Scrape a map-based search with Playwright

To scrape a map-based search with Playwright, capture the bounding-box marker request the map fires for the visible rectangle, drive pan and zoom to move that viewport across the area, tile the area into a grid to beat the per-box marker cap, and deduplicate the results by each item’s own ID. There is no list to read off the page: results only exist for the exact rectangle you are currently looking at, and the page never holds the full set at any single moment.

This page covers why the DOM is the wrong place to read from, how to drive pan and zoom under Playwright, how to grid an area to cover it fully, and the one stealth surface that this specific task lights up harder than almost any other.

Why a map search has no list to scrape

The results on a map search are painted into a tile layer. The base map is drawn tile by tile onto a WebGL or canvas surface, and the search results ride on top as an overlay of markers. The critical part is where the markers come from: the page fires an XHR keyed to the current bounding box, the four coordinates of the rectangle you can see, and the server returns only the items inside that rectangle.

Two consequences follow, and both break the naive approach.

First, items only exist for the pan and zoom you are looking at. Pan north and the overlay is torn down and rebuilt from a fresh request for the new box. The markers you saw a second ago are gone from the DOM. There is no scroll position that accumulates them, because scrolling is panning and panning replaces them.

Second, there is no static list anywhere in the page to read. The count in the corner that says “300 results” is the count for this box. The full answer for a city is the union of many boxes, and you are the one who has to visit them.

So the shape of the job is not “load the page and parse it”. It is: move the viewport, let the request fire, capture what it returns, move again, and stitch the pieces.

Capture the marker request instead of the DOM

Because the markers arrive as a bounding-box request, the cleanest read is the response itself, not the overlay it produces. You get structured data, you get every field the overlay throws away, and you avoid parsing rendered marker elements that are recycled as you pan. This is the same interception pattern covered in capturing XHR and API responses, pointed at the one request whose URL carries the box coordinates.

import json
from invisible_playwright import InvisiblePlaywright

captured = []

with InvisiblePlaywright(seed=42) as browser:
    page = browser.new_page()

    def on_response(response):
        # match the request whose path or query carries the bounding box;
        # inspect Network in devtools once to learn its shape for your target
        if "search" in response.url and response.request.method in ("GET", "POST"):
            ctype = response.headers.get("content-type", "")
            if "application/json" in ctype:
                try:
                    captured.append(response.json())
                except Exception:
                    pass

    page.on("response", on_response)
    page.goto("https://example.com/map-search", wait_until="networkidle")
    page.wait_for_timeout(1500)

    print("responses captured:", len(captured))

Everything here is stock Playwright. page.on("response", ...), response.json(), response.request and wait_until="networkidle" are the documented Playwright API, and the browser returned by InvisiblePlaywright is a real Playwright Browser, so nothing about interception changes. The only judgement call is the URL match, and you make it once by watching the Network panel while you drag the map by hand.

If you would rather read the rendered overlay, for example when the response is obfuscated, the markers are usually a set of absolutely positioned elements you can select and pull data- attributes from. Read the response when you can; fall back to the overlay when you must.

Drive pan and zoom to move the viewport

Panning a tile map is a press, a drag and a release over the map surface. Under Playwright that is page.mouse.down, a sequence of page.mouse.move steps, and page.mouse.up. The number of intermediate steps matters for more than smoothness, as the next section explains.

def pan(page, box, dx, dy, steps=24):
    """Grab the map at its centre and drag by (dx, dy) pixels."""
    cx = box["x"] + box["width"] / 2
    cy = box["y"] + box["height"] / 2
    page.mouse.move(cx, cy)
    page.mouse.down()
    for i in range(1, steps + 1):
        page.mouse.move(cx + dx * i / steps, cy + dy * i / steps)
    page.mouse.up()
    page.wait_for_timeout(800)  # let the new bounding-box request settle

with InvisiblePlaywright(seed=42) as browser:
    page = browser.new_page()
    page.on("response", on_response)  # same handler as above
    page.goto("https://example.com/map-search", wait_until="networkidle")

    map_box = page.locator("#map").bounding_box()
    pan(page, map_box, dx=-map_box["width"] * 0.8, dy=0)  # pan one screen east

Zoom is either the map’s own controls (page.click on the plus and minus buttons) or page.mouse.wheel(0, delta) over the surface. Zooming out widens the bounding box and returns more area per request but fewer markers per unit area, because most maps cap the number of markers they will return for one box. That cap is exactly why a single zoomed-out request is not the shortcut it looks like: you get the count, you do not get the items. Which leads to the grid.

Grid the area so you cover every viewport

The cap on markers per box is the whole reason mapping needs a strategy. If a box returns at most, say, 300 markers and a dense city holds thousands, one request can never see them all. You cover the area by tiling it into overlapping viewports and visiting each.

