Ad Verification Proxies

Ad verification fails quietly. You open the campaign URL from the office, the creative renders, the landing page loads, the spreadsheet says "live in all markets". Then a customer in another city sends a screenshot of a competitor's ad in the slot you paid for, or a blank page, or a consent wall that never offers a way through. The check passed because it was never a check: one IP, one city, one connection type, one cookie state.

What you are actually verifying is not the creative. It is the delivery decision — and that decision is made from three signals you can only reproduce with the right exit.

Three layers decide what renders

The first layer is the one teams handle. The second is where most budgets leak: programmatic inventory is priced and allocated by metro, so a country-level check cannot see a campaign that under-delivers in one city and over-delivers in another. The third layer is the one that produces the "it works for me" argument — a fresh profile with no consent cookie sees a different page from a returning visitor, and both are legitimate.

Match the exit to the verification task

Note the pattern: the more the task depends on being treated as a person, the more it needs a residential exit and a stable session. Only the first row is safe on datacenter, and even there the platform may serve a stripped page.

Keep one exit for the whole chain

An ad check is a sequence, not a request. Search or feed → impression → click → redirect → landing page → CTA. Frequency capping, A/B assignment and geo-redirects are all session-scoped. Rotate in the middle and every downstream measurement is taken against a different user than the one who saw the ad. In practice that produces the two most confusing reports in the discipline: "creative matched but landing page 404" and "ad served in a city we did not target".

# One verification pass = one sticky exit for 20 minutes
# Gateway form: country-city-session-ttl
PROXY = "http://user-country-de-city-berlin-session-8f21a:pass@gate.iplight.net:8000"
# session id stays fixed; ttl keeps the same exit alive across requests

# Playwright: drive the real chain, capture what rendered
from playwright.sync_api import sync_playwright

with sync_playwright() as p:
    br = p.chromium.launch(proxy={"server": "http://gate.iplight.net:8000",
                                 "username": "user-country-de-city-berlin-session-8f21a",
                                 "password": PASS})
    ctx = br.new_context(locale="de-DE", timezone_id="Europe/Berlin",
                         viewport={"width": 390, "height": 844},
                         is_mobile=True)
    pg = ctx.new_page()
    pg.goto(CAMPAIGN_URL, wait_until="networkidle", timeout=45000)

    # 1. did we leave the domain (redirect chain changed the market)?
    print("final", pg.url)
    # 2. does the geo layer agree with the exit?
    print("geo", pg.evaluate("fetch('/cdn-cgi/trace').then(r=>r.text())"))
    # 3. evidence, one per city per day
    pg.screenshot(path="de-berlin-2026-11-03.png", full_page=True)
    br.close()

Two details in that snippet do more work than the proxy itself: locale/timezone_id matching the exit, and is_mobile. A German exit with an English locale and a desktop viewport gets a different layout and often a different creative, and you will chase a "missing ad" that was a viewport mismatch.

Measure the things that can actually be wrong

Geo mismatch is the metric that pays for the programme. If the exit says Berlin and the page's own geo layer says Amsterdam, your targeting is leaking spend into a market you did not buy — and that is a budget conversation, not a technical one. Track it per city, weekly, and put the number in the same deck as delivery.

What a programme costs

The arithmetic: 200 placements × 8 cities × 2 passes a month = 3,200 page loads, at an average of 700 KB with images blocked, is about 2.2 GB a month, so roughly $2.20 at $1/GB. Even at 10x that volume the proxy line stays smaller than the screenshot storage and the engineer's time. Treat traffic cost as noise and spend the effort on coverage: more cities, more device profiles, more consent states.

Six failure modes worth pre-empting

• Country-level checks only. Buys and delivery are metro-level; a national pass reports healthy while a city is dark.
• Rotating mid-chain. Attribution, frequency caps and A/B buckets all reset. Sticky, or the number is meaningless.
• Desktop-only viewport. Most social and app-install inventory is served to mobile; a desktop pass cannot see it.
• Fresh cookie jar every run. You verify the first-visit experience forever and never the returning one, which is the one most customers have.
• No consent-state variants. Accept, reject and undecided produce three different pages in the EU; two of them are what most visitors see.
• Verifying competitors by clicking. That is click fraud, not verification, and it costs the other side real money.

Frequently asked questions