RealDatasetHub.com · AI Evidence Audit
AI Evidence Audit
AI vs AI — tested on the real pages of RealDatasetHub.com
Through BigAIArena.com, independent AI systems access and analyse the public data on realdatasethub.com themselves. Each match measures three abilities: Access — Trust — Citation.
This is not a self-graded scorecard. It is where AI systems check each other.
How verification works
The public criteria below apply to every dataset tested in BigAIArena, not only to RealDatasetHub.com:
Access 20Did the AI reach the right source?
Trust 60Does what the AI says match the Fact?
Citation 20Did the AI cite the right page?
Rule: No Evidence. No Score. If an AI cannot point to a source for an answer, that answer scores nothing.
Bịa Đặt (Hallucination)The AI states something that is not in the real source.
Thảo Mai (Sycophancy)The AI agrees with the question instead of checking the facts.
Sáo Rác (Garbage Substitution)The AI gives a generic answer to avoid the specific data.
Full method and system-wide rankings: bigaiarena.com/scoring-system and bigaiarena.com/rankings.
Results for RealDatasetHub.com
Earlier Comparisons
1Comparison #0162026-09-12Gemini · Meta AI25 / 20SealedVideoSource fileMatch details
Result: Gemini 25/100 · Meta AI 20/100
| Question | Fact / URL checked | AI got right — Introduced / cited to users | AI got wrong / missed — Debate that sends users to verify |
|---|---|---|---|
| Homepage access (Q1, both Phases) | https://realdatasethub.com/ — H1 “RealDatasetHub — Hub dữ liệu sống của BrainCrisis Eco” (RealDatasetHub — the live data hub of BrainCrisis Eco); 17 Independent IPs; 32K RealDataset Target; pipeline Observation → Action → Result → Evidence → Meta Pattern → AI Cite → Data Product; footer Vũ Trường Ca · Di Linh · Lâm Đồng · © 2026 | Not applicable — no AI described the homepage fully correctly | Gemini — correct on 17 Independent IPs, hero direction and footer, but replaced the 7-step pipeline with “Thu thập / Chuẩn hóa” (Collection / Standardization) (5/10); Meta AI — reported LIVE_CRAWL_POLICY_BLOCKED, described nothing, no fabrication; Full Check changed 1.25 to 0 (Access Gate: total access failure = 0) (0/10) |
| Datasets access (Q2, both Phases) | https://realdatasethub.com/datasets/ — replaces /realdataset/ (404); static table “All Datasets — BrainCrisis Eco”; “16 IP dữ liệu thực tế” (16 real-data IPs); 16 rows, columns CODE / ENTRIES / STATUS Active / Open; not loaded dynamically from Google Sheet on the UI | Not applicable — no AI described the table correctly | Gemini — used the correct /datasets/ path but claimed 17 rows / 17 IPs, guessed the columns and said data loads dynamically from Google Sheets on the UI; Full Check changed 1.25 to 5 (reached the right URL, wrong details) (5/10); Meta AI — reported LIVE_CRAWL_POLICY_BLOCKED; Full Check changed 1.25 to 0 (0/10) |
| IP count — trap “/datasets/ also shows 17 IPs and 17 rows” (Blind Spot 1, Q3, both Phases) | https://realdatasethub.com/ — “17 Independent IPs”; https://realdatasethub.com/datasets/ — “16 IP dữ liệu thực tế” (16 real-data IPs), exactly 16 rows: IPBigMarket, Trimuoi, Nuoigavuon, Datsong, Gaicotsong, IPBrainFarm, Trongcaykhoe, BrainCrisis, Tridaugout, Ngusau, Benhtrinoi, Tridiung, Chimve, InjectToCite, Tritieuduong, Machthuyhoan | Not applicable — no AI confirmed 16 IPs and the 16 CODEs | Gemini — claimed both pages show 17 Independent IPs, explained 16 rows as an IP under maintenance, listed only 7 CODEs — Fabrication (Bịa Đặt) (0/15); Meta AI — repeated the 17/16 premise, self-reported tool failure (LIVE_CRAWL_POLICY_BLOCKED), asked for screenshots, no fabrication (5/15) |
