Every rate has a tier
When a rate comes in, we label it by how it was proven. The label decides how much it counts.
| Tier | How a rate gets there | Counts as |
|---|---|---|
| Verified | You uploaded a policy declarations page (the summary page of your policy) or renewal offer, every check passed, and you changed the premium by 2% or less. | 1 |
| Reviewed | It failed an automatic check, and a person looked at it and approved it. | 1 |
| Premium verified | A bill or escrow statement proves the price, but the coverage was typed in. | 1 for price, ½ for coverage |
| Shared | Typed in, with no document. | ¼ |
| Public | State regulators, the Census Bureau, FEMA and the U.S. Treasury. | The starting point |
The weighting, in plain words
- We start from the public number. For each area, the best public figure (a state regulator's average where there is one, otherwise the Census survey) is the starting point. It counts as if it were 5 policies. Where there's no public figure, an area borrows from the area around it: a ZIP code from its county, a county from its state.
- Verified rates pull hardest. Each one counts in full.
- Shared rates nudge. Each one counts a quarter as much as a verified rate.
- Shared rates have a ceiling. In any area, all the shared rates together can never count for more than the verified rates plus the public starting point. A flood of made-up numbers can't move a figure far.
- We work with the shape of real prices. Insurance prices are lopsided, with a long tail of very expensive policies, so we do the math in a way that keeps a few huge bills from dragging the typical figure around, and we set aside extreme outliers.
- We correct for habits. If people who type their rate in tend to round down, and the verified rates show it, we adjust shared rates by that amount.
These weights are a starting point. Each month, in areas with enough of both, we compare shared rates with verified ones, measure how far apart they run, and update the weights and corrections. Every change is logged and posted here. We don't go back and restate old figures.
What you'll see on a figure
| When an area has | We show | Labeled like |
|---|---|---|
| 10 or more verified policies | The verified figure only | "Based on 23 verified policies" |
| Fewer than 10 verified, but enough data counting everything | A blended estimate | "Estimate from 6 verified policies and 31 shared rates" |
| Less than that | The public figure, or nothing | Its source and year |
Headline numbers we give to the press use verified figures or public data only, never blended estimates.
The checks
An AI model reads your document: Anthropic's Claude, through its API. Anthropic may keep what it's sent for a limited time to monitor for abuse, and we're requesting zero data retention. The model never does the math. Plain code checks that:
- the line-item premiums add up to the total;
- the coverages make sense together, and match the policy form;
- the policy period is a normal length and ended no more than 18 months ago;
- the ZIP code matches the one you gave.
We also look for documents that were edited, the same file or the same policy sent twice, and prices far outside the normal range for the county. Anything that doesn't line up goes to a person.
Deletion
We read your document, keep the numbers, and delete the document within 24 hours. You can delete it sooner. We keep it longer only if you choose to store it in your vault, if you separately let us keep it for up to 2 years to improve how we read documents, or while it waits for a person to review it, for no more than 7 days. Read the privacy policy for the details.
What we publish
Totals and ranges for areas, never anything about one home, household, driver or policy. Crowd figures stay hidden until enough policies are behind them, and nothing from shared rates is published below the ZIP code level. Insurance company names appear only next to a regulator's own figures, never in our crowd data.
Want the longer story? Read how we verify a rate, and why we delete your document.