Most Google Search ads in the Transparency Center are screenshots. Here is how I get the text out.
The Google Ads Transparency Center (adstransparency.google.com) shows every ad an advertiser has run, by region, format and date. For competitor research it is the best free source there is.
The problem shows up the moment you try to export it: for most Search ads, Google does not have the ad text as text. It has a picture of the ad.
This post covers how the Transparency Center serves its data, why you end up with screenshots, and the two-step extraction I use to get headline, description, display URL and sitelinks as fields. I built this into an Apify actor, linked at the end. Disclosure: I am the developer.
The site runs on a small RPC API
The Transparency Center is a JavaScript app that talks to one base URL:
POST https://adstransparency.google.com/anji/_/rpc/<Service>/<Method>
The methods you need:
SearchService/SearchSuggestions: resolve a name to advertiser IDs.SearchService/SearchCreatives: list ads for an advertiser or a domain, with region, format and date filters. About 100 ads per call, with a cursor for the next page.LookupService/GetCreativeById: details for one ad, including all its variations and the regions it was shown in.
The bodies are protobuf-style JSON with numbered keys instead of names ("1" is the advertiser ID, "2" the creative ID, and so on), sent as a form field. No login, no cookies, no signed token. The field layout has been stable for years; every open-source client I looked at uses the same numbers.
One operational note: Google blocks this endpoint from many cloud IPs. From a home connection it just works. From a server you will likely need a residential IP for the search calls. The image and preview hosts do not have that restriction.
Each ad has a preview, and the preview is the problem
A creative in SearchCreatives comes with a preview in one of two shapes:
- A
content.jsURL ondisplayads-formats.googleusercontent.com. This is a script that renders the ad, and the headline and description strings are inside it. - An
<img>pointing attpc.googlesyndication.com/archive/simgad/.... A PNG of the rendered ad.
For image ads, the PNG is the ad, so that is fine. For video ads, the content.js contains the YouTube thumbnail URL, so you can regex out the video ID and build a YouTube link. For text (Search) ads, you want shape 1, and you mostly get shape 2.
In my samples:
- GEICO: 39 of 40 text ads had only a screenshot.
- Nike: 43 of 50 ads had no structured text.
So any tool that only reads the preview script returns metadata for most Search ads: advertiser, dates, format, a link. That matches complaints I found in reviews of existing scrapers along the lines of "only returns metadata."
Step 1: read the preview data when it exists
When a creative has a content.js preview, fetch it and pull the strings out. It is a big script (around 150 KB for video ads, smaller for text), but the ad text appears as plain string literals. This path is exact and cheap, so it always runs first.
Step 2: OCR the screenshot
The screenshots are good OCR material: Google renders them at a consistent size, on a white background, in a standard font. I run Tesseract on them (English and Spanish, plus the region's main language), which takes about 0.3 to 0.5 seconds per image.
Raw OCR gives you lines, not fields. A rendered Search ad has a predictable layout, so I split lines by position and size:
- The first line is the advertiser name row. Skip it.
- Line 2 or 3 that looks like a domain (
www.geico.com/) is the display URL. - Lines labeled
"Sponsored","Ad","Anuncio"or"Anzeige"are skipped. - The headline is the first block of lines with the largest text height. A drop in line height to under 70% of the headline's, or a vertical gap, starts the description.
- A larger gap after the description starts the tail, where short lines (60 characters or less) are sitelinks.
- Footer lines like
"Top pages"or"Rating for..."end the ad.
Then a quality gate: lines under 60% confidence are dropped, and if the mean confidence of what is left is under 80%, the whole result is discarded. Local map ads and ads with a product photo overlapping the text produce noise, and returning garbage as "ad copy" is worse than returning nothing.
Here is a GEICO text ad that only exists as a screenshot, after OCR:
{
"advertiserName": "GEICO Corporation",
"format": "text",
"headline": "GEICO® Car Insurance - Get a Free GEICO® Quote Today",
"description": "GEICO® was Named the Best Car Insurance Company for 2026-2027. Get a Free Quote Today.",
"displayUrl": "www.geico.com/",
"adTextSource": "ocr",
"ocrConfidence": 95
}
I keep adTextSource (preview_data or ocr) and ocrConfidence on every row, so you can decide how much to trust each one.
A cheaper question: does this domain run Google Ads at all?
A lot of people do not need ad copy. They have a list of 5,000 prospect domains and want to know which ones advertise on Google. That is one SearchCreatives call per domain with the domain filter.
Google returns an approximate ad count range (for example 10,000 to 20,000 for shopify.com) and the ads themselves, from which you get the last shown date and formats. One row per domain: runsGoogleAds, the count range, activeLast30Days, formats seen. It is a useful lead qualification signal for agencies and SaaS sales teams.
Watching competitors over time
Every creative has a first shown and last shown date. If you store the set of creative IDs per advertiser and run daily, you can output only what changed: new ads, and stopped ads whose last shown date is more than a few days old.
Be careful with partial fetches. If you only pulled 200 of an advertiser's 2,000 ads, a missing ad means nothing, so only judge "stopped" from complete fetches.
The hosted version
The actor is Google Ads Transparency Scraper: Ad Text, Domain Check on Apify. It does the RPC calls (through residential proxy by default, at no extra charge), preview data first, OCR second, YouTube links for video ads, the domain check mode and the new/stopped monitor.
Pricing per 1,000: $1.20 per ad row, plus $1.00 per ad where text was actually found, and $2.00 per domain check. OCR attempts that find nothing are not charged. 1,000 ads from a Search-heavy advertiser with text found on 75% of them comes to about $1.95.
from apify_client import ApifyClient
client = ApifyClient("YOUR_APIFY_TOKEN")
# 1) ad copy for a competitor's Search ads
run = client.actor("zrk1212/google-ads-transparency-ad-copy").call(
run_input={
"mode": "ads",
"domains": ["geico.com"],
"region": "US",
"adFormats": ["text"],
"maxAdsPerTarget": 100,
"includeAdText": True,
}
)
for ad in client.dataset(run["defaultDatasetId"]).iterate_items():
print(ad.get("adTextSource"), "|", ad.get("headline"))
# 2) which prospects run Google Ads
run = client.actor("zrk1212/google-ads-transparency-ad-copy").call(
run_input={
"mode": "domainCheck",
"domains": ["shopify.com", "geico.com", "example-small-shop.com"],
"region": "US",
}
)
for row in client.dataset(run["defaultDatasetId"]).iterate_items():
print(row["domain"], row["runsGoogleAds"], row.get("adCountMin"), row.get("adCountMax"))
Limits worth knowing: the date filter only reaches about the last 12 months of shown history, ad counts are ranges because that is all Google publishes, and text hidden behind an ad's own product image cannot be recovered. The free Apify plan allows OCR on up to 200 ads per run.
This is public transparency data that Google publishes for accountability. The actor is not affiliated with Google, does not log in, and does not solve CAPTCHAs.
Comments
No comments yet. Start the discussion.