Residential exits for targeted eval and RAG source diversity

When residential is the right tool for an AI workload

Most training-corpus collection is better served by datacenter — it's cheaper, faster, and most open-web sources don't care about origin ASN. Residential becomes the right tool in four narrower cases:

1. The source geoblocks cloud ASNs

A non-trivial fraction of the open web — regional press, some .gov, some enterprise knowledge bases — returns empty or degraded content to AWS/GCP/Azure subnets. For a RAG ingestion pipeline that needs consistent coverage, residential is the only viable answer.

2. Regional evaluation

Testing how a commercial LLM responds when the request appears to originate from Berlin vs. Riyadh vs. São Paulo is a methodology requirement for any safety or regional-bias eval. Residential provides the IP-layer authenticity these evals require. Datacenter IPs are trivially classified and some model APIs apply different content policy by origin region.

3. Training corpus diversity

Models trained predominantly on US-datacenter-visible web content inherit a specific compositional bias (heavy on enterprise/marketing long-form, under-weighted on regional news and community content). Supplementing with residential-sourced collection from the long tail of countries improves eval scores on multilingual benchmarks — modestly, but measurably.

4. Competitive AI intelligence

Scraping public model outputs (chat logs shared to forums, exported Gradio spaces, reasoning-model trace dumps) from residential IPs avoids the rate-limit behaviour most model-hosting platforms apply to bulk-cloud-origin requests.

Coverage and routing

• Pool size. 4M+ distinct exit IPs observed in rolling 30-day windows, varying by country. Depth is concentrated in the 10 focus markets rather than thinly spread across 190+ countries.
• Targeting. Country, region, city in 30+ metros across the 10 focus countries, ASN filtering on the top 6 carriers per country.
• Rotation. Per-request is the default for training corpus sampling. 10-minute sticky is typical for sequential eval loops where you want the same IP across a multi-turn conversation.

Clean provenance

The pool is built on an opt-in SDK integration — peers receive value (SDK features, rewards, ad-free tiers) in exchange for bandwidth when their device is idle. No stealth installations, no bundleware, no children's devices. For AI teams that publish their data sources and want to avoid the provenance problems that have surfaced with some training-data vendors, this matters.

Request shape still matters

Residential IP does not save a scraper with a bad request fingerprint. A realistic TLS fingerprint, plausible HTTP/2 headers, and timing that doesn't look like a tight loop are still required. See the blog post on eval-grade scrapers for a full reference stack.