How the order is set

What a Loan Comps search asks

Amount, purpose, kind of loan, property or business type, number of units, place, and an optional loan to value. Nothing else is accepted. No name, Social Security number, tax ID, credit score, income, revenue or account number can be entered, by a person or an AI tool.

The whole deal lives in the URL. It is not saved against anyone unless you save a list yourself.

What "a loan like this one" means

Five things are checked against every lender's recorded loans:

• Kind. The same lane of loan. Multifamily leaves out manufactured-home parks. For SBA loans a business type narrows it by NAICS code.
• Size. 0.6x to 1.5x the amount you typed. A $2.4M loan reads as $1.4M to $3.6M. Just outside that, from 0.36x to 2.25x, counts as near.
• Purpose. Purchase, refinance, cash-out or improvement. Only HMDA records carry purpose, so in other lanes we say so.
• Place. The state, and a county or metro when given. Loans in a neighboring state count as near.
• Recent. The latest HMDA year (2025), or the last 24 months of the SBA file. A loan in the prior year or an earlier window counts as near.

The levels

Every lender gets one level, the highest its evidence supports.

Beside the level, five dots: Kind, Size, Purpose, Place, Recent, each with its word. Filled: a recorded loan shows it. Half: a balance sheet or office suggests it. Empty: nothing public tells us.

There is no score out of 100, and no percentage. A number like 87 reads as odds of a yes, and we do not know that.

The order

Grouped by kind of lender, because banks, credit unions and non-banks are not comparable. Inside a group, in this order: if the deal names a county or a metro area, loans like this one made there, capped at 10, so a lender with more of them comes first; then level; then loans like this one, capped at 10; then the share of the lender's own loans in this lane that are like this one, in 5-point steps; then the most recent evidence; then room above the loan amount on the size estimate; then name. Every step uses public filings and the deal as typed. Nothing a lender has paid or told us is an input.

1. Place, when you name a county or a metro area. Loans like this one made there, capped at 10. A lender with more of them comes first, so the place you typed changes the list in a way you can see.
2. Level.
3. Loans like this one, capped at 10, so national volume does not swamp a local lender.
4. The share of the lender's own loans in this lane that are like this one, in 5-point steps. A lender for whom this deal is normal sorts above one for whom it is rare.
5. The most recent evidence.
6. Room above the loan amount on the size estimate.
7. Name.

Groups are not comparable, so results never mix them: Banks and credit unions with an office in your state; then Banks and credit unions with no office in your state; then Non-banks.

Sentences on a result come from templates filled with data, never from a language model at request time, so nothing can be invented.

Where each lane's evidence comes from

Bank call reports carry balances, not originations, so a bank's "loans made" come only from HMDA and SBA files. Those files show nothing about office, retail, industrial or construction loans at banks. Bank 1-4 family loans cannot be split between owner-occupied and investor, and we label that plainly.

Rankings and the screener

A ranking is one measure, one period and one source. Each is defined here and on its page. We never rank banks by concentration, or by anything that reads as bad news about a bank.

• Multifamily dollars originated in the latest HMDA year. National and by state.
• Multifamily loans made in the latest HMDA year, manufactured-home parks excluded. National.
• 1-4 unit investor dollars originated in the latest HMDA year. National.
• 1-4 unit investor or business-purpose loans made in the latest HMDA year. National.
• SBA 7(a) dollars on record in the 36 months of the file, net of cancelled approvals. National and by state.
• SBA 7(a) loans on record in the 36 months of the file, net of cancelled approvals. National.
• SBA 504 first-lien dollars on record in the 36 months of the file, net of cancelled approvals. National.
• SBA 504 projects where the lender held the first lien, 36 months of the file, net of cancelled approvals. National.
• Credit union real estate commercial loans made in the first half of 2026, in dollars. National.
• Multifamily loans held. National and by state.
• Construction and land loans held. National.
• Owner-occupied commercial real estate loans held. National.
• Non-owner-occupied commercial real estate loans held. National and by state.
• Farmland loans held. National.
• Commercial and industrial loans held. National.
• All commercial loans held. By state only.

