Methodology

Every number on this site should be traceable to a source you can check. This page is the contract: where ratings come from, how the benchmark works, what gets recorded on timelines, and the conflict of interest we disclose instead of hiding.

Where ratings come from today

Until our own benchmark data is published, app ratings are derived from public sources in this priority order:

  1. iOS App Store rating (primary, US storefront)
  2. Google Play Store rating (fallback)
  3. Trustpilot, G2, or a similar aggregator
  4. Average of two or more independent review sites

The source used is named on each app's page. We never invent a score, and we never claim hands-on testing that hasn't happened. General-purpose LLMs (ChatGPT, Claude, Gemini, Grok, DeepSeek, Pi) are listed unranked, because no app-store rating measures how good they are as companions.

The memory benchmark (in progress)

Memory is the axis this category markets hardest and measures least. We are building a fixed, repeatable, published test: the same facts planted in conversation with every app, then probed at increasing distance (roughly 10, 100, 500, and 1,000 messages back) and across session boundaries. Each planted fact is scored as recalled, forgotten, corrupted, or leaked; probes are repeated to account for randomness, and each app runs at its best available memory configuration, disclosed.

Three commitments when scores go live:

  • Every published score links to the raw test run.
  • Every score states the methodology version it was produced under (v1, v2, ...), and old scores are re-run when apps change.
  • The full test design is public, so anyone, including the apps we test, can reproduce it.

Until a real run exists for an app, its page says "not yet benchmarked." No exceptions.

How the change tracker works

Companion apps mutate: prices, models, content policies, free-tier limits, ownership. Each recorded change is a dated entry with a type (pricing, model, policy, feature, legal, shutdown, company) and, wherever possible, a source link. Entries not yet verified against their sources are labeled "being verified." The tracker updates continuously and rolls up into a monthly changelog email.

The conflict of interest, disclosed

CompanionHunt is built by the maker of Softly, an AI companion app for iOS. That is exactly the kind of conflict most review sites hide. We handle it in the open instead:

  • No paid placement, ever. No app can pay for a ranking, a score, or a mention, and rankings are never influenced by affiliate status.
  • Softly gets no special treatment. When benchmark scores exist, Softly is ranked exactly where the data places it, with the same raw run published.
  • The methodology is reproducible. You do not have to trust us; you can re-run the test.

Corrections and update cadence

Facts (pricing, platforms, policies) are re-verified on a published cadence and whenever a change entry lands. The site-wide "last updated" date is in the footer; each app page shows when its facts were last current. Something wrong? Tell us and we will fix it and record the correction: see about for contact.

Frequently asked questions

Are CompanionHunt's rankings influenced by affiliate links?

No. Rankings are never influenced by affiliate status or payment of any kind. Where an affiliate relationship exists it is disclosed on the affected page; currently there are none.

Does CompanionHunt do hands-on testing?

Our hands-on memory benchmark is in progress. Until a real run exists for an app, its page shows a 'not yet benchmarked' state, and current ratings are clearly attributed to public sources like App Store aggregates.

Who runs CompanionHunt?

CompanionHunt is built by the maker of Softly, an AI companion app. The conflict is disclosed openly: no paid placement, Softly is never ranked above where data places it, and the benchmark methodology is public and reproducible.

How often is the data updated?

The change tracker updates continuously as events are recorded and verified, and rolls up into a monthly changelog. App facts are re-verified on a published cadence.