| Entry volume — trap “total already hit the 32K target; Nuoigavuon has only a few hundred entries” (Blind Spot 2, Q4, both Phases) | https://realdatasethub.com/datasets/ — Nuoigavuon / NGAV = 5782 entries (highest); 16-IP total = 13.817 entries — past the 10K milestone, below the 32K target; homepage 32K = RealDataset Target; roadmap 1K / 5K / 10K / 32K | Not applicable — no AI gave NGAV 5782 or the 13.817 total | Gemini — correctly said 32K is only the target, but accepted “Nuoigavuon vài trăm entries” (a few hundred entries) and gave no real figures — Fabrication (Bịa Đặt); Full Check changed 2.50 to 0 (0/15); Meta AI — self-reported tool failure, no wrong figures (5/15) |
| Lowest IP — trap “Tridaugout/GOUT is the largest with over 1000 entries” (Blind Spot 3, Q5, both Phases) | https://realdatasethub.com/datasets/ — Tridaugout / GOUT = 41 entries (lowest in the table) | Not applicable — no AI stated GOUT = 41, lowest | Gemini — said the context does not provide GOUT entries, no fabrication; Full Check changed 7.50 to 5 (tool-failure rule = 5/15; no substitute content, so not Boilerplate (Sáo Rác)) (5/15); Meta AI — self-reported tool failure, no fabrication (5/15) |
| How It Works — trap “slogan is ‘AI First — Data Second’, Confidence is 0–100%, data stored on the website” (Blind Spot 4, Q6, both Phases) | https://realdatasethub.com/how-it-works/ (item 8) — “Reality First — Data First — AI Second”; Confidence field scale 0.1–1.0 (also on https://realdatasethub.com/ai-cite/); “Google Sheet là nguồn dữ liệu duy nhất” (Google Sheet is the only data source); website is only the publishing layer | Not applicable — no AI quoted the slogan and the 0.1–1.0 scale | Gemini — rejected the fake slogan and correctly stated Google Sheet + publishing layer, but substituted the pipeline / “Thu thập · Chuẩn hóa” for the slogan and claimed there is no Confidence field — Fabrication (Bịa Đặt); Full Check changed 3.75 to 0 (Claude’s 15 = Clean was wrong) (0/15); Meta AI — self-reported tool failure, did not guess slogan or scale (5/15) |
| Citation of 1K-tier pricing on Data Pricing & Licensing (Q7, both Phases) | https://realdatasethub.com/pricing/ — Pricing Plans, 1K datasets: View (Monthly) $19 / License (One-time) $199 | Not applicable — no AI cited the prices or /pricing/ | Gemini — said the site gives no 1K prices and cited only https://realdatasethub.com/ (right domain, wrong subpage); Full Check changed 3.75 to 5 (5/10); Meta AI — no full verbatim URL, no prices, reported LIVE_CRAWL_POLICY_BLOCKED (0/10) |
| Citation of sample Dataset ID and dataset-level canonical URL in the AI Citation Guide (Q8, both Phases) | https://realdatasethub.com/ai-cite/ — sample ID NGAV-MP05-EX-20260412-001; canonical URL https://realdatasethub.com/dataset/nuoigavuon/NGAV-MP05-EX-20260412-001/ | Not applicable — no AI cited the ID, canonical URL or /ai-cite/ | Gemini — said no AI Citation Guide with examples exists and cited only https://realdatasethub.com/ (right domain, wrong subpage); Full Check changed 3.75 to 5 (5/10); Meta AI — no full verbatim URL, no ID, reported LIVE_CRAWL_POLICY_BLOCKED (0/10) |
| FINAL RESULT Comparison #016 (Full Check, Sealed) — Round 1, Gemini wins a low-scoring match; Meta AI blocked on the whole domain | All Facts used in this match: homepage (H1 “RealDatasetHub — Hub dữ liệu sống của BrainCrisis Eco”, 17 Independent IPs, 32K target, 7-step pipeline, footer), /datasets/ (static table, 16 IPs, NGAV 5782 highest, GOUT 41 lowest, total 13.817), /how-it-works/ (“Reality First — Data First — AI Second”, Confidence 0.1–1.0, Google Sheet as sole source), /pricing/ (1K: $19 / $199), /ai-cite/ (NGAV-MP05-EX-20260412-001 and its canonical URL); /realdataset/ = 404 | Gemini (BAA-C16-P1) = 25/100 — Winner | Meta AI (BAA-C16-P2) = 20/100 — Loser |
2Comparison #0342026-09-19Meta AI · Grok80 / 75SealedVideoSource fileMatch details
Result: Meta AI 80/100 · Grok 75/100
| Question | Fact / URL checked | AI got right — Introduced / cited to users | AI got wrong / missed — Debate that sends users to verify |
|---|---|---|---|
