Methodology
This page explains how AI Deals verifies data, orders access options, computes objective champions, and separates provider facts from platform calculations and community reports.
1. No overall Deal Score
We do not combine price, free allowance, privacy, speed and reliability into one 0–100 score. Fixed weights would impose a hidden utility function on users.
2. AccessValueGroup
- Permanent unlimited free
- Permanent recurring free
- Other no-expiry free
- Temporary unlimited free
- Temporary metered free
- One-time free credit
- Paid
Numeric sorting only happens when units are genuinely comparable. For example, 10M tokens/month can be compared with 5M tokens/month; 100 prompts/day is not forcibly converted into tokens/month. Public Deal views do not use the migration-era legacy comparableMetric by default; without a trusted typed-derived metric they keep only objective grouping and deterministic order.
Price and Quota objective champions require two positive verification timestamps: the typed fact row itself must have last_verified_at, and the exact fact must also have OFFICIAL/PROVIDER Primary Evidence with its own last_verified_at. A verified parent AccessOption, evidence-only freshness, or fact-only freshness is not sufficient.
A Quota proves how much allowance is included; it does not prove that access is free. Stable free access requires an ACTIVE FREE_TIER Plan with its own positive last_verified_at and exact OFFICIAL/PROVIDER Primary Evidence. Quota evidence then independently proves allowance size. A paid subscription with included monthly usage does not become a free option or free-allowance champion.
A stable Plan is established only when official or Provider evidence explicitly identifies the Plan and its type. AI Deals does not infer FREE_TIER from Quota, noExpiry, a `$0` usage component, Deal type or marketing language. Those signals may trigger further verification, but they are not sufficient proof of a stable free tier.
Quota.noExpiry only means the allowance has no stated expiry; it does not by itself prove “free forever.” A one-time ONCE allowance remains one-time even if its balance does not expire. Finite ongoing allowances must also have no duration or unresolved condition and are compared only within the same metric × reset period × credit-currency qualifier. Provider-specific credits without an explicit currency are not compared across providers.
Best Free AI separates stable FREE_TIER access options from current free Promotions. They are not forced into one blended overall ranking.
3. Pareto dominance
If option B is no better than option A on every known comparable dimension and worse on at least one, B may be collapsed into Other options. Any meaningful trade-off keeps both options visible. Price contributes to Pareto only when price type, currency, unit and price-region qualifier all match; otherwise price is incomparable.
4. Verification evidence hierarchy
Third-party lists and community posts can discover candidates, but cannot independently create Verified facts. AI is a fallback interpreter when deterministic extraction is insufficient.
Use-case capability filters also require fact-level evidence. Any CapabilityAssignment that changes eligibility must itself have positive OFFICIAL/PROVIDER Primary Evidence and its own capability last_verified_at. A verified parent AccessOption cannot substitute for capability evidence, and discovery-only SUPPORTED or UNSUPPORTED rows cannot qualify or disqualify an option.
Model intent pages do not borrow parent verification either. /pricing uses only current PriceComponents with both fact-row and evidence verification. /free additionally requires the access path to belong to an evidence-backed ACTIVE FREE_TIER Plan; Quota alone cannot prove free access. The Quota still needs its own fact-and-evidence verification to be shown as a verified allowance. Legacy currentPrice/freeAllowance cannot create “currently verified” eligibility, and the SEO candidate builder uses the same conditions.
Plan-owned PriceComponent and Quota rows are valid stable commercial facts. Public comparison reads project those Plan-owned facts onto related AccessOptions while preserving exact Plan/Price/Quota provenance; facts are not duplicated merely for display.
5. Checked ≠ Verified ≠ Changed
- last_checked_at — last attempted check
- last_verified_at — last positive verification evidence
- last_changed_at — last material business change
A 429, 403, 5xx or timeout only means the check failed; it is not evidence that an offer expired. Fact categories also cannot inherit the parent AccessOption last_verified_at and present it as their own freshness.
6. Data provenance labels
- Provider data — sourced from official/provider evidence
- Calculated by AI Deals — derived from structured facts, e.g. Cheapest
- Community reported — real-world Worked / Didn’t work reports
- Sponsored — paid placement that cannot change natural ranking
7. Fastest / Most Reliable
Fastest and Most Reliable require real comparable measurements for the same model, access type, region and time window with a minimum sample size. Provider marketing claims are not benchmark evidence.
8. Commercial separation
Referral revenue, affiliate commission and sponsored bids never enter AccessOptionRanking, Pareto logic or Objective Champion calculations.