# E-signature teardown: coding notes, codebook v2.0 (collected September 26, 2026)

The published brand codes are the reader codes in coded_v2/ (codebook v2.0, protocol gold2/READER_PROTOCOL_v2.md). The rules-based coding in coded/ (codebook v1.5) is retired for this teardown and is not published.

## Coding method and agreement

- Every one of the 100 answers was coded by two independent AI readers (Claude agents, labels A and B), each working blind from the written protocol, with link addresses removed from the answer text. Each brand in each answer got one of six codes: picked, recommended, listed, passing, anchor, not_a_mention.
- Where the two readers disagreed, a third AI reader (J03) read the answer and decided. The readers are AI models, not people.
- Reader vs reader agreement in this teardown: 96.1% on the six codes, 98.8% at the broad level (recommended or not). Codes decided by the third reader: 16, in 7 answers. (gold2/CODED_V2_SUMMARY.md, row 18)
- The rules-based coder (v1.5, the one that produced coded/) matched the final reader codes on 71.5% of brand codes at the broad level. Misses already documented under v1.5 and resolved by the readers include head-to-head forks the cues could not see ("Choose Adobe if", "better if", a "Bottom line" lead-in) and citation link text counted as answer text; the rules also coded many fit-labeled options as listed (110 listed mentions under v1.5 against 5 now), and counted 15 answers with no recommendation against 6 now.

## Collection

All 100 ChatGPT tasks and 100 Google tasks were posted and fetched through the DataForSEO standard queue on the evening of September 26, 2026 (US Mountain time; the raw files carry UTC timestamps, early September 27). No task failed and none was re-run.

Six answers came back as a single sentence that announces or offers a comparison and then stops, with no sources: q014 (best self-hosted e-signature software), q034 (docusign vs dropbox sign), q048 (pandadoc vs adobe sign vs dropbox sign), q054 (pandadoc or dropbox sign for a startup), q060 (adobe acrobat sign or docusign for a company on microsoft 365), q068 (cheaper alternative to adobe acrobat sign). They are kept as returned and counted as answered. They are the 6 answers with no recommended brand, 6 of the 11 answers with no pick, all 4 head-to-heads with no pick, and all 6 answers with no citation.

## Judgment calls by the third reader (gold2/out/J03.jsonl)

- q034 and q054 (stubs): the brands in the one sentence are passing, "one-sentence comparison announcement repeating question". In each, one reader had them listed. This follows the protocol rule for announcement-only answers.
- q079 (docusign alternative for healthcare): one reader coded six brands picked as "use-case picks"; the third reader coded all six recommended: "best-fit table and use-case map, no verdict". This answer is one of the 5 full answers with no pick, and the only DocuSign-alternative answer without a pick.
- q007 (cheapest e-signature software): DocuSign picked ("My practical take" assigns it the brand-recognition case); Dropbox Sign and SignNow recommended, "domain table row with best-for label only".
- q012 (best e-signature api for developers): Adobe Acrobat Sign recommended, not picked: it has a best-fit table row but is not in the answer's "If I were building" choice.
- q065 (signnow alternatives): DocuSign listed, "feature description, no buyer case assigned". This answer also has no pick.
- q088 (law firms): Clio picked, "in answer's narrowed set; prioritize if on Clio".

