Why IP Geolocation Accuracy Makes—or Breaks—Ad Tech in 2025

Why precise IP geolocation is mission-critical—and how 1% swings your budget

For geo-sensitive planning, bidding, pacing and compliance, IP geolocation is your most privacy-respectful signal at internet scale. When location is wrong, waste creeps in everywhere: audience eligibility, creative relevance, bid modifiers, frequency control and attribution all degrade—often invisibly.

A quick budget example (city-level targeting)

• Monthly geo-targeted media spend: US$1,000,000
• Average CPM: US$10
• Delivered impressions = (Budget ÷ CPM) × 1,000 = (1,000,000 ÷ 10) × 1,000 = 100,000,000 impressions
• City-qualified conversion uplift vs non-qualified: +20% (illustrative)
• AOV: US$80

If city-level accuracy improves by 1 percentage point (say 70% → 71%), that’s +1,000,000 impressions now served in the right geography. Even with conservative CTR/CR assumptions, that 1% shift compounds into meaningful incremental revenue—typically tens of thousands of US dollars per month and six figures annually at enterprise budgets.

Bottom line: 1% accuracy isn’t a rounding error. It’s money.

MaxMind (paid vs free): small deltas, big dollars

MaxMind’s own materials make clear that the paid GeoIP2 datasets/services are generally more accurate than the free GeoLite2. In many markets the uplift is single-digit percentage points rather than a huge leap—but those few points move millions of impressions at scale. With US$1,000,000/month at US$10 CPM (i.e., 100,000,000 impressions), a +3–5% accuracy gain shifts 3–5 million impressions into the correct geography. If city-qualified traffic converts even 20% better, you’re looking at six-figure annual ROI before changing anything else in the funnel.

What accuracy can IP geolocation theoretically deliver?

First principles: IP ≠ GPS

IP geolocation estimates where an IP serves users; it doesn’t track a device. The estimate is bounded by how the internet works:

• CGNAT / dynamic pools: Many users share a public IP; assignments change. “One IP = one person at one address” is false by design.
• Anycast & load-balanced egress: The same IP can be announced from multiple locations, steering clients to the nearest POP—not a single, static point.
• VPNs/proxies/Tor: Deliberately obscure the user’s true location; you see the exit node.
• Routing physics & policy: Paths are asymmetric; triangle-inequality violations and policy-based detours make latency a noisy proxy for distance.

For a non-technical backgrounder on limits and what “good” looks like, see How accurate can IP Geolocation get?

Country vs city—realistic expectations in 2025

• Country-level: With a well-engineered dataset, country identification on typical web audiences is very high—often 97–99%+. Edge cases (roaming, enterprise VPNs, anycast) exist.
• City-level:
  • Fixed broadband (more static): Often strong at city, sometimes suburb—though privacy-respecting providers cap to an area.
  • Dynamic consumer broadband: Generally good at city; suburb is situational.
  • Mobile/CGNAT: Frequently city/metro at best; sometimes region/state-scale when egress is centralised.

Applied where it actually matters: at the receiving end (focus on the source IP)

In ad tech, IP geolocation is almost always applied server-side—on the incoming request that reaches your systems. You geolocate the source of the connection (the end user’s egress: home router, CGNAT gateway, enterprise proxy or VPN exit).

• You do not need to locate destination infrastructure such as anycast edges, CDN POPs, web servers or routers. Those are primarily targets of client requests, not initiators. If a vendor’s headline strength is “identifying anycast locations”, that’s useful for network diagnostics—but not a buying advantage for audience localisation.
• What you do need is realistic handling of end-user egress: robust confidence areas (not just a point), network-type classification (mobile vs fixed vs hosting), and hazard/anonymiser signals to price, throttle or exclude accordingly.

Two macro trends that make city perfection unrealistic

1. Anonymisers are mainstream. Consumer VPN usage is now a normal part of the mix rather than a corner case.
2. Mobile dominates—and keeps growing. Mobile generates the majority of page requests in many regions, and global mobile data volume continues to surge. Mobile is exactly where CGNAT and centralised egress most constrain city-level precision.

How leading providers talk about “accuracy” (and how to read the claims)

Accuracy statements vary in what is measured and how it’s reported. Here’s a straight-talk comparison and a single, consolidated reality check so the argument reads cleanly.

Reality check

A realistic global traffic mix (why >95% city claims deserve scrutiny)

To ground expectations, consider an optimistic yet realistic worldwide mix of web requests:

• VPN / hosting: ~5% of traffic — by definition this segment has 0% true-city, since you see the exit node or a server location, not the user.
• Wired residential / business: ~35% of traffic — assume an optimistic 90% correct city rate.
• Mobile (CGNAT): ~60% of traffic — assume an optimistic 70% correct city rate.

Aggregate city hit-rate ≈ 0.05×0 + 0.35×0.90 + 0.60×0.70 = 73.5%.

In other words, a global city-level accuracy in the ~60–70%+ range is credible across mixed traffic. Claims of >95% city-level accuracy worldwide should trigger rigorous scrutiny of sampling, traffic mix, and methodology. For a deeper explanation of these limits, see our explainer: How accurate can IP geolocation get?

