Google Images results: who ranks, how big, and what you can download

Google Images looks like one big pool of pictures. If you build anything on it, such as a content tool, a moodboard, a dataset or a catalogue, three practical questions matter: who are the results actually from, how large are the images, and can you fetch them?

We ran 20 searches through the Google Images API on 16 October 2026 and measured all three across 2,000 results. This guide has the numbers, the code to check your own searches, and a note on what the API does not tell you. Image results change constantly, so yours will differ.

How we ran it

We used 20 searches in five groups, each with country set to us:

• Animals and nature: golden retriever, monarch butterfly, northern lights, mount fuji.
• Products: running shoes, standing desk, espresso machine, wireless headphones.
• Places and travel: eiffel tower, santorini, tokyo skyline, grand canyon.
• Food: sourdough bread, margherita pizza, chocolate cake, sushi platter.
• Design and business: minimalist logo, dashboard ui design, office interior, infographic template.

It cost 20 credits. Every search returned exactly 100 images, 2,000 in total. These are searches we picked, checked once, so the numbers describe this set and not Google Images in general.

Who the results come from

The 2,000 results came from 1,266 different page domains, which is a very long tail:

Instagram came first, with images in 13 of the 20 searches. Even so, the five domains with the most images held only 98 of the 2,000, 5%. And 897 of the 1,266 domains appeared exactly once.

Who leads depends on the subject:

Shopping sites led products, recipe sites led food, and design marketplaces led design. The image files themselves were hosted on 1,149 different hosts.

How big the images are

Every result has original_width and original_height, so size is easy to measure:

• Median width: 1,200 pixels. The range was 210 to 9,504.
• 67% were at least 1,000 pixels wide, 20% at least 1,920, and 4% under 500.
• 54% were landscape, so 46% were square or portrait.
• Design and business searches had the smallest median, 1,046 pixels, and only 59% reached 1,000 pixels. Food had the largest share at 1,000 pixels or more, 73%.

File types, from the extension in the URL: 1,404 .jpg (70%), 200 .png, 110 .webp, 82 .jpeg, and 192 with no extension at all. Do not depend on the extension. It is often missing, and what matters is the Content-Type the server sends when you fetch the file.

No image URL appeared under more than one of the 20 searches, and no response contained the same URL twice.

Can you actually download them?

The original URL points at someone else’s server, and a search result is not a guarantee that the file is fetchable. We took a random sample of 80 original URLs from the 2,000 and requested each:

So about five in six worked. One host stood out. The URLs on lookaside.instagram.com and lookaside.fbsbx.com, 58 of the 2,000, are the ones Instagram and Facebook results use. We probed 15 of them, and all 15 answered with a web page instead of an image. If you download images, skip those hosts up front.

The code

Put your key in an environment variable and use the google_images function from Is there a Google Images API?. The report and the download check look like this:

import statistics
from collections import Counter
from urllib.parse import urlparse

import requests


def domain(link):
    return (urlparse(link or "").hostname or "").removeprefix("www.")


def image_report(results_by_query):
    """Sizes and sources across searches. Each value is a list of images_results items."""
    items = [item for images in results_by_query.values() for item in images]
    widths = [i["original_width"] for i in items if i["original_width"]]
    sources = Counter(domain(i["link"]) for i in items)
    return {
        "images": len(items),
        "source_domains": len(sources),
        "top_sources": sources.most_common(5),
        "median_width": int(statistics.median(widths)),
        "share_1000px_plus": round(100 * sum(w >= 1000 for w in widths) / len(widths)),
    }


def is_image(url, timeout=12):
    """True if the URL answers 200 with an image content type."""
    try:
        resp = requests.get(url, timeout=timeout, stream=True, headers={"User-Agent": "Mozilla/5.0"})
        ok = resp.status_code == 200 and resp.headers.get("Content-Type", "").startswith("image/")
        resp.close()
        return ok
    except requests.RequestException:
        return False


def downloadable(images):
    """Drop hosts that serve a web page instead of the image, then check the rest."""
    candidates = [i for i in images if not (domain(i["original"]).startswith("lookaside."))]
    return [i for i in candidates if is_image(i["original"])]

On our 20 searches, image_report returned:

{'images': 2000, 'source_domains': 1266,
 'top_sources': [('instagram.com', 41), ('facebook.com', 16), ('etsy.com', 14),
                 ('tripadvisor.com', 14), ('magnific.com', 13)],
 'median_width': 1200, 'share_1000px_plus': 67}

is_image uses stream=True and closes the connection, so it reads the headers without downloading the whole file. Run it on a sample before a bulk job. The first six golden retriever results all passed in our test.

What the response does not tell you

Each result has the image URL, its size, a thumbnail, and the page it appears on, but no license. Images belong to their owners, and being in a search result does not make one free to reuse. Check the source page before you publish, sell or train on anything you download.

What to do with it

• Filter on size at the start. original_width and original_height are in the response, so drop small images before you fetch anything.
• Skip the unfetchable hosts. Drop lookaside.*, and expect roughly one in ten of the rest to fail anyway: in our sample, 9 of the 76 URLs that were not on those hosts did.
• Keep the source page. Store link next to every file, so you can check the licence and give credit later.
• Vary the query, not the page. The endpoint returns page 1 only, and a related query gives you a different 100.

What it costs

Every search is 1 credit, on every queue, and returns up to 100 images. Starter is $1.00 per 1,000 requests and the largest pack is $0.30 per 1,000, see pricing.

Failed requests are not charged. Downloading the files is your own bandwidth and not part of the credits.

Next steps

You now have numbers on where Google Images results come from and how large they are, a report you can run on your own searches, and a check that stops you wasting time on URLs that serve a page instead of a picture. The Google Images API page lists every parameter with a full example response, and Which sites AI Overviews cite does a similar count for AI Overviews.