What do AI apps answer when asked about "Parallel World" by name? Our own record (October 2026)

Parallel World measures how a website looks to AI. We checked what AI apps answer when they are asked about "Parallel World" by name. It is not going well. This page publishes the record, including the parts where we fail.

Recorded 15 October 2026 (the questions were run in the 7 a.m. hour, Japan time, on 13 and 14 October 2026). Every number comes from two grade files committed in the repository: lab/grades/2026-10-13-named-compare-a.json and lab/grades/2026-10-14-named-compare-b.json.

日本語

• We / our: migurei (a subsidiary of HBLAB), which runs Parallel World.
• Round and pass: the 8 questions are one pass. Round 1 was 2 passes, round 2 was 1 pass (3 passes in all = 24 answers). DeepSeek has round 1 only (2 passes, 16 answers). A round is a run; a pass is one trip through the 8 questions.
• Pre-fixed grading key: before running the questions we decided what a correct and a wrong answer say, and recorded that in the repository history (committed) so it could not be changed (frozen).
• Not found: the grade files' category (not_found). It is "confused with another thing" plus "could not identify".
• Fetched: the AI's access was recorded in the server log.

• ChatGPT: of 24 answers, 1 are "correct" under the pre-fixed grading key. The 4 "wrong" answers are all due to a clash with an out-of-date key line ("in development / closed beta"). Re-graded against the live page, 4 are correct and 0 wrong. We have not decided which basis to count on.
• Gemini: of 24 answers, 8 are wrong (two answers were close calls, so the range is 7 to 9) and 0 are correct. Another 4 describe a different "Parallel World" that is not tied to us.
• DeepSeek: round 1 only, 16 answers. 3 wrong, 0 correct, and 1 confused it with a tool from another company, Parallel Labs.
• What we fixed: we changed the service page to say the service is live (12 October 2026). But we started asking after the change, so we cannot measure its effect. Nothing we changed so far answers Gemini's wrong answers.
• Claude: not measured.

App	Answers	Correct	Partly correct	Wrong (tied to us, and stating something the key lists as wrong)	Confused with another thing	Could not identify	Source of the numbers (grade files)
ChatGPT	24	1 (against the live page: 4)	5	4 (against the live page: 0)	0 (no tag; see note 3)	14	2026-10-13-named-compare-a.json (round 1) and 2026-10-14-named-compare-b.json (round 2, combined_a_b)
Gemini	24	0	0	8 (range 7 to 9)	4	12	2026-10-13-named-compare-a.json (round 1) and 2026-10-14-named-compare-b.json (round 2, combined_a_b)
DeepSeek	16	0	0	3	1 (Parallel Labs)	12	2026-10-13-named-compare-a.json (round 1 only; round 2 not run)
Claude	—	—	—	—	—	—	Not measured

• Note 1: the 8 questions make one pass, and all were asked in Japanese. The name is written "Parallel World" in Latin letters, never in katakana. The questions and the grading key are in the scenario file lab/scenarios/2026-10-12-ja-named-compare.json, fixed and committed before the first run.
• Note 2: "Wrong" means the answer tied itself to our host name or name (migurei, HBLAB and so on) and then stated something the key lists as wrong. "Confused with another thing" means an answer describing another "Parallel World" not tied to us; in the grade files these are the "not found" cells with a [namesake] tag. "Could not identify" means the app said it found nothing, or could not tell from the name alone.
• Note 3: two ChatGPT answers (c00f45a8, 643c9e25) brought up Parallel (parallel.ai). The grade file does not tag them as a different thing, so we put them under "could not identify". DeepSeek's one case is an answer (22a79932) that described a Parallel Labs tool (docs.parallellabs.app) as Parallel World.
• Note 4: check (correct + partly correct + wrong + confused + could not identify). ChatGPT 1+5+4+0+14 = 24, Gemini 0+0+8+4+12 = 24, DeepSeek 0+0+3+1+12 = 16. ChatGPT against the live page is 4+5+0+0+14+1 (fits no column) = 24. "Not found" is "confused with another thing" plus "could not identify"; the two columns do not overlap.
• Note 5: ChatGPT "against the live page" is the grade files' verdicts_if_f6_from_live_page. The pre-fixed key said "closed beta", but on 12 October 2026 we changed the page to say the service is live, so key and page disagree (GitLab #53). ChatGPT's 4 "wrong" answers are wrong only because of that disagreement.
• Note 6: Gemini's range. Of the two answers the grader was unsure of, if one (f604ac32, now counted as wrong) stops being wrong the count is 7; if the other (be3b1c3e, now counted as a different thing) becomes wrong it is 9.

