Methodology

11.1 How the clause dataset was built

1. For each company, we located its current consumer agreement containing the dispute-resolution section — deposit agreement, cardmember agreement, terms of service, user agreement, customer agreement — preferring the company’s own page over any copy.
2. We fetched it with a single-request, no-browser, no-login tool that records the requested URL, the final URL after redirects, HTTP status, content type, byte count, SHA-256 of the raw response, the capture timestamp in UTC, and the exit IP and country of the machine that fetched it. Every capture in this report was made from a United States exit, verified at the start and end of each run. A date given for a capture is the UTC date of its recorded timestamp, so captures made on the evening of September 18, 2026, Central time, carry the date September 19, 2026.
3. We read the captured text and wrote a structured extraction: 22 fields per document, each field carrying either a verbatim quotation from the capture or an explicit null, because silence is a finding.
4. A verifier string-matched every quotation back to the capture. A field that did not match was rejected and rewritten against the document. Only passing extractions are in the dataset.
5. The quotation behind every flag in the fourteen fields most likely to generate a headline was then re-read against the document and assigned a legally meaningful category under written rules, every field recorded as silent in those fourteen was re-read the same way, and the counts were restated. Where the machine flag and the quoted language disagreed, the language won. The counts in § 4 are the reviewed counts, not the machine counts.
6. The mass-arbitration thresholds in § 4.10 were established a different way, and the difference matters. Each of the 110 clause-bearing documents was handled individually rather than by pattern search, and where a document carries mass-filing terms that section was read. Each number, unit, time window and same-counsel condition was recorded with the sentence it came from, each quotation was machine-verified against its capture, and a reviewed category was then assigned under written definitions, because a number that batches claims and a number that bars a filing are different facts that a keyword scan reports identically. The 46 documents recorded as silent were spot-checked by regex across the whole capture.

Why that section was rebuilt rather than extended. Earlier drafts of this report built the threshold table by pattern search, and the pattern was wrong. It looked for “N or more demands”. Contracts say “claims”, they say “cases”, they say “in the event N or more”, and one says “at least 24 other customers”. Companies that plainly state a trigger were recorded as having none, and at one point a sentence in the draft turned each of those silences into an affirmative statement that the company’s contract set no threshold. A search that finds nothing is not a finding — it is a failed search until the section has been read. That is why the figures in § 4.10 come from a document-by-document read, and it is the reason this report labels keyword-indicated evidence as such wherever it appears.

