text blindly, you count that content twice and corrupt any frequency or reply-count analysis you run later.

So the job is not “collect all the comment nodes”. It is “walk the tree, keep each node’s own text separate from what it quotes, and carry the sibling metadata along with each body”.

Set up one identity for the whole crawl

Before touching the DOM, launch the browser once with a fixed seed and keep it for the entire sweep. A community site rarely rate-limits a single thread. It watches identity across a broad crawl, and a harvest that spans dozens of threads over dozens of pages is exactly the pattern it is built to catch.

from invisible_playwright import InvisiblePlaywright

# One seed, one identity, reused for every thread and every page in the run.
with InvisiblePlaywright(seed=42) as browser:
    page = browser.new_page()
    page.goto("https://example.com/forum/thread/1001")
    page.wait_for_selector(".comment")
    # ... walk the tree (below)

The browser here is a real Playwright Browser, so new_page, goto, query_selector_all and every other method behave exactly as they do upstream. The only difference from plain firefox.launch() is that the GPU, fonts, audio device, screen and the many other fields behind them are all derived from that one seed and stay identical for the whole run. Requesting fifty threads from one consistent member looks ordinary. Requesting fifty threads each from a freshly randomised device does not.

Walk the reply tree depth-first

Recursion mirrors the structure. At each node, read the node’s own fields, then descend into its direct children before moving to the next sibling. Depth-first keeps a branch together in your output, which is what you want when the meaning of a reply depends on the reply it answers.

def extract_node(handle):
    """Read one reply's own fields, then recurse into its direct children."""
    # Metadata lives in sibling nodes to the body, so read it off the container.
    author = handle.eval_on_selector(".comment__author", "el => el.textContent.trim()")
    score = handle.eval_on_selector(".comment__score", "el => el.textContent.trim()")

    # Direct children only: the container each reply keeps for its own replies.
    child_handles = handle.query_selector_all(":scope > .comment__children > .comment")

    return {
        "author": author,
        "score": score,
        "body": original_text(handle),          # defined in the next section
        "replies": [extract_node(c) for c in child_handles],
    }

with InvisiblePlaywright(seed=42) as browser:
    page = browser.new_page()
    page.goto("https://example.com/forum/thread/1001")
    page.wait_for_selector(".comment")
    roots = page.query_selector_all("#thread > .comment")
    tree = [extract_node(r) for r in roots]

The load-bearing detail is :scope > .comment__children > .comment. The :scope pseudo-class anchors the match to the node you called the query on; without the :scope > combinator a plain descendant selector matches every reply anywhere below the node, including grandchildren, and your recursion visits them again at the wrong depth. The direct-child combinator is what keeps each post counted once and placed correctly.

Expand collapsed branches before you read

Deep branches do not ship in the initial HTML. The server sends the thread down to some depth and replaces the rest with a “load more replies” control that fetches the missing subtree over XHR when clicked, the same load-more button pattern seen on flat listings but applied to nested replies. Read the DOM before expanding and those branches simply are not there. There is no error, just a silent hole in your data.

So expansion has to happen first, and it has to repeat, because expanding one branch can reveal another collapsed control nested inside it. Loop until none remain.

def expand_all(page, pause_ms=1200):
    """Click every 'load more' control, repeatedly, until the tree is fully open."""
    while True:
        controls = page.query_selector_all(".load-more-replies")
        if not controls:
            break
        for control in controls:
            if control.is_visible():
                control.click()               # arcs to the control on a Bezier curve
                # Each click is an XHR that injects a subtree. Wait for it to land.
                page.wait_for_load_state("networkidle")
        page.wait_for_timeout(pause_ms)       # pacing, not decoration - see below

If you would rather read the injected subtrees straight off the wire instead of re-scraping the DOM after each click, the response bodies those controls fetch are plain data, and capturing the XHR responses directly is often cleaner than waiting for the DOM to settle. Either way, the branch has to be requested before it can be read.

Separate quoted text from original text

A quote is a copy. When someone replies “as the parent said, X” and their client embeds the parent’s paragraph inside a blockquote, that paragraph now exists in two nodes: the original post and the quoting reply. Extract the reply’s full text and you have double counted X.

The fix is to read the body with the quoted blocks removed, and, if you want them, capture the quotes separately as references rather than as content.