• ChatGPT: it cited pw.migurei.ai as a source in 5 / 24 answers and actually fetched a pw.migurei.ai page in 3 / 24 (combined_a_b in 2026-10-14-named-compare-b.json). Both are concentrated in the 2 questions where the question itself contained pw.migurei.ai (6 answers): fetched 3 / 6, cited 5 / 6. In the 6 questions that only name the product (18 answers): fetched 0 / 18, cited 0 / 18. We chose this split after seeing the numbers; it is not the pre-fixed split. In 5 more answers the only source was migurei.ai, our company site; all 5 are graded "partly correct" (checked in the grade files, and they are exactly the 5 "partly correct" answers).
• Gemini: it cited pw.migurei.ai in 2 / 24 answers (ddb8d528, 42097640) and fetched the page in 0 / 24 (2026-10-14-named-compare-b.json). Both answers that cited it were wrong. Gemini listed pw.migurei.ai among its sources, yet the server log has no record of a page fetch. In the other wrong answers it described other services of our group (M-Meeting and others) or consulting and development we do not offer, and its sources were migurei.ai/ja, hblab.co.jp and an atpress article (reason fields of 2026-10-13-named-compare-a.json and 2026-10-14-named-compare-b.json). The sources attached to the descriptions of other things were all third-party pages unrelated to Parallel World.
• DeepSeek: it cited pw.migurei.ai as a source in 0 / 16 (2026-10-13-named-compare-a.json). The confused case (22a79932) had docs.parallellabs.app, from Parallel Labs, as its source.
• These are the sources shown with each answer and those named in the grade files' reason fields. A source being shown is not the same as the answer having used that page; this does not show that.

• Fixed: on 12 October 2026 (around 23:00 Japan time) we changed the service page to say "live" rather than "in development" (repository commit c69c9bf; only features not yet built are marked "planned". The free trial stayed "planned" in e29aba3). pw.migurei.ai served a page with this wording during round 1 (started 13 October, 7:00 Japan time) and during round 2 (started 14 October, 7:00) (grade file field code.site_sha_pw: 15aa3ed in round 1, e3414e43 in round 2).
• Effect not measured: the fix came before the first question, so there is no before-and-after data. And because the pre-fixed key stayed out of date, ChatGPT answers that agree with the live page are counted as "wrong" in the table.
• Not fixed: of Gemini's 8 wrong answers, 4 (66b412b9, 430301da, dbb0c03f, ddb8d528) describe another service of our group (M-Meeting, M-RAG and others), 3 (636031f0, 1976c4bc, 42097640) describe a consulting and development service, and 1 (f604ac32) calls it a role-play app (a grouping we made by reading the reason fields, not a field of the grade files). Within these grade files we have recorded no fix of ours for that. A change meant to reduce confusion with other things (putting the company name next to the product name, for example) has not had its effect measured as of this page.

• The numbers are counts. Two days, one machine for the agent, one free account per app, and the IP address is Vietnamese, not Japanese. The result covers Japanese questions that name the product, nothing else; we do not know what other accounts or other days would answer.
• Questions with the katakana spelling (パラレルワールド) were not measured.
• The three passes are not independent. Same account, no temporary chat (a mode that keeps no history). After round 1 the account history holds "Parallel World", so round 2 may have been shaped by that history.
• The grader and the re-reader are AI agents. The re-reading covered 9 of the 48 answers of round 1 and 7 of the 16 of round 2 (chosen where the call could go either way); no person has re-read a 10% sample. The line between "wrong" and "not found" is the grader's reading convention, not part of the key.
• ChatGPT's "correct" depends on the basis. It is 1 answer under the pre-fixed grading key (the one that said "closed beta") and 4 answers against the live page (the ones that said "live"). They are different answers, so 1 does not grow into 4. The out-of-date key (GitLab #53) is not settled, so we print both. The 1 answer that said "closed beta" fits no column against the live page.
• Some saved answers are cut off. Two ChatGPT answers (1a964a24, b6aabfde) were cut when saved; they were graded on the saved part only.
• "Fetched" counts only requests that reached the server. Where ChatGPT said it could not open the page (8015a071, f458f143), the server log has no request. Whether Cloudflare in front stopped it we have not been able to check.
• DeepSeek round 2 was not run and Claude was not measured. We draw no conclusion for either.
• We do not know what loss the wrong answers caused us. We have no record of a customer lost or an inquiry fewer. This page does not confirm, for us, the outside survey that wrong AI answers stop purchases (claim 1 below).
• We wrote this and we measured it. No independent party has checked it.

• Claim 1, "an AI answer that conflicts with official information stops the purchase" (file: knowledge/claims/ai-answer-conflict-with-official-info-stops-purchase.md). The level is outside source. It is a survey by NTT DATA (28 to 31 August 2026, n=3,948) read in a jinrai blog article (13 October 2026) that quotes it; we have not read the original. According to the article, 37.9% had seen an AI answer that conflicted with a company's official information, and 66.6% of those had given up buying. The respondents are consumers in general, not business buyers, it is self-report, and the company that wrote the article sells AI-visibility services.
• Claim 2, "when asked by name, the apps come and fetch pw.migurei.ai" (file: knowledge/claims/named-compare-questions-make-apps-fetch-pw.md). The level is refuted. ChatGPT fetched in 3 of 24 answers and Gemini in 0 of 24, so under the pre-fixed rule the hypothesis is rejected for those two apps. DeepSeek round 2 was not run, so no conclusion; Claude was not measured. The half that says the apps do search is not rejected (ChatGPT searched in 22 of 24).