Methodology

How summaries and ratings on AIgree are produced, and how to flag a mistake.

1. Sources

We only work from publicly available legal documents published by each service on its own website — Terms of Service, Privacy Policy, and, where applicable, Cookie Policy. The source URL for every document is linked on the service page so you can compare our summary against the original.

We do not scrape user data, member-only content, or anything behind authentication.

2. Fetching & update cadence

• Every tracked document is fetched once a day using a headless browser.
• Extracted text is hashed. If the hash is unchanged, nothing else runs — no new summary, no AI call.
• When the hash changes, we re-summarize the document and produce a change summary that describes what was added, removed, or clarified compared to the previous version.
• The history of every change is visible on each service's History tab, including the date and a severity shift indicator.

3. The AI pipeline

Summaries are produced by a large language model (currently an OpenAI model) driven through three stages. Each stage has a fixed system prompt and a strict JSON schema — the model is not allowed to emit free-form prose outside the schema, which makes the output predictable and auditable.

• Stage 1 — Per-document summary. Each legal document is reduced to a list of jargon-free bullet points covering the key user-facing terms.
• Stage 2 — Points of interest & overall rating. The per-document bullets are combined to surface the clauses that matter most to users (favorable or unfavorable), each tagged with a kind (positive or negative) and a severity score, plus an overall 1–5 rating for the service.
• Stage 3 — Change summary. When a document changes, we diff the new and previous versions and produce a single-sentence change note, a severity shift, and a bullet-level delta list.

Every AI call is retried on transient errors, validated against its schema, and logged with the token usage. If a document genuinely cannot be parsed, we preserve the previous summary rather than publish a degraded one.

4. Severity scale

• 1 — minor or informational.
• 2 — low impact on user rights.
• 3 — moderate impact, worth being aware of.
• 4 — high impact (e.g. extensive data collection, auto-renewal without easy cancellation, mandatory arbitration).
• 5 — critical impact on user rights.

A clause can be either positive (user-favorable, e.g. clear data-export rights, short retention, granular consent) or negative (user-unfavorable). Colors on the site follow the same convention — green for positive, orange for moderate concern, red for critical.

5. Overall rating

The 1–5 overall rating is the model's synthesis of all the points of interest it surfaced, weighted by severity. A short rationale is shown next to the rating on every service page so you can see why that score was assigned. The rating does not include factors outside the documents themselves (security incidents, press coverage, regulatory fines) — it reflects the text of the policies, nothing more.

6. Limitations

• Summaries and ratings are AI-generated opinions. They may be incomplete, out of date, or simply wrong.
• The legal effect of any contractual clause depends on the applicable jurisdiction, the other party, and the context. A term that is enforceable in one country may be void in another.
• AIgree is not legal advice. Before relying on any clause for a decision that matters, read the source document and, if the stakes are high, ask a lawyer.
• We summarize what the documents say, not what the company does. A clean Privacy Policy does not prove clean practices.

7. Right of reply for listed services

If you represent a service listed on AIgree and believe our summary, a specific point of interest, or the overall rating misrepresents your Terms or Privacy Policy, please reach out:

[email protected]

Include the service name, the specific item you are flagging, the exact quote from your document that you believe we mischaracterized, and a short explanation. We review every request within 7 business days and respond with one of:

• An update to the summary, point of interest, or rating, if we agree we got it wrong.
• An explanation of why we stand by the current version, with reference to the source text.
• An annotation on the page noting the dispute, where the facts are genuinely ambiguous.

We do not charge for corrections, we do not accept payment to change a rating, and we do not remove listings on request — services with critical low ratings are precisely the ones users need to see.

8. Independence

AIgree is an independent service. We have no affiliation or partnership with any of the services we review. We do not accept money, promotional consideration, or any other benefit to change a summary or a rating. Service names and logos shown on this site belong to their respective owners and are used solely to identify the services under review — no affiliation or endorsement is implied.

9. Reporting issues & suggesting services

Spotted something wrong, missing, or misleading on a service page? Use the Report a problem with this summary button on that page — your report goes straight to our moderation queue, anonymously by default, with an optional email if you'd like a reply.

Don't see a service we should cover? Search for its name on the homepage — if we don't track it yet, a Suggest this service button appears right there.

For anything else — press enquiries, accessibility issues, takedown requests — contact us.