ORIGINAL_TEXT_JS = """
el => {
    // Clone so we do not mutate the live page, then drop embedded quotes.
    const clone = el.cloneNode(true);
    clone.querySelectorAll("blockquote, .quoted-post").forEach(q => q.remove());
    return clone.textContent.trim();
}
"""

QUOTED_REFS_JS = """
el => Array.from(el.querySelectorAll("blockquote, .quoted-post"))
        .map(q => q.textContent.trim())
"""

def original_text(handle):
    return handle.evaluate(ORIGINAL_TEXT_JS)

def quoted_refs(handle):
    return handle.evaluate(QUOTED_REFS_JS)

Now each node carries only the words its author actually wrote, and the quotes it leans on are kept as a separate list you can match back to the posts they came from. Any later analysis, top authors, most-repeated claims, reply volume, is counting real contributions instead of an echo.

Pace the sweep to stay under the per-identity budget

Here is the honest caveat, and it is the part that a fingerprint alone does not solve. Matching a real browser removes the automation and device tells, so the crawl stops looking like a bot on inspection. It does nothing about volume. Expanding every collapsed branch multiplies requests: one thread that arrived as three HTTP responses can become forty once you have clicked open every subtree, and a hot thread spans dozens of pages on top of that. A stable identity that behaves like a member is still a member who can be throttled, and a member firing forty XHRs a second is a member who gets rate limited.

So pace it. The wait_for_timeout in the expansion loop is not cosmetic, and the same budget applies across threads, not just within one. The mechanics of holding a crawl under a per-identity request budget have their own page on rate limiting your scraper, and for threads that paginate rather than expand in place, the paginated-pages walk covers advancing page to page without tripping the same signal. The identity keeps you from looking like fifty devices. The pacing keeps you from looking like one very busy one.

Conclusion

A thread is a tree, so scrape it as a tree. Walk it depth-first so each branch stays together, expand the collapsed branches over XHR before you read anything, and strip quoted blocks so a copied paragraph is not counted as new content. Run the whole multi-thread, multi-page sweep under one seeded identity so the site sees a single member, and pace that sweep so the single member does not exhaust its own request budget. The fingerprint is what lets the crawl look ordinary. The structure and the pacing are what make the data correct and the crawl sustainable.

Short answers to the questions that lead here

Why does my forum scraper miss deep replies? Because they are not in the initial HTML. They sit behind a “load more replies” control that fetches them over XHR, so you have to click every such control, repeatedly, before reading the DOM.

How do I avoid counting quoted text twice? Clone each reply node, remove the blockquote and quoted-post elements, and read the text from the clone. Keep the quotes as a separate list of references if you need them.

Should I flatten the thread into a list? Only after you have walked it as a tree. Flatten during extraction and you lose the parent-child relationships and mis-attach metadata. Walk depth-first, then flatten your finished structure if you want.

Why do I get throttled halfway through a big crawl? Community sites throttle by identity across the whole crawl, not per thread. Expanding branches multiplies requests, so a deep sweep needs pacing even when the browser fingerprint is clean.

Does a stable seed help or hurt across many threads? It helps. A fresh random device per thread is a stronger signal than one consistent member browsing many threads. Reuse one seeded identity for the whole run.

Where does the vote score live in the DOM? In a sibling node to the post body inside the same reply container, alongside the author and timestamp. Read it off the container, not out of the body text.

Sources

  • The real invisible_playwright API as documented in Quickstart and Configuration: InvisiblePlaywright(seed=...) returns a stock Playwright Browser.
  • Standard Playwright element-handle and evaluation methods, used exactly as documented upstream, for the tree walk and the quote-stripping.
  • MDN: the :scope pseudo-class, for why :scope > .comment matches direct children only, not every descendant.
  • This project’s own measurements on identity-based throttling across broad crawls, summarised in the pacing section above.

See also: capturing XHR and API responses for reading injected subtrees off the wire, rate limiting your scraper for the per-identity budget, and scraping paginated pages for threads that span many pages.

Written while maintaining invisible_playwright, a Firefox patched at the C++ level driven by stock Playwright. The tree walk and the quote-stripping are stock Playwright; the one seeded identity across a broad crawl is the part the wrapper adds.

Back to top

MIT licensed. Every page here is written against the current source of the thing it describes, and several record something we got wrong first.

This site uses Just the Docs, a documentation theme for Jekyll.