Public, hashed raw answers, published as dated factual records; our compilation is CC BY 4.0. Five findings from a record of 12,180 dated rows, 2026-09-01 to 2026-10-13. Each one prints its number, its denominator, its window and the file it comes from.
The rows are 8,160 AI engine answers, 2,010 Google result pages and 2,010 engine agreement rows. Each day a program asks four AI engines and Google the same fixed business questions, in 25 US local markets and 2 nationwide question sets, across 38 market and niche cells that the metrics read, and saves every answer word for word. These are API calls, not the apps people open on a phone. Perplexity alone searches the web in this setup, so the engines are not like for like.
Perplexity, 2026-09-01 to 2026-09-27. 2,398 of 5,486 business-day namings were returned again the next day.
Two consecutive days with names on both: the median overlap between the two lists was 30 percent, over 832 day pairs. Same window and method for the others: OpenAI 31.2 percent (1,130 of 3,621) with a median overlap of 17 percent (699 pairs); Anthropic 50.0 percent (2,015 of 4,030) and 33 percent (988 pairs); Gemini 19.6 percent (402 of 2,052) and 10 percent (174 pairs, a small count). The 43.7 percent counts names returned again the next day. The 30 percent is the overlap between two days' lists. They are different statistics.
Source: metrics/metrics_v2.csv, methods P.raw.pername and J.raw.pair_bothnamed.median
Perplexity businesses named on 20 to 80 percent of answered days (method S.gated.coinflip, gated names): 173 of 564 in 2026-09-01 to 2026-09-27 (WA); 98 of 351 in 2026-09-28 to 2026-10-13 (WB, 15 run days).
Persistence on the raw list (method P.raw.pername) fell from 43.7 percent (2,398 of 5,486 business-day namings returned again the next day) to 27.1 percent (382 of 1,411). Median list overlap fell from 30 percent (832 pairs) to 17 percent (492 pairs). On 2026-09-27 Perplexity retired Sonar chat completions and the collector moved to the Agent API with web search forced on. The two windows are not compared as one series: WB is 15 run days, and the method changed between them.
Source: metrics/metrics_v2.csv, methods S.gated.coinflip, P.raw.pername, J.raw.pair_bothnamed.median and Y.answered.day, relabeled as the share of asked days with a parsed answer and at least one extracted string (rows WA and WB; WB run days = 570 asked cell days over 38 cells)
Mean overlap of two engines' raw name lists on the same day, OpenAI with Perplexity, 779 days, 2026-09-01 to 2026-09-27.
Perplexity with Gemini 2.7 percent (280 days). OpenAI with Anthropic 0.9 percent (781 days). Perplexity with Anthropic 0.1 percent (903 days). On names that pass the name gate, the highest pair is Perplexity with Gemini at 6.1 percent (280 days). The engines were called with different settings and Perplexity alone searches the web, so this describes answers from API calls. It does not say the engines disagree about the market.
Source: metrics/metrics_v2.csv, methods A.raw.pair and A.gated.pair
A gated business name appeared on 34 of 1,026 answered cell days, 2026-09-01 to 2026-09-27. To 2026-10-13: 54 of 1,596 (3.4 percent).
In the same window, 2026-09-01 to 2026-09-27, Perplexity named a business on 852 of 903 answered cell days (94.4 percent) and OpenAI on 351 of 781 (44.9 percent). Anthropic's raw lists hold generic phrases, which is why its raw persistence reads 50 percent. Use the gated figure for how often an engine names a business, and the raw one for what the engine wrote.
Source: metrics/metrics_v2.csv, methods N.gated.day and P.raw.pername (rows WA and WE)
Perplexity answers that cite a Namebeam owned domain, 2026-09-01 to 2026-10-13. 20 are one Kansas City health insurance page on namebeam.ai, 4 are trunkline.money and 2 are huggingface.co/Namebeam.
Google organic result pages listing one of our domains: 0 of 2,010. Google AI Overviews that carry reference links and cite one: 0 of 506 (the field exists from 2026-10-03, 10 run days). OpenAI, Anthropic and Gemini return no source list, so a zero for them means nothing and is not counted. The pack counts the two Hugging Face hits as ours (our_group huggingface.co/Namebeam).
Source: data/customer_zero_daily.csv (fields sources_cited, organic_top_results and ai_overview_references) and data/customer_zero_hits.csv
The pack is public: the answers table, data/answers_2026-09-01_to_2026-10-13.csv (12,180 rows, each with the SHA-256 of its raw file), the sources table (45,637 rows), the metrics table (204 rows), the code that computes the metrics, and a manifest of 32 files that a script checks every file against.