The screener filters and sorts on the fields in the data dictionary, and it shows its sort. Every sort is a column of public data, with the lender's permanent number as the tie-break. A search by name is ordered by how closely the name fits what you typed, then by assets. Payment is not a field.

The Receipts: how a recorded loan is printed

• Never a borrower name, street, ZIP or census tract, no loan status, no HMDA demographics, and never joined to deed records.
• Place is a county only if that county has 5 or more loans of that program in that year or window across the whole public file, otherwise a metro by the same test, otherwise the state.
• Amounts are rounded to 2 significant figures. SBA dates are month only. HMDA is year only.
• A NAICS code shows at 6 digits only if it has 5 or more loans nationally in the window, otherwise the sector.
• HMDA multifamily amounts are the midpoint of a $10,000 band, as HMDA publishes them.

Capacity figures: our arithmetic, not a rating

Our calculation from public filings, as of June 30, 2026. Not a rating, and not a statement about safety or deposits.

• Commercial real estate concentration, banks: construction, multifamily and non-owner-occupied commercial real estate, divided by tier 1 capital plus the allowance. It approximates the measure in the 2006 interagency guidance, which uses total risk-based capital and a 300% screening level that also requires 50% growth over 36 months. We do not use that growth test. The screening level is not a limit, and meeting it does not mean a bank has stopped lending.
• Credit union concentration: the same loans over net worth, so it cannot be compared to the bank figure.
• Size estimate, banks: 15% of tier 1 capital plus the allowance, from the national bank lending limit (12 CFR 32.3). State-chartered banks can have different limits. Labeled an estimate.
• Member business loan cap, credit unions: the lesser of 1.75 times net worth or 12.25% of assets (12 CFR 723.8). Some credit unions are exempt. The single-borrower estimate is the greater of 15% of net worth or $100,000 (12 CFR 723.4).

Where payment is, and where it is not

Payment appears nowhere in the order of results, and in no public figure. Verified is a paid listing. Order is not affected by payment. Three separate things hold that in place, so one mistake cannot undo it:

1. The ordering code is handed only public data. Lender accounts, plans and payments live in a different database. The code that sets the order is never given a way to read it.
2. A test reads the code on every build. It fails the build if the ordering code, or any file it uses, mentions an account, a plan, billing or a paid flag, or reaches a file that does.
3. A comparison runs the real data each quarter. Every ordering is run with no lender paid, then with the lower half of every list marked paid, then with a random half marked paid. The order must be identical all three times.

The latest comparison ran on November 9, 2026: passed, 80 orderings compared, 0 differences. The ordering code itself and the test are printed in full on the ordering code page, so anyone can read them.

The neutrality test, in plain words

Ask one question about any change to this site: if a lender started paying today, or stopped, would any list, search result or ranking come out in a different order, or would any figure change? The answer has to be no, and it has to be no by construction, not by promise.

• Labels go on last. "Verified (paid)" and "Ad" are read after the order is final and printed on the finished list. A label never moves a lender.
• Listing state is a filter, never a sort. You can choose to see only Unverified, Claimed or Verified (paid) listings. No list is sorted by it, starts filtered by it, or breaks a tie with it.
• Ads are separate. An ad never sits inside a result list. It says Ad before the name, in its own box, and names who paid. Our API and MCP answers carry no ads by default, and any paid item carries is_paid and its label.
• A paid page gets the same data, in the same order. A Verified (paid) listing adds the lender's own confirmed details, its own people and a contact form to its own page. It changes no figure from a public record.
• What a lender tells us can only ever appear as that lender's own statement, marked as such. It is never an input to an order.

More: How we make money, the Advertising Policy and the Verification Policy.

What is counted

This site covers every US bank and credit union that files a call report, plus the non-banks visible in public filings. That is 10,906 lenders: 3,840 banks, 473 savings institutions, 4,299 credit unions and 2,294 non-banks. Another 197 depositories with no current call report keep their pages for the record. A lender seen only in HMDA or SBA files counts as a non-bank unless it is a depository on record. This is the count this site uses.