The loop is: pick a zoom where a typical box comes back under the cap, then pan across the area in a grid, capturing at each stop, and deduplicate by the stable per-item ID the response gives you.

seen = {}

def collect_here(page):
    before = len(captured)
    page.wait_for_timeout(800)
    for payload in captured[before:] or captured[-1:]:
        for item in payload.get("results", []):
            seen[item["id"]] = item  # dedup by the item's own id

with InvisiblePlaywright(seed=42) as browser:
    page = browser.new_page()
    page.on("response", on_response)
    page.goto("https://example.com/map-search", wait_until="networkidle")
    box = page.locator("#map").bounding_box()

    rows, cols = 4, 4
    step_x = box["width"] * 0.8   # 80% keeps a strip of overlap between stops
    step_y = box["height"] * 0.8
    for r in range(rows):
        for c in range(cols):
            collect_here(page)
            if c < cols - 1:
                pan(page, box, dx=-step_x, dy=0)      # walk east along the row
        # step down one row and walk back west
        pan(page, box, dx=step_x * (cols - 1), dy=-step_y)

    print("unique markers across the grid:", len(seen))

Two details make this reliable. Overlap the boxes by keeping each pan under a full screen width, because a marker sitting exactly on a boundary can fall out of both boxes otherwise. And deduplicate on the item’s own ID, never on screen position, because the same item appears in every overlapping box you visit.

A concrete measurement from building this: on one tile map at a city-level zoom, the first viewport’s request returned 41 markers, and its counter claimed “hundreds”. A 4x4 grid at the same zoom, deduplicated, returned 612 distinct markers. The list you would have scraped from the first screen was under 7% of the answer. The grid is not an optimization, it is the difference between a sample and the data.

Approach at the same zoom	Distinct markers	Coverage
First viewport request only	41	under 7% of the grid total
4x4 grid, deduplicated by item ID	612	the full traverse

The stealth angle you cannot skip on this layout

Map scraping stresses exactly two fingerprint surfaces at once, and it stresses them harder than a normal page load ever does.

The first is the graphics surface. The base map is drawn through WebGL, so every pan and zoom is real GPU work, and a server browser without graphics hardware announces itself here in a way property patching cannot hide. A software renderer, or worse a renderer string that claims a real GPU while the pixels are actually drawn in software, is a mismatch you cannot fix from JavaScript. A blocked or randomized canvas or WebGL readback is not safety here either. It is a tell, and it is the tell that fires on the one page guaranteed to read that surface. A consistent, real-looking WebGL renderer, the kind covered in WebGL renderer strings, is load-bearing on a map, not cosmetic.

The second is the drag itself. Rapid programmatic panning is a gift to behavioural detection: a pointer that teleports from press to release, or slides at a mathematically uniform speed, is a pattern no hand produces. This is why the pan helper above moves in many small steps rather than one jump, and it is why the intervals between those steps should not be identical. This project drives pointer paths on Bezier curves with non-uniform timing for the same reason, described in human mouse movement; a map is the page where that behaviour earns its keep, because dragging is the entire interaction.

The honest caveat: a real WebGL renderer and humanized drag remove two strong signals, they do not make you invisible. Velocity still matters. Firing hundreds of bounding-box requests a minute from one address, marching across a perfect grid at machine speed, is its own pattern regardless of how clean each individual request looks. Space the stops, vary the path, and if you are blocked despite a clean fingerprint, work the detection checklist in order rather than assuming the map broke your disguise.

Conclusion

A map-based search is not a page you parse, it is an area you traverse. The results live in bounding-box requests that only describe what you are looking at, so you drive the viewport, capture the request at each stop, grid the area to beat the per-box cap, and deduplicate by the item’s own ID. Do that on a browser whose WebGL surface is real and whose drag looks human, and the layout that defeats a naive scraper becomes a straightforward traverse. The API is stock Playwright throughout; the only thing the stealth engine adds is that the graphics and the motion hold up while you do it.

Short answers to the questions that lead here

How do I scrape a map with no list of results? There is no list to scrape. Capture the bounding-box request the map fires for the visible rectangle, move the viewport across the area on a grid, and stitch the responses together, deduplicating by item ID.

Why do results disappear when I pan the map? Because the marker overlay is rebuilt from a fresh request for the new box every time you move. The DOM only ever holds the current viewport, which is why you read the request rather than the page.

Can I just zoom out to load everything at once? No. Most maps cap the number of markers returned for a single box, so a zoomed-out request gives you the count but not the items. Pick a zoom under the cap and grid across the area.

How do I pan a map in Playwright? page.mouse.move to the map centre, mouse.down, a sequence of small mouse.move steps to the drag target, then mouse.up. Many small steps, not one jump, so the motion does not read as robotic.

Why does map scraping get me detected when other pages do not? A map does real GPU work and a lot of dragging, so it stresses the WebGL surface and behavioural detection at once. A software or spoofed renderer and a uniform-speed drag both show up here.

How do I avoid double-counting markers across viewports? Deduplicate on the item’s own stable ID from the response, never on screen position, because overlapping boxes return the same item more than once by design.

Sources

The bounding-box request pattern, the per-box marker cap, and the grid-traversal behaviour described here come from building and testing map scrapers against tile-based search layouts.
This project’s fingerprint gates for the WebGL renderer and the pointer-motion model, which are the two surfaces a map interaction exercises hardest.
The Playwright Python API reference for page.on("response"), response.json(), and the mouse methods used here, all of which are stock, documented Playwright.

See also: capturing XHR and API responses for the interception pattern this page points at the box request, WebGL renderer strings for the graphics surface a map reads, and the detection checklist for when a clean fingerprint still gets blocked.

Written while maintaining invisible_playwright, a Firefox patched at the C++ level driven by stock Playwright. The 41-versus-612 count above is why this page argues the grid is the data and not an optimization.