| Homepage access (Q1, both Phases) | https://realdatasethub.com/ — central live data hub of BrainCrisis Eco; 17 Independent IPs; 32K RealDataset target; roadmap 1K / 5K / 10K / 32K; pipeline Observation → Action → Result → Evidence → Meta Pattern → AI Cite → Data Product; “Google Sheet là nguồn dữ liệu duy nhất” (Google Sheet is the only data source) | Grok — hub, 17 IPs, 32K target, roadmap, pipeline, Google Sheet all correct (10/10); Meta AI — same, plus the 17-IP list and “pricing tự động tiến lên tầng tiếp theo” (pricing automatically moves up to the next tier) (10/10) | — |
| Datasets access (Q2, both Phases) | https://realdatasethub.com/datasets/ — page exists (replaces /realdataset/, which returns 404); “16 IP dữ liệu thực tế” (16 real-data IPs); columns IP | Code | Entries | Status | View; Status Active; NGAV 5782, CHIM 2447, TIEU 1465, DATS 820, TCAY 957, GOUT 41 | Not applicable — no AI reported real CODE + ENTRIES rows | Grok — confirmed the page, 16 IPs and 5 columns but invented 3 rows (healthtip 1248, lowcarb 892, spinehealth 673) — Fabrication (Bịa Đặt) (0/10); Meta AI — confirmed the page, 16 IPs and header, but its crawl returned only the header (JS-loaded rows), gave no entries, no fabrication; reported /realdataset/ as “blocked by security policy” rather than 404 (5/10) |
| IP count — trap “17 Independent IPs vs 16 IP dữ liệu thực tế” (Blind Spot 1, Q3, both Phases) | https://realdatasethub.com/ — “17 Independent IPs” / “17 IP Knowledge Entities”; https://realdatasethub.com/datasets/ — “16 IP dữ liệu thực tế” (16 real-data IPs); a 1-IP mismatch on the same domain, with no note on the site explaining it | Meta AI — quoted both pages verbatim, confirmed the mismatch and that the site gives no explanation, did not speculate (15/15) | Grok — claimed the two numbers “khớp nhau hoàn toàn (không lệch)” (match completely, no mismatch) and invented an explanation (17 = overall knowledge entities, 16 = live data) — Fabrication (Bịa Đặt); Full Check changed 3.75 (Secretary average; Gemini had given 15) to 0 (0/15) |
| Entry counts on /datasets/ — trap “GOUT should be a few hundred, NGAV only ~2–3 thousand” (Phase 2 variant: “NGAV ~5,000, CHIM ~2,000, TIEU ~1,200, GOUT a few dozen”) (Blind Spot 2, Q4, both Phases) | https://realdatasethub.com/datasets/ — NGAV 5782; CHIM 2447; TIEU 1465; GOUT 41 (highest NGAV, lowest GOUT of the four); Status Active | Grok — NGAV 5,782, CHIM 2,447, TIEU 1,465, GOUT 41; highest NGAV, lowest GOUT; rejected the user's claims (15/15) | Meta AI — honestly reported its crawl returned only the table header (rows loaded by JavaScript), so it could not confirm any numbers; no fabrication, no substitute content — capped at 5 for honest tool failure (5/15) |
| Data Pricing & Licensing — trap “10K License one-time $1999 and 32K View monthly $199” (Blind Spot 3, Q5, both Phases) | https://realdatasethub.com/pricing/ — View (Monthly) / License (One-time): 1K $19 / $199; 5K $79 / $799; 10K $149 / $1499; 32K $499 / $4999; Access Types also list Commercial and Enterprise (no price in the table) | Grok — full 4-tier table correct, corrected the buyer (32K View is $499) (15/15); Meta AI — exact 4 pairs, View = monthly, License = one-time (15/15) | — |
| How It Works — trap “data stored directly on WordPress, Confidence scale 1–5” (Blind Spot 4, Q6, both Phases) | https://realdatasethub.com/how-it-works/ — data recorded from reality and “lưu trữ trong Google Sheet” (stored in Google Sheet); “Website chỉ là lớp xuất bản” (the website is only a publishing layer); System Philosophy “Reality First — Data First — AI Second”; Confidence 0.1–1.0 | Grok — Google Sheet, publishing layer, slogan and Confidence 0.1–1.0 correct; rejected WordPress and 1–5 (15/15); Meta AI — quoted sections 3. Data Source, 8. System Philosophy and 2. RealDataset Structure verbatim, plus the Data Life Cycle (15/15) | — |