• Uniform “97% city-level globally” is implausible at true web scale. Today’s traffic mix includes substantial mobile/CGNAT, dynamic reassignment, and non-trivial VPN/proxy usage. Treat any global city figure as a marketing ceiling, not a planning baseline.
• Prefer methods you can encode into bidding. Accuracy radius (MaxMind) and confidence areas (BigDataCloud) make it straightforward to apply confidence-aware rules.
• Ask for ongoing, real-traffic comparisons. One-off case studies are easy; a daily comparative report is harder to fake and more relevant to your audience. 

How to choose your IP geolocation provider (and why probe/latency triangulation is overhyped)

Choose evidence over slogans

1. Transparent methodology: Prefer vendors who document how accuracy is measured. MaxMind’s accuracy-radius model and per-country charts are one approach; IPinfo explains its Probe Network; BigDataCloud publishes a daily provider-by-provider accuracy dashboard and returns confidence areas with each lookup.
2. Confidence-aware outputs: Radius/polygon + confidence lets you encode uncertainty in bidding and eligibility.
3. Network classification & hazards: City accuracy behaves differently on mobile vs fixed vs hosting. You also need anonymiser and hosting-likelihood signals to protect budgets.
4. Continuous refresh & explainability: The internet is dynamic; accuracy decays without frequent recalculation and explainable corrections.

Why latency triangulation gets oversold (IPinfo and similar claims)

Main point — Responders are often infrastructure, not people. Active probing primarily reaches routers, servers and other interfaces that are happy to respond; the vast majority of consumer IPs either sit behind NAT/CGNAT or drop unsolicited probes. That means you get good coverage of infrastructure geography, not user geography. For ad tech, that misses the brief: the job is to localise users’ egress, not to chart where routers live. Even if probe-based positioning were flawless on infrastructure, it still wouldn’t deliver the end-user geolocation you need for targeting, eligibility and budgeting.

• Latency inconsistencies: Internet paths are asymmetric and policy-driven; congestion and triangle-inequality violations distort RTTs—latency is a noisy proxy for distance.
• Anycast complications: A single IP can be announced from multiple POPs; probes reveal edges, not users, and results shift with routing.
• Dynamic and real-time issues: Consumer addresses move as pools are reassigned; mobile egress centralises and changes. Periodic probing can’t keep pace, leaving stale precision that looks confident but is wrong on the day.
• Blocking and intrusiveness: Many endpoints drop ICMP by default; clouds often block inbound ping unless explicitly allowed. Samples skew toward “friendly” networks.
• Non-response from most IPs: Firewalls, NAT/CGNAT and enterprise policies mean vast numbers of consumer IPs will never answer unsolicited pings—triangulation is impossible there.
• Regional performance gaps: Probe-based positioning underperforms where latency is volatile (rural, satellite, mobile)—precisely where address pools are most dynamic.
• Resource demands & coverage bias: Global probe fleets are costly; coverage clusters in easy places, biasing samples toward urban cores and popular ASNs.

Buyer’s checklist for ad-tech use cases

• Per-impression interpretability: Demand location + confidence (radius/polygon) + network type.
• Confidence-aware bidding: Apply stricter geo-eligibility for high-value campaigns; price lower-confidence traffic down.
• Segment by network: Treat mobile vs fixed vs hosting differently; consider separate CPA targets.
• Protect budgets with hazards: Filter/throttle VPN/proxy/Tor/hosting via a hazard signal (e.g., BigDataCloud’s Hazard Report API).
• Test on your audience: Run A/B tests using confidence areas rather than points.
• Monitor externally: Track vendor drift with an independent daily comparison.

Why BigDataCloud is the best choice today

• Patented, Proprietary Technology. Our geolocation engine isn’t a recombination of others’ data—it’s built entirely in‑house, powered by proprietary, automated algorithms and supported by U.S. Patent No. 11,792,110 B2 (and its Chinese counterpart). That means we’re not dependent on outdated third‑party databases—every update, every resolution is entirely ours.
• Accuracy you can verify daily. We don’t ask you to trust a slide. Our Daily IP Geolocation Accuracy Report compares providers using live, user-verified reference data—broken out by wired, cellular and hosting/mixed segments—so you can see how everyone performs on real audiences, any day you like.
• Evidence-based modelling with confidence areas. Our IP Address Geolocation with Confidence Area returns both a point and a polygon confidence area, plus a confidence value. Your bidding systems can finally optimise to uncertainty instead of ignoring it.
• Privacy-first by design. We deliver actionable geography without invasive signals, aligning with modern privacy expectations while retaining commercial utility. For a clear, non-technical tour of what’s realistic and why, read How accurate can IP Geolocation get?
• Full-stack risk intelligence. Pair location with our Hazard Report API to detect VPNs/proxies, Tor and hosting-likelihood. This protects budgets from low-quality or policy-violating traffic before it drains spend.
• Transparent pricing and scale. From free to enterprise, with optional pay-as-you-go overage and no artificial throttle at the top end—see the IP Geolocation package page.

Your next steps

1. Audit your geo layer: Replace single-point geos with confidence-aware rules; add hazards to pricing/eligibility.
2. Run a head-to-head: On a live campaign, compare your incumbent against BigDataCloud. Track eligible impressions, CPA/ROAS and wasted geo-impressions.
3. Watch the dials move: If you see even +1–3% more impressions correctly delivered to the target city, you’re looking at material ROAS lift at enterprise budgets.