11.2 What the dataset is not

• It is not a sample. Companies were selected by U.S. consumer reach within each sector, largest first, plus companies headquartered in or heavily serving Minnesota within each sector. Do not read a percentage here as a percentage of American consumer contracts.
• Documents are not companies, and unless a sentence says otherwise the number is documents. Four corporate families hold more than one document in this dataset — PayPal Holdings (PayPal, Venmo), Synchrony Financial (Synchrony Bank, CareCredit), JPMorgan Chase (Chase, and Zelle, which incorporates Chase’s clause) and Big Picture Loans (two documents). By family, the 139 documents are 135 companies, the 110 clause-bearing documents are 107, and the 29 without a clause are 28. Where this report gives a company figure — in § 3.1, § 4.10 and § 7 — it says so and counts corporate families. That is the only definition of “companies” used anywhere in this report.
• The same Amazon document was captured twice, by two different research passes, and counted twice until late in this project. Two sector agents fetched the same Conditions of Use page; the texts are byte-identical. It was found when a discrepancy between a document count and a company count turned out not to reconcile. The duplicate is recorded in data/duplicates.json, the second copy is excluded from every count in this report, and a pairwise full-text comparison of all 9,591 document pairs found no other duplicate — a comparison controlled by scoring it against the known Amazon pair, which it flagged at 1.000. Every count in this report is over distinct documents.
• A website’s terms of use are often not the contract that binds the customer. Gym memberships, retail installment contracts, loan agreements signed at closing and residency agreements signed at move-in are generally not published. Where we captured website terms, we say so and name the document. Senior-living residency agreements are private paper contracts; this project captured none, and this report describes none.
• Some documents did not come from the company’s own site. American Express, Bread Financial/Comenity, CareCredit and Synchrony were fetched from the CFPB’s credit card agreement database; Synchrony’s copy, the American Eagle Outfitters co-brand card agreement, carries the printed form date “9/2023”. Sallie Mae’s sample promissory note was fetched from the Maryland Department of Labor’s Office of Financial Regulation site, and Sunrun’s agreement from the New York Department of Public Service’s document site. The archival snapshots are listed in the next bullet.
• Thirteen documents are archival snapshots, not live captures — PNC Bank, TransUnion, College Ave, Mayo Clinic, Chime, Coinbase, Roblox, H&R Block, ADT, Expedia, Tesla, Anytime Fitness and Cox. An archival capture is weaker evidence of current terms and is labelled wherever it is used. Tesla’s is a September 2022 order form recovered from a 2026 snapshot. TransUnion’s page is dated 2019.
• One document is on a different evidentiary footing from all the others. Ticketmaster/Live Nation’s clause text comes from a browser-rendered partial capture — 19,264 characters of a 45,820-character rendered article, with the SHA-256 of the full rendered text recorded but no raw HTML saved — because the clause is injected client-side and four ordinary captures returned 897-to-1,417-character page shells with zero occurrences of “arbitrat”. It is labelled every time it is used. The page was rendered on September 19, 2026, in a browser that was not logged in, from a United States exit, without accepting any terms or submitting any form.
• One document in this dataset has a successor version with a different answer. On the September 19, 2026 capture date, X Corp.’s terms page carried both the U.S. terms then in force, dated “Effective: April 10, 2026” on the page, and a version stating that it “will go into effect on October 9, 2026” that adds a conditional arbitration provision. This report counts X Corp. among the documents with no arbitration clause because it counts the terms in force on the capture date. The October 9, 2026 version is quoted in § 6.2 and is in no count.
• DoorDash, MoneyLion, OneMain, Ticketmaster and TitleMax were not read in full, and § 4 counts them that way. Where the part of a document that would answer a question was not captured, § 4 shows it as not captured and never as silent. TitleMax’s page is its “Texas Additional Application Terms”; the arbitration agreement it describes was not captured. MoneyLion’s Terms of Service incorporate a separate “Agreement for Resolving Disputes”, which was not captured; the captured Terms of Service state the delegation and modification terms themselves, and those two are counted. In OneMain’s sample loan agreement, the opening sections of the arbitration agreement did not survive text extraction; the captured later sections support the rest. Ticketmaster’s saved text omits sections that do not address disputes (above), so where it says nothing on a question, § 4 records that question as not captured. DoorDash’s capture is the help-center page that carries Sections 12 through 14 of its Consumer Terms (payment terms, DashPass subscriptions and the arbitration agreement); any provision elsewhere in those terms on how they may be changed was not captured, so § 4.11 shows DoorDash as not captured on that question.
• Upstart’s document is a website’s terms, not a loan agreement. The document captured is a co-brand Website Terms of Use at mbc.upstart.com stating “Effective Date: March 19, 2019.” It was read in full, and it is counted as what it is.
• Blocked is not “no clause.” Spectrum/Charter, Rent-A-Center, Jefferson Capital’s consumer portal, Portfolio Recovery Associates and Lexington Law could not be captured and are absent from every count. Four debt-collection and credit-services companies missing from a consumer-arbitration dataset is itself a limitation worth stating.
• The SoFi, Bank of America, Tesla, American Express and U.S. Bank texts were re-extracted from the saved PDFs. In the first text extraction of the SoFi, Bank of America, Tesla and American Express PDFs, words were fused or split, or characters the PDFs’ fonts encode as ligatures were lost, so language that is in the document did not match its text. On September 25, 2026 each of the five, U.S. Bank’s for consistency, was re-extracted from the untouched raw file, whose SHA-256 matches the hash recorded at capture, and each re-extraction is saved beside the original with the extractor and every character mapping it applied. The re-extracted text is the text of record for all five, and every quotation this report takes from them verifies against it.

11.3 The AAA analysis

Source: AAA’s Q2 2026 consumer disclosure workbook, downloaded directly from adr.org, 24,522,699 bytes, SHA-256 recorded. Our non-blank row count for the full file reconciles exactly to AAA’s own stated total of 211,748 rows. Filter: AAA’s own literal field value Typedispute == 'Consumer' (171,546 rows). The filter was tested against rows it must reject, using AAA’s employment-only Salary Range field: populated in 0 of 5,000 sampled retained rows, and in 1,932 of 2,000 and 1,998 of 2,000 sampled rows of the two employment categories. The script was run twice and the output diffed: no differences.

Every count is a claim-row unless stated otherwise. AAA publishes one row per named non-consumer party per case and says so in its legend. 171,546 rows correspond to 168,433 distinct Case IDs.

Two data-quality notes about AAA’s own file. Its Source Of Authority column carries the value AAA Clause on every one of the 211,748 rows, with no other value anywhere, so it does not in fact distinguish clause-based from non-clause-based cases as the statute contemplates. And at least seven awarded rows carry consumer-side claim amounts above $1 billion on their face, the largest $469,822,000,000. We did not edit AAA’s data; we report medians, which are far less sensitive to that kind of error, and we publish no claim-amount total.