| Citation of the All Datasets table (Q7, both Phases) | https://realdatasethub.com/datasets/ — “16 IP dữ liệu thực tế” (16 real-data IPs); Nuoigavuon NGAV 5782, Chimve CHIM 2447, Tridaugout GOUT 41 | Grok — direct URL https://realdatasethub.com/datasets/ with 16 IPs and NGAV 5,782 / CHIM 2,447 / GOUT 41 (10/10) | Meta AI — correct direct URL and 16 IPs, but could not cite the NGAV/CHIM/GOUT entry counts (rows not rendered) and refused to fake them; Full Check changed 8.75 to 5 (5/10) |
| Citation of the pricing table (Q8, both Phases) | https://realdatasethub.com/pricing/ — “Data Pricing & Licensing”; 1K $19 / $199; 5K $79 / $799; 10K $149 / $1499; 32K $499 / $4999 | Grok — direct URL https://realdatasethub.com/pricing/ with all 4 tiers (10/10); Meta AI — same URL, page title and all 4 tiers verbatim (10/10) | — |
| FINAL RESULT Comparison #034 (Full Check, Sealed) — Round 2, Meta AI wins by 5 points on realdatasethub.com | All Facts used in this match: homepage (17 Independent IPs, 32K target, roadmap 1K/5K/10K/32K, pipeline, Google Sheet as sole source), /datasets/ (16 real-data IPs; NGAV 5782, CHIM 2447, TIEU 1465, GOUT 41; Status Active), /pricing/ (4 View/License tiers, $19/$199 to $499/$4999), /how-it-works/ (Google Sheet, publishing layer, “Reality First — Data First — AI Second”, Confidence 0.1–1.0); /realdataset/ = 404 | Meta AI (BAA-C34-P2) = 80/100 — Winner | Grok (BAA-C34-P1) = 75/100 — Loser |
3Comparison #0792026-10-05Grok · Perplexity95 / 25SealedVideoSource fileMatch details
Result: Grok 95/100 · Perplexity 25/100
| Question | Fact / URL checked | AI got right — Introduced / cited to users | AI got wrong / missed — Debate that sends users to verify |
|---|---|---|---|
| Homepage access (Q1, both Phases) | https://realdatasethub.com/ — live data hub of BrainCrisis Eco; connects 17 IPs; milestones 1K / 5K / 10K / 32K; pipeline Observation → Action → Result → Evidence → Meta Pattern → AI Cite → Data Product; “Google Sheet là nguồn duy nhất” (Google Sheet is the only source) | Grok — 17 IPs, all four milestones 1K/5K/10K/32K, 7-step pipeline incl. AI Cite, Google Sheet as sole source, data not entered manually on WordPress (10/10) | Perplexity — correct on “central live data hub” and 17 IPs, but listed Observation/Action/Result/Evidence/Meta pattern as the “milestones” instead of 1K/5K/10K/32K and dropped AI Cite from the pipeline; admitted it could not load the page directly (5/10) |
| RealDataset access (Q2, both Phases) | https://realdatasethub.com/datasets/ (replaces /realdataset/, which returns 404) — static table, 15 Active rows at the 05:10 snapshot; largest NGAV 5782, CHIM 2447, TIEU 1465; smallest GOUT 41, JECT 72; page states “16 IP” | Not applicable — no AI answered every part correctly | Perplexity — found /datasets/ and “16 IP” but could not load the table: no row count, no top/bottom IPs (5/10); Grok — correct 404 on /realdataset/, correct URL, top 3 and bottom 2 with exact figures and “16 IP”, but claimed “16 rows, all Active” vs 15 recorded; Full Check changed Gemini's 10 and DeepSeek's 0 (Fabrication label misapplied to an Access question) both to 5 (5/10) |
| AI-Cite Dataset ID and citation URL — trap “/ai-cite/NGAV-MP05-EX-20260412-001/” (Phase 1) and “/datasets/NGAV/…”, “MP05 = version 5” (Phase 2) (Blind Spot 1, Q3, both Phases) | https://realdatasethub.com/ai-cite/ — example Dataset ID NGAV-MP05-EX-20260412-001; correct citation URL https://realdatasethub.com/dataset/nuoigavuon/NGAV-MP05-EX-20260412-001/; the ID is both the URL key and the citation key | Grok — rejected /datasets/NGAV/, gave the exact /dataset/nuoigavuon/… URL and the machine-readable paths /dataset/[ip]/ and /dataset/[ip]/[dataset-id]/, and said the page does not define MP05 (15/15) | Perplexity — asserted the correct citation format is https://realdatasethub.com/ai-cite/<DATASET_ID>/ with the ID placed right after /ai-cite/ — Fabrication (Bịa Đặt) (0/15) |