Open the pack in the public repository. Dated changes after release are logged in CORRECTIONS.md.
cd namebeam-visibility-index python3 releases/edition-1/verify.py python3 releases/edition-1/verify.py --rebuild --metrics
Seven things a reader can check, each with the file it comes from. Paths are inside the public repository; the pack is under releases/edition-1/. Counts on this page are read from PACK.json of the 2026-10-13 build (window 2026-09-01 to 2026-10-13). Added 2026-10-13, refreshed 2026-10-15.
The prompts table data/prompts.csv lists every question asked in the window, word for word, with its id, the SHA-256 of its text and the days it was in the roster: 73 questions. Every raw answer file in corpus/e1 also carries the question as sent in its prompt_text field, next to the answer, and the fixed question for each business or segment is the prompt_1 field of roster/e1_roster.json. One of them reads: "What is the best service to check if AI engines like ChatGPT recommend my small business?"
The rows record the model name that was requested, not the model that answered, for OpenAI, Anthropic and Gemini; Perplexity rows from 2026-09-28 on record the served model in engine_meta. The model strings are set in the collector code (scripts/engines.py) or in repository variables: OpenAI gpt-5-mini; Anthropic claude-haiku-4-5 unless a variable overrides it; Gemini gemini-3.6-flash; Perplexity ran on Sonar chat completions to 2026-09-27 and on the Agent API with model perplexity/sonar from 2026-09-28. Every row carries its engine name, so the API runs with no search tool (OpenAI, Anthropic, Gemini) and the web-searching run (Perplexity) can be read apart. Source: pack README, "How the questions were asked" and "Served model".
Window 2026-09-01 to 2026-10-13, UTC dates. 42 run days. Days with no run are absent from every table, not counted as zero; 2026-10-05 had no run. Two windows are reported apart: WA, 2026-09-01 to 2026-09-27, and WB, 2026-09-28 to 2026-10-13, split where Perplexity changed its API. Every row carries its date in the date column. Dated changes to the pack after release are listed in CORRECTIONS.md. Edition 1 is dated 2026-10-15.
United States. The engine answer rows cover 25 US local markets (Atlanta GA, Branson MO, Charlotte NC, Chicago IL, Columbus OH, Dallas TX, Denver CO, Fort Lauderdale FL, Houston TX, Joplin MO, Kansas City MO, Liberty MO, Los Angeles CA, Maryland, Miami FL, Nashville TN, Phoenix AZ, Pleasanton CA, San Francisco CA, Santa Cruz County CA, St. Charles MO, St. Louis MO, St. Peters MO, Tyler TX, West Palm Beach FL) plus 2 nationwide question sets (US nationwide; US nationwide, online), across 37 niches. Engine agreement rows carry no market. Every question is asked in English and no other language is in the record. Market and niche are columns on every row of data/answers_2026-09-01_to_2026-10-13.csv.
Sampling rules. One run per UTC day. Each engine is asked the same fixed question for every roster entry, once, and one raw file is written per answer. These are API calls, not the apps people open on a phone. Settings per engine (call method, search tool, temperature) are in the pack README table, as set in scripts/engines.py. The metrics read the market and niche cells asked on or before 2026-09-27; US nationwide questions are left out of the cell metrics; API rows count and SERP and AGREE rows do not. Scoring method. Every metric names its method id, numerator and denominator in metrics/metrics_v2.csv; the definitions are in metrics/metrics_v2_methods.csv. Names pass through a rule-based gate in method/Brain_Kit/status/staged/1005A_AI_PICKS_CHART/gen_ai_picks_chart.py (SHA-256 611894c48c21a512bc856463c2e4b6b7a33155375aebcce71a036e820a4edde6), and every persistence and overlap metric is given twice, raw and gated. The known under-count of names in the derived tables (DQ-5) and the text-mention recount that corrects it are stated in the pack README.
Every answer is saved word for word as one JSON file in corpus/e1, in the answer_verbatim field, with the engine, date, market, niche and prompt beside it. The answers table, data/answers_2026-09-01_to_2026-10-13.csv (12,180 rows), has one row per file with source_file and sha256 columns, so any row can be matched to its file and the file checked against its hash. verify.py --rebuild rebuilds the table from the files.
The record is released under Creative Commons Attribution 4.0 (CC BY 4.0): LICENSE at the repository root and in the pack. Quote and reuse it with a credit line to Namebeam, Carter Enterprise LLC, and a link to the record. A citation form is in CITATION.cff.