One Product, Forty Prices: A Methods Paper on Measuring Geo-Pricing Honestly
Regional price discrimination is one of the most studied and worst-measured phenomena in e-commerce — because the researchers who study it usually collect from one IP, one VPN exit, one moment, and call it data. This methods paper is for market-research and consumer-advocacy teams who want geo-pricing measurements that survive peer review: sampling design, collection fidelity, and the cost of doing it right, which turns out to be a few hundred dollars.
The Measurement Problem
Geo-pricing claims follow a familiar arc: a screenshot from one country, a cheaper screenshot from another, a headline. Then the rebuttal: “that was currency conversion,” “that was a sale,” “that was one data point.” Both sides are arguing about anecdotes because the underlying measurement is hard, and it’s hard in four specific ways:
- Price varies by more than geography. Currency, tax display rules, local promotions, device class, and even time of day all move the number. A study that doesn’t hold these constant measures everything at once.
- One observation per market is noise. Prices A/B test now; the same market can serve different prices to different sessions on the same day.
- Collection method contaminates the result. A datacenter IP, a VPN exit shared with thousands of researchers, a residential IP from the wrong city — each gets a different treatment, and the datacenter version often gets a different *price*, not just a different page.
- Markets drift. A monthly screenshot per country measures last month’s pricing policy with this month’s error.
The Methods Section We’d Want to See
A defensible geo-pricing study, section by section:
Sampling frame. Define the markets (not “40 countries” but which 40, and why), the products (a fixed basket), and the observation window. Pre-register the basket before collection — post-hoc basket selection is where most bias enters.
Collection fidelity. Each observation must come from a residential IP pinned to the target market, at city level where the retailer prices by metro, with device held constant. This is the section where the study lives or dies: shared VPN exits and datacenter IPs are measurement instruments that change the reading. Thordata’s residential network is the kind of instrument this section needs — 100M+ IPs across 190+ countries, country/state/city/ASN targeting at no extra charge, sticky sessions for cart-flow measurements, on a published slider from $2.00/GB down to $0.65/GB at volume. For a research budget, that last part matters as much as the fidelity: methodology that costs thousands per study doesn’t get replicated.
Replication. N observations per market per product per week, minimum — enough to see A/B splits as *distributions* rather than being fooled by them as outliers. Weekly cadence over the study window, timestamped, so drift is a finding instead of a confound.
Normalization. Currency conversion at the observation timestamp, tax-inclusive vs. tax-exclusive handling declared, promotion flags captured rather than filtered out — a flash sale is data about pricing strategy, not noise to discard.
The Instrumented Version
import requests
from datetime import datetime
# One endpoint per market, generated with city-level pinning
MARKETS = {
"DE-Berlin": "http://USER-geo-de-berlin:PASS@res.thordata.com:18000",
"BR-SaoPaulo":"http://USER-geo-br-saopaulo:PASS@res.thordata.com:18000",
"JP-Tokyo": "http://USER-geo-jp-tokyo:PASS@res.thordata.com:18000",
}
def observe(product_url, market, proxy):
r = requests.get(product_url, proxies={"http": proxy}, timeout=30)
page = parse(r.text)
return {
"market": market,
"price": page.price, "currency": page.currency,
"tax_included": page.tax_included,
"promo_flag": page.on_sale, # promotions are findings, not noise
"observed_at": datetime.utcnow().isoformat(),
}
# N replicates per market per week, stored raw — normalize downstream, never at capture
The comment on the last line is the methodology in miniature: capture raw and normalize in analysis, because a capture-time normalization decision can’t be audited later.
What a Real Study Costs
| Component | Scale | Rough cost |
|---|---|---|
| Residential collection | 20 products × 10 markets × 5 obs/week × 8 weeks | ~40 GB ≈ $60–100 on the slider |
| Cart-flow price checks (sticky sessions) | 20% of observations | Included above |
| Structured marketplace pricing | Via Scraper API where targets are pre-built | ~$0.50–$1.00 per 1,000 results |
| Search-side price signal (shopping placements) | Via SERP monitoring | ~$0.70 per 1,000 responses at volume |
The total lands in the low hundreds of dollars — which is the point of the methods framing: a defensible geo-pricing study costs less than the coffee budget of the meeting where someone proposes the screenshot version. The search-side line deserves emphasis: shopping placements and localized ads are where pricing strategy becomes visible *before* the product page, and continuous SERP data crawling captures that layer with the same geo fidelity as the page itself.
The FAQ Every Reviewer Asks
Why residential IPs specifically — isn’t that expensive for research?
It’s the fidelity requirement, not a luxury: geo-pricing detection systems treat connection class as an input, so a datacenter observation measures “the price for datacenter traffic,” which is usually not the research question. At published slider rates, the fidelity costs less than the error it removes.
How do you distinguish geo-pricing from legitimate cost differences?
You often can’t from the price alone — which is why the methods section captures tax display, shipping anchors, and promotion flags: the claim that survives is “observed price differs by X% after normalization,” and the interpretation section is where cost-based explanations get weighed. Studies that skip the capture of those fields can’t have that conversation honestly.
What about ethics and terms of service?
Research collection at polite frequency from public product pages, with rate discipline built into the managed layer, sits in the same posture as a comparison-shopper visiting manually — but each team should run its own review, minimize personal data capture (prices and pages, not sessions), and publish methodology openly. Replicability is the ethical argument: a methods section others can rerun is a study others can trust.
What’s the minimum publishable design?
Ten products, five markets, three observations weekly, four weeks — a few dozen GB and a month of patience. Start there, publish the methods even if the price findings are boring, and the boring baseline is what makes the next study’s interesting finding credible. For the search-side layer, SERP monitoring pilots on free trial credits, geo-pinned to the same five markets, so the page data and the query data describe the same world.
The one-sentence abstract of this whole methods paper: geo-pricing is measurable, the instruments are rentable at research-budget prices, and the difference between a screenshot and a study is nothing but sampling discipline — plus the willingness to let the data say “no significant difference,” which, inconveniently, is what a defensible design finds about half the time.