We also downloaded and analyzed JAMS’s consumer file. Those figures are not in this report.

11.4 The CFPB card scan

The complete 2026 Q2 bulk archive (1,492,022,880 bytes; SHA-256 ebe51a41e81da6a974146c293e132c7d006a723a4533152371cb3d28a83f4f4f), extracted to 3,264 agreement PDFs across 561 issuer folders, with a manifest carrying per-file path, issuer, byte count and SHA-256. Text extraction by pdftotext -layout with two independent fallback libraries. 184 files (5.6%) returned zero characters under all three methods; a random sample of them, checked page by page, held only scanned images. They are excluded from every share and are never counted as “no clause.”

The scan was spot-checked twice against the agreements’ text, on two different random seeds, 100 checks in total. Round one found one false negative — a Spanish-language agreement whose only English “arbitrat” strings were the proper nouns “American Arbitration Association” and “Federal Arbitration Act”. We added Spanish patterns and reran the full corpus; round two returned 25 of 25 correct on each side. Both full runs produced byte-identical output.

We separately confirmed that the CFPB’s live per-issuer listing and its bulk quarterly archive are the same underlying data — 561 issuers each, zero in one and not the other, confirmed programmatically and spot-checked down to a matching file identifier.

11.5 The clause-history analysis

Dated archival snapshots of 46 companies’ own documents at 2016-07-01, 2019-07-01, 2021-07-01, 2023-07-01 and 2025-01-01, plus the live capture, yielding 169 usable snapshots of 276 attempted. A snapshot is unusable — not “no clause,” unusable — if its actual timestamp is more than about twelve months from the requested date, if the extracted text is under 1,500 characters, or if it is a failure or redirect page. A snapshot that is not the English-language document is unusable too, which excludes Netflix’s 2016 snapshot (the archive served its Dutch page), Dropbox’s 2016 and 2019 snapshots (its Chinese pages) and YouTube’s 2023 snapshot, whose saved text is not readable. Every indicator in § 6 is keyword-indicated, every reported first appearance was checked against its source snippet, and two demonstrable false positives were caught and excluded rather than published.

That analysis covered four sectors. Healthcare, education, home services and most newer marketplace and gig companies are not in it, because their captures arrived after the target list was frozen. And its null result on removals was wrong: it found no company that dropped a clause, because Amazon’s page had moved to a new address and Valve was not on its 46-company list. Both removals are documented in § 6.3 from other evidence. We report that limitation because we found it ourselves, and because a null result from an incomplete grid is not a finding.

11.6 What we did not do, and what we are not saying

• We did not contact any company named in this report before publication.
• We make no statement about why any company adopted, removed, changed or restored a clause; who drafted any shared template; or that any company acted in response to litigation, a statute or a mass filing. The dataset contains dates and text. It contains no motive.
• We make no statement that any card issuer failed to submit, stopped submitting, or is exempt from submitting agreements to the CFPB. Absence from one bulk collection establishes none of those things.
• We do not characterize any named company’s conduct as unlawful, deceptive or wrongful. Where a court used a word, we quote the court and name the case.
• Every matter described in § 5 states its status — granted, denied, withdrawn, reserved, vacated, or unknown. “Unknown” appears where it is the truth.

11.7 Open questions we intend to close

1. Whether the CFPB’s earlier quarterly archives contain agreements from the large issuers absent from the 2026 Q2 collection. If they do, the story is a change over time. If they do not, the story is about the structure of the database and about no company at all. This is the highest-value unanswered question in the project.
2. The date on which Ticketmaster’s terms replaced New Era ADR with JAMS.
3. Whether Bank of America, JPMorgan Chase and U.S. Bank removed the consumer’s choice of administrator between their 2023 and 2025 terms.
4. A Minnesota-specific analysis of AAA’s full 211,748-row disclosure file: Minnesota respondents, Minnesota consumer counsel, repeat-player counts, fee allocations.
5. The exact dates of Netflix’s move from AAA to JAMS, and of the arbitration section appearing in Google’s YouTube paid terms for NFL Sunday Ticket.

11.8 How to get the data

The dataset behind this report is available on request, without charge, from Madgett Law, LLC: the clause matrix, one row per document and one column per field, with the verbatim quotation and the capture’s SHA-256 for every entry; the reviewed-category files behind § 4 and § 4.10; the scripts that recomputed the AAA disclosure file and scanned the CFPB credit-card collection; and the capture manifests, with every URL and hash. Write to Madgett Law, LLC, IDS Center, 80 South 8th Street, Suite 1650, Minneapolis, MN 55402, or call 612-470-6529.

---