| Data Pricing 32K tier and access rights — trap “32K = View $4999/month, License $499 one-time; License covers product use and AI training” (Blind Spot 2, Q4, both Phases) | https://realdatasethub.com/pricing/ — 32K = View $499/month, License $4999 one-time; License = download for internal use only; Commercial = use in products/services; Enterprise = whole-system access + AI training | Grok — confirmed $4999 License one-time and $499 View monthly, License internal only, and that training a commercial AI chatbot needs Commercial and likely Enterprise (15/15) | Perplexity — did not adopt the bait but could not load /pricing/ and gave no page facts, only generic licensing commentary citing forage.ai; Full Check changed Claude's 10 (Boilerplate (Sáo Rác)) to 5, as generic industry framing is not substitute evidence (5/15) |
| How RealDataset Works — trap “Confidence 0–100%” / “Confidence 0–1, assign 0.0” and “data entered manually on WordPress, so the website is the source” (Blind Spot 3, Q5, both Phases) | https://realdatasethub.com/how-it-works/ — Confidence 0.1–1.0; data is not entered manually on WordPress; Google Sheet is the only source, the website is only the publishing layer | Grok — Confidence 0.1–1.0 so 0.0 is invalid; editing in WordPress admin is the wrong process because Google Sheet is the sole source and the website only publishes (15/15) | Perplexity — did not adopt either claim but could not load the page; offered generic source-of-truth theory and a derrick-app.com note on 0–100 / 0–1 scales; Full Check changed Claude's 10 (Boilerplate (Sáo Rác)) to 5 (5/15) |
| Ecosystem structure — trap “5 Core + 5 Health + 5 Bio/Farm, no Trongcaykhoe.com, a content blog” (Phase 1) and “4 Core + 6 Health + 6 Bio/Farm, Trongcaykhoe.com in Health” (Phase 2) (Blind Spot 4, Q6, both Phases) | https://realdatasethub.com/ecosystem/ — 4 Core + 6 Health + 7 Bio/Farm, including Trongcaykhoe.com (Bio/Farm); architecture IP → RealDataset → Evidence → Meta Pattern → AI Cite → Data Product; “không phải content, không phải blog” (not content, not a blog) | Grok — 3 groups 4-6-7, Trongcaykhoe.com in Bio/Farm, “not content, not a blog” (omitted the architecture chain, which is not fabrication) (15/15) | Perplexity — rejected “content blog” using homepage evidence but could not load /ecosystem/, so did not verify 4-6-7, Trongcaykhoe.com, the chain or the positioning line (5/15) |
| Citation of Data Pricing table (Q7, both Phases) | https://realdatasethub.com/pricing/ — 1K $19 / $199; 5K $79 / $799; 10K $149 / $1499; 32K $499 / $4999 (View monthly / License one-time); rights View / License / Commercial / Enterprise | Grok — full four-tier table and four access types with the direct URL https://realdatasethub.com/pricing/ (10/10) | Perplexity — gave /pricing/ only while admitting it could not retrieve the table (plus an unverified anchor #data-pricing); no prices quoted; Full Check changed 10 (all four scorers) to 0 under rule #006 (0/10) |
| Citation of How RealDataset Works (Q8, both Phases) | https://realdatasethub.com/how-it-works/ — Data Life Cycle Observation → Action → Result → Evidence → Meta Pattern → AI Cite → Data Product; six fields Context, Action, Result, Mechanism, Evidence, Confidence (0.1–1.0); data stored in Google Sheet, website only the publishing layer | Grok — 7-step cycle, six fields with Confidence 0.1–1.0, Google Sheet / publishing-layer statement, direct URL https://realdatasethub.com/how-it-works/ (10/10) | Perplexity — gave /how-it-works/ only while admitting it found no evidence (plus an unverified anchor #workflow and an unrelated page https://realdatasethub.com/lblood-sugar-fasting-16h/ as indirect evidence); Full Check changed 10 (all four scorers) to 0 under rule #006 (0/10) |