The recorded loans are the 258,919 you can open on a lender page: HMDA multifamily loans for 2024 and 2025, and SBA 7(a) and 504 loans made Oct 2023 through Sep 2026. Cancelled and not-funded SBA approvals are never counted or shown.

What each count includes

A count on this site means one thing wherever it appears, and has a name. These are the names and what each includes, so two pages that print different numbers are counting different things, and say so. The numbers below are read from the same data the pages use.

Lenders on record: 10,906

Every bank, savings institution and credit union that filed the call report for June 30, 2026, plus the non-banks seen in HMDA or SBA files. This is the lender count on the home page and in page headlines.

Lender pages: 11,103

One page for every lender record: the 10,906 on record, plus 197 depositories with no 2026 call report, which keep a page for the record. The A to Z index counts these.

Lenders listed in Loan Comps and the screener: 10,854

The lenders on record that are still active. 52 institutions filed the June report and then merged, closed or stopped filing. They stay on record, and Loan Comps lists them apart, under "Not an option, and why".

Recorded loans in a place

Every recorded loan located there, whichever lender made it and whatever that lender's status is now. Ohio had 1,361 multifamily loans recorded for 2025. A page that counts only the loans of lenders still active says so beside its figure.

Lenders that made a loan like this

The lenders Loan Comps lists with at least one recorded loan like the deal: 53 for the Ohio example on the home page. The funnel above that figure also counts the lenders that are no longer an option, 55 in all, and shows the difference on its own line.

The checks we run

Every build reconciles the file against its sources. Category totals are checked against each filer's reported totals. SBA figures cover Oct 2023 through Sep 2026, and net figures leave out approvals that were later cancelled.

Limits on reading and on the API

Reading the site is not metered like the API. On the Loan Comps, search and compare pages, a visitor may view 2,000 pages a day, or 5,000 once signed in free. A hosted assistant such as Claude or ChatGPT shares 20,000 a day, on the addresses it publishes. The name suggestions and result counts those pages fetch while you type have allowances of their own, 5,000 a day each, and are not counted as page views. A person does not meet these numbers in ordinary use, and signing in is not needed to read.

The API and the MCP server have daily limits that depend on how you ask. With no key, 50 calls a day and 10 rows a call. With a free account, 500 calls a day. With a developer license, 5,000 calls a day, or 50,000 for a team. Hosted assistants such as Claude and ChatGPT share a pool of 5,000 calls a day, because their users arrive from a few addresses. The limits gate depth and bulk, not single lookups, and the full table is on the developers page.

When you reach a limit, the page says so first, in plain words, offers the free sign-in where it raises the number, and tells you when the allowance resets: midnight UTC, which is 7 p.m. Eastern in winter and 8 p.m. in summer.

What this data cannot tell you

Figures are compiled from public filings and government data, as explained on our methodology page. They are shown as filed, may be out of date or wrong, and do not cover every loan a lender makes. Search results are not a recommendation. Nothing here is a credit rating, a prediction, or a statement about the safety of any institution or its deposits. Unless a page says otherwise, institutions shown have not reviewed it. Names identify institutions and imply no affiliation.

• Whether anyone is taking new loans today. Filings look backward.
• What a lender will quote, or its underwriting limits, unless it has told us.
• Loans a lender made that no public file records.

Source and date on every figure

A figure on this site is printed with the filing it came from and the date of that filing, beside the figure or at the head of its table. A sentence that states a figure names its source in the sentence. When the public record has no figure, the page says "Not reported". The build checks lender pages for this, and the sources page lists every file and the day it was pulled.

Corrections and changes

Anyone can ask for a correction, free, on the corrections page. We reply within 2 business days and fix confirmed errors within 10. If we misread a filing we fix it and log it in the corrections log. Where the filing itself is wrong, the figure stays as filed with the lender's correction beside it. While a credible dispute is checked, that one figure is taken out of the site and the lender's page says it is under review. Every change to data or to this method is on the changelog.