Data philosophy

Built for retrieval, not for scrolling

Agent Retrieval Index treats machine readers — search engines, AI answer systems, retrieval bots, and agents — as first-class users. Humans get the same pages, and benefit from the same discipline.

The thesis

An increasing share of product research is done by software on a person's behalf: an AI assistant answering "which power station should I buy for home backup," a search engine composing an answer box, an agent filling a procurement spreadsheet. Those systems don't need persuasive prose — they need identifiable entities, normalized attributes, explicit verdicts, timestamps, and sources they can verify. That is what this index publishes.

How this differs from a review blog

DimensionTypical review blogAgent Retrieval Index
Unit of content An article optimized for a keyword An entity (product, comparison, category) with a stable URL
Specifications Prose mentions, inconsistent units Normalized attribute tables, same keys and units across brands
Recommendations "Top 10 best X" rankings Explicit verdicts plus scenario → product mappings
Freshness Opaque; often years stale Visible last-updated dates on every entity and in the sitemap
Machine access Incidental; sometimes bot-blocked Deliberate: crawlable HTML, machine summary blocks, open robots.txt
Monetization Affiliate links shape the content Affiliate/merchant links are a clearly-labelled secondary layer; the facts layer is merchant-agnostic

Commitments to machine readers

  • All core content is server-rendered HTML — nothing important requires JavaScript to read.
  • robots.txt allows crawling broadly by default, including AI and agent crawlers.
  • URLs are stable identifiers; entities are added, not silently rewritten into different things.
  • Product and comparison pages open with a visible key-value machine summary block.
  • Every entity states when it was last updated and where its facts came from.
  • Semantic markup throughout: real headings, real tables, real definition lists.

Product pages also carry a clearly-separated Where to buy block listing retailer offers, some of them affiliate links. That commerce layer sits on top of the facts layer and never feeds it: retailer listings are not a source for any normalized specification, and prices are hand-entered with a checked-on date and withheld once stale — see the affiliate disclosure.

The methodology page documents the collection and normalization rules; safety and editorial standards document the limits we hold ourselves to; AI & research use and llms.txt tell machine readers how to use the site. Structured JSON exports are live for every collection: products, generators, solar-kits, categories, comparisons, use-cases, and spokes — plus per-entity JSON at {entity-url minus trailing slash}.json; every machine summary block links to its own.

Current scope

The index covers US-market portable power across three product families, each with its own schema, source policy, and normalization rules — this index never pretends watt-hours, gallons, and panel wattage are the same property:

  • Battery power stations — 34 indexed. Sealed, indoor-safe, rechargeable.
  • Portable generators — 14 indexed. Fuel-burning inverter and conventional units; outdoor-only, with stricter safety-field rules (a model with ambiguous carbon-monoxide-shutdown documentation is not published at all).
  • RV solar kits — 11 indexed. Panel kits that recharge a battery bank; charging systems, not standalone power sources.

Above the families sits the Portable Power hub, with 10 cross-category use-case guides that weigh the three families against each other on shared decision axes (indoor safety, fumes, noise, output, runtime model, setup) — grouped by category, never merged into one ranking. Alongside them, 7 single-category use-case pages rank power stations in depth, and 11 comparisons set same-category products against the same attribute keys.

Standalone base models only. Every core specification is normalized from official manufacturer sources with field-level provenance (raw claim text, source URL, date checked, confidence), and records that fail the publishability gate are not published — the build itself refuses them. Coverage expands family by family.