## Judgment calls that shape the findings (both readers agreed)

- Conditional verdicts count as picks. "Choose Adobe if PDF-heavy" (q032, q033), "Choose Adobe Acrobat Pro if PDF-heavy" (q029), "my take: Acrobat Pro if PDFs heavy" (q085) are picks under the protocol. Of Adobe Acrobat Sign's 48 picks, 38 have a reader reason that names PDF work or an existing Adobe or Microsoft setup; 12 of its 13 recommended-not-picked codes are fit labels of the same kind ("best if you already use Adobe", "PDF-heavy workflows"). These counts are from the readers' free-text reasons (10 words max), matched with a regex; they are a description of the reasons, not a separate code.
- Head-to-head forks: both brands picked in q042 (DocuSign vs BoldSign: a "Bottom line" that gives BoldSign the embedded-signing case and DocuSign the enterprise-ecosystem case) and q055 (DocuSign or dotloop, "better if"). The only one-sided head-to-head is q053: "For real estate, I'd lean DocuSign", Adobe recommended ("strong if already on Adobe workflows").
- q017 (best free way to sign a pdf online): the answer's pick is PDF24, which is not in the dictionary; Adobe (via "Adobe Acrobat", Fill & Sign) and DocHub recommended, Smallpdf listed. Counted as an answer with no pick.
- q009 (most popular), q019 (contract management): every brand carries a fit ("suited to", "useful if", "geared toward") and no verdict, so all recommended, no pick.
- q091 (freelancer, a few contracts a month): Dropbox Sign picked, PandaDoc recommended ("worth considering if contracts start as proposals"), DocuSign listed. It is the only advice answer without a DocuSign pick.
- Citation labels at the end of a sentence (for example "[PandaDoc]" after another brand's line) are passing under the protocol, so a vendor's blog citation no longer creates a recommendation for that vendor.

## Brands the readers added

The two readers added 12 codes outside alias matching (6 each), every mention flagged by both readers (they differed on the code only in q007): 6 brand mentions in 4 answers, each named by a vendor domain used as link text in subject position.
- q007: Dropbox Sign (sign.dropbox.com table row), recommended after adjudication.
- q022: Dropbox Sign (sign.dropbox.com, "best all-around" for a small firm) and Adobe Acrobat Sign (adobe.com, "best if PDF-heavy"), both picked from the answer's shortlist.
- q063: Dropbox Sign (sign.dropbox.com) and Zoho Sign (zoho.com), both picked from "quick picks".
- q097: Dropbox Sign, recommended (table row: simple, occasional signature collection).

## Dictionary notes that still apply

- Two domains were added after collection, before the final coding run: dropboxsign.com for Dropbox Sign and authentisign.com for Authentisign. Both are the vendors' own sites that had been classed as other third-party. No mention changed. Details and effect in DICTIONARY_CHANGES.md; the drafted dictionary is brands_original.csv.
- "adobe acrobat" is an alias of Adobe Acrobat Sign (decided before collection, INPUT_NOTES.md). In 8 answers it is the only thing that credits Acrobat Sign, and the product named there is the PDF editor (Acrobat Pro, Acrobat Reader, or "Adobe Acrobat" with Fill & Sign): q003, q017, q023, q029, q070, q075, q085, q095. The readers coded the alias as Adobe Acrobat Sign in all 8 (picked in q003, q023, q029, q070, q075, q085; recommended in q017, q095); none used not_a_mention. Without those 8 answers Adobe Acrobat Sign is 53 named / 53 recommended / 42 picked instead of 61 / 61 / 48, behind Dropbox Sign (55 / 54 / 44); 13 instead of 14 advice picks and 16 instead of 18 industry-question picks. A ninth answer, q097, writes "Adobe Acrobat Pro/Sign" and is a genuine Acrobat Sign mention. Disclosed on the page (method note and limitations).
- Bare "Adobe" is not an alias. Under v2 it no longer costs Adobe picks: in q032 and q033 ("Choose Adobe if ...") the readers coded Adobe Acrobat Sign picked because the question names Adobe Sign.
- No brand outside the dictionary is named in 5 or more answers. LibreSign appears in the out-of-dictionary list for q006, q070 and q072. The accountants answer (q026) leads its shortlist with SafeSend and also lists TaxDome and Canopy, none of which are in the dictionary.
- Case-sensitive aliases (capitalized only): Signable, Ironclad, Clio. Clio is a law practice suite, in the dictionary for law-firm questions.
- dupple.com, esigncompare.com and similar comparison sites are classed other third-party because they are not on the codebook's review-and-media list, the same list every teardown uses.
- Citations are unchanged by the re-coding (same domains, classes and Google overlap as coded/ after the dictionary fix).