Method

What a row is

For each question, on each run day, we send the same words to each source and log one row per source: the status, the names found in the answer, and the SHA-256 of the raw answer file. The sources are OpenAI, Anthropic, Perplexity and Gemini (by API) and Google search.

What the status words mean

Status words: OK means an answer was logged. QUOTA_BLOCKED means the engine refused the call over its quota. FAILED means the call did not complete. EXTRACTION_FAILED means the response came back without readable results (in the logged rows this is Google search pages blocked by a consent wall or captcha). Every status is printed at the same size as the wins.

Names and merging

Names are extracted from each answer. Spelling variants that the Edition 1 files group together are shown as one name. Dashes and curly quotes are normalized to plain characters in names. Extraction is automatic, so some entries are headings or terms and not businesses. We log that on the corrections page.

Two counting rules, and which page uses which

Extracted-list rule (the market pages on namebeam.ai, the Edition 1 answers table, the per-market hashed files): a business counts for an engine on an answer day when it appears in the business list the collector extracted from that answer (bold names and listed names, each verbatim in the answer). Text-mention rule (the homepage explorer, data/explorer_data.json, the named lookup): a business counts for an engine on every answer day whose saved answer text names it, whole words, full name, case and punctuation ignored. The two rules give different numbers for the same business and window: for Chicago personal injury attorneys, OpenAI, Salvi, Schostok and Pritchard counts on 18 of 39 answer days under the text-mention rule (Sep 1 to Oct 11, 2026) and on 4 of 71 answers under the extracted-list rule (Aug 9 to Oct 19, 2026). Every page states which rule it uses beside its numbers; never compare a count from one rule with a count from the other.

DQ-5, the extracted-list undercount. The name lists under-count the businesses an answer mentions, on every engine and by different amounts. For Anthropic the collector reads bold spans as names, so a plain-text mention is missed; for the other engines a name outside the list the engine wrote is missed. A recount of the 71 OpenAI answers for Chicago personal injury (Aug 9 to Oct 19, 2026) found Salvi in the text of 36 of 71 while the extracted lists named it on 4. This is documented in the Edition 1 pack README as DQ-5 and applies to every extracted-list count on these pages. A recount of the market pages under the text-mention rule is prepared and will land as a dated correction if adopted.

Dated method changes

2026-09-27: Gemini API returned HTTP 402 on every call from this date (prepayment exhausted); Gemini rows are logged as FAILED, never as zero, until the route is restored. 2026-09-28: Perplexity moved from Sonar chat completions to the Agent API with web search forced on, after Perplexity retired Sonar chat completions on 2026-09-27; windows before and after are reported apart (WA and WB in Edition 1) and are never quoted as one trend. 2026-10-14: the name extractor became a union of the earlier readers, forward only; rows before that date are not re-extracted. 2026-10-15 and 2026-10-16: a free-tier Gemini route returned answers on a few questions; 2026-10-17 on, FAILED again. 2026-12-11: OpenAI retires the gpt-5-mini-2025-08-07 snapshot; a dated change row will be added when the collector moves.

Units, so numbers are not mixed

47 questions: the September 2026 roster, asked on each of 30 days. 73 questions: the Edition 1 roster, Sep 1 to Oct 13, 2026 (55 market or category questions and 18 named-business probes), 42 run days, 12,180 answer rows. 46 market cells: the slice of the explorer file behind the Digital Information World study, a market plus a niche, Sep 1 to Oct 11, 2026, 24 markets and 33 niches. The live explorer file (data/explorer_data.json, read 2026-10-24) holds 52 cells, 24 markets and 34 niches over Aug 5 to Oct 21, 2026, and grows with each run day. 27 market pages: one question each, from Aug 5, Aug 9 or Aug 26, 2026 by market; each page states its window. Edition 1 ends 2026-10-13; Gemini answered only on 2026-10-15 and 2026-10-16 after it. A run day is a date with at least one row; an answer day for an engine is a date with an OK row from that engine.

Who entered and who dropped

Each run day is compared with the previous run day that has rows. The engines that answered OK on both days are the ones compared, so an engine skipping a day cannot look like a business dropping out.

What the counts are not

A count is not a ranking, not a score and not a promise about traffic or bookings. AI answers vary from run to run.

Check it yourself

Each CSV lists the raw file name and its SHA-256. The raw files are in the public corpus. manifest.sha256 lists the hash of every CSV on these pages.

Compiled by Terry J Carter, Carter Enterprise LLC. Page built 2026-10-21. Namebeam, a Carter Enterprise property. This is the first leg.