# Business intelligence teardown: coding notes, codebook v2.0 (collected September 26, 2026)

Published data: coded_v2/ (mentions.csv with a 0/1 "picked" column, observations.csv, citations.csv, stats.json at the broad level, picked_stats.json at the picked level). The rules-based files in coded/ are superseded and are not the source of any published number.

## Collection

All 100 ChatGPT tasks and 100 Google tasks were posted and fetched through the DataForSEO standard queue between about 8:15 and 8:30 PM US Mountain time on September 26, 2026; the raw files carry UTC timestamps, early September 27. Zero failed tasks, no retries. Raw responses are saved as returned.

## How the answers were coded

- Every one of the 100 answers was coded by two independent AI readers (Claude), each working blind from the written protocol (gold2/READER_PROTOCOL_v2.md) without seeing any machine coding. Each dictionary brand in an answer gets one of six codes: picked, recommended, listed, passing, anchor, not_a_mention.
- Published levels: picked = the answer's own verdict; recommended (broad) = picked plus brands assigned to a stated case or fit; named = recommended plus listed.
- Reader vs reader agreement in this teardown: 98.5% across the six codes, 100.0% at the broad level.
- A third AI reader decided the 7 codes the two readers disagreed on. All 7 were the line between picked and recommended; the broad level was never in dispute.
- A rules-based first pass (codebook v1.6 rules) was checked against the final reader codes: 95.6% agreement at the broad level. Reader codes are what is published.

## Third-reader decisions (gold2/out/J02.jsonl)

- q065 "domo alternatives", Metabase: picked. The line is a decision rule for a replacement scenario (inexpensive and simple); reader A had it as a fit label.
- q087 "which BI tool is best value for money", Looker, Metabase and Tableau: picked. All three sit in the answer's own "My practical take" section, one per case (large data organization, developer or startup, highly visual analytics).
- q090 "what business intelligence software should I pick for SaaS startups", GoodData and Sisense: recommended. Both are named as examples of the embedded analytics category, not chosen by the answer; reader B had read the table column as a pick.
- q095 "which BI tool should I use if my data is in postgres", Apache Superset: recommended. It is on the answer's shortlist with a fit (open source teams) but left out of its "If I were choosing" verdict. Metabase, Power BI, Tableau and Looker are picked there ("Pick Metabase if...", and the same line for each).

## Judgment calls that shape this category

- Stack lines. "If you already use Microsoft 365 / Excel, Power BI" inside a verdict section is picked; the same fit sentence in a catalog of options is recommended. Power BI's 4 recommendations without a pick (q027, q075, q089, q100) are all that Microsoft-fit line in answers whose verdict went to a Google or Snowflake tool.
- Fit tables. Many answers carry an options table with a "best for" column. Rows in those tables are recommended, not picked, unless the answer's own verdict repeats them. Domo's 5 recommendations in questions that do not name it (q002, q004, q010, q017, q023) are all such rows; q017 labels the Domo row by its domain.
- Products being replaced. In q006 "best open source BI tool", Tableau and Power BI appear only as what an open source tool would replace and are passing.
- Head-to-heads. 29 of 30 pick every brand named ("choose A if..., choose B if..." makes both picked). The exception, q040 "thoughtspot vs tableau", is a one-sentence offer to compare; both brands are passing.
- Clarifying questions. q076 "sisense alternatives" only asks what kind of alternative the buyer wants; Sisense is the anchor and nothing else is named.
- Citation-only names. A brand that appears only as a citation label or link address is passing (Holistics in q027 and q065, Omni in q079, Domo in q035, Databox in q098).

## Reader-added mentions

Readers coded 3 brand mentions that the candidate list (dictionary alias matching) missed, all from reader A: q095 Apache Superset (recommended; named as "Superset" in a shortlist row), q098 ThoughtSpot (picked; "Consider ThoughtSpot if..." in the answer's closing advice) and q098 Databox (passing; a citation label only).

## Dictionary

- Looker and Looker Studio (formerly Google Data Studio) are one brand, "Looker". They could not be kept apart: "looker" matches inside every "Looker Studio" mention. Looker's counts are therefore the Google Looker family. Counted from the answer text (link labels and link addresses ignored):

| Level | Looker Studio only | Looker only | both | total |
|---|---|---|---|---|
| named | 25 | 32 | 14 | 71 |
| recommended | 25 | 30 | 14 | 69 |
| picked | 22 | 26 | 13 | 61 |

  Split apart, neither product would be second: enterprise Looker is named in at most 46 answers and picked in at most 39, Looker Studio at most 39 and 35, both below Tableau (56 named, 47 picked).
- Four answers (q039, q068, q071, q089) say Google renamed Looker Studio to Data Studio in April 2026, and one citation points at docs.cloud.google.com/data-studio. All four answers also use the name Looker Studio, so none was missed. Bare "Data Studio" is not an alias (it would also catch Metabase's Data Studio feature in q037, and Azure Data Studio).
- Two domain fixes after collection, before the final coding run: GoodData also matches gooddata.ai (its current site), Power BI also matches powerbi.com. Details and before/after numbers in DICTIONARY_CHANGES.md; brands_original.csv is the dictionary as first fixed.
- Excel and Google Sheets are not in the dictionary (not BI software; decided before collection). Excel is named in 65 answers, in 48 of them in the same sentence or line as Power BI.
- google.com is Looker's own site, microsoft.com Power BI's, apache.org Apache Superset's (INPUT_NOTES.md).
- Case-sensitive aliases (capitalized only): Superset, Preset, Sigma, Hex, Omni, Amplitude. Every capitalized match was checked in context: all refer to the products (no "Six Sigma", no hex colors).