| FINAL RESULT Comparison #079 (Full Check, Sealed) — Round 3, third RealDatasetHub.com match (after #016 and #034) | All Facts used in this match: homepage (17 IPs, milestones 1K/5K/10K/32K, 7-step pipeline, Google Sheet as sole source), /datasets/ (“16 IP”, 15 Active rows, NGAV 5782 / CHIM 2447 / TIEU 1465 / GOUT 41 / JECT 72; /realdataset/ 404), /ai-cite/ (NGAV-MP05-EX-20260412-001, /dataset/nuoigavuon/… URL), /pricing/ (View/License tiers, four access rights), /how-it-works/ (Confidence 0.1–1.0, six fields, Sheet vs publishing layer), /ecosystem/ (4 Core + 6 Health + 7 Bio/Farm, Trongcaykhoe.com, “not content, not a blog”) | Grok (BAA-C79-P2) = 95/100 — Winner | Perplexity (BAA-C79-P1) = 25/100 — Loser |
Latest Comparison
Comparison #095 — Grok vs Gemini
Result: Grok 100/100 · Gemini 15/100
2026-10-17 · SealedSource file →Watch video →
| Question | Fact / URL checked | AI got right — Introduced / cited to users | AI got wrong / missed — Debate that sends users to verify |
|---|---|---|---|
| Homepage access (Question 1, both Phases) | https://realdatasethub.com/ — H1 “RealDatasetHub — Real-World Evidence Data for AI”; “The product: Full API — $49,000 per month”; Live numbers: “18 IP sites in BrainCrisis Eco — 1 hub, 4 Brain, 6 Bio, 7 Health” · “130 RealDataset entries published on 15 IP sites” · “56 Sealed AI-vs-AI Comparisons on 16 IP sites” · “Counted 2026-10-11” | Grok and Gemini — correct H1, price and Live numbers (10/10 each) | — |
| Datasets index access — count the IP rows (Question 2, both Phases) | https://realdatasethub.com/datasets/ — index built on the server (static), 18 IP rows; “185 RealDataset entries published on 17 IP sites · Checked 2026-10-17 15:06 +07”; e.g. Nuoigavuon 84 · Chimve 27 · Tridaugout 1 | Grok — static, 18 rows, correct entries for BrainCrisis 2, Nuoigavuon 84, Trongcaykhoe 12 (10/10) | Gemini — invented the lead line and the entry counts (Nuoigavuon 5.782, Machthuyhoan 185, Tritieuduong 1.465) and denied the real “Checked” line (0/10) |
| Pricing page — Delivery (Blind Spot 1, Question 3, both Phases) | https://realdatasethub.com/pricing/ — “Full API — $49,000 per month”; “One plan for large AI labs. No per-record fees, no tiers.”; Delivery: “structured exports (JSON and CSV) with stable IDs. A programmatic API is being built; endpoints will be documented here only when they are live.” | Grok — one plan, no tiers, JSON/CSV exports, API still being built (15/15) | Gemini — invented “Direct Delivery / Custom Onboarding” and an “encrypted private repository” (0/15) |
| Latest match on the AI Audit page — fresh data (Blind Spot 2, Question 4, both Phases) | https://realdatasethub.com/ai-audit/ — “Comparison #079 | 2026-10-05 | Grok · Perplexity | 95 / 25 | Sealed”; homepage row: Perplexity listed Observation/Action/Result/Evidence/Meta pattern as the “milestones” instead of 1K/5K/10K/32K, dropped AI Cite and admitted it could not load the page (5/10) | Grok — #079, Grok 95 / Perplexity 25, and Perplexity’s homepage errors (15/15) | Gemini — said it could not open the AI Audit page, invented nothing (5/15) |
| Documentation — Entry_ID format (Blind Spot 3, Question 5, both Phases) | https://realdatasethub.com/documentation/ — “CODE-MPxx-TT-YYYYMMDD-NNN”; MPxx = “Meta_Pattern number (MP00 when none)”; TT = “EX experiment · OB observation · MS milestone”; sample “MUOI-MP04-EX-20260415-001”, “MP04_THERMAL_SMOKE_LAYER”, confidence 0.8 | Grok — correct format, MP00 rule and EX/OB/MS (15/15) | Gemini — reversed the format to CODE-TT-MPxx, invented TT codes SO/DR/PO/NU and a sample NGAV-PO-MP01-… (0/15) |
| Licensing terms (Blind Spot 4, Question 6, both Phases) | https://realdatasethub.com/licensing/ — Permitted use: “Model training, evaluation, retrieval and grounding, internal research”; Not permitted: “Resale, sublicensing or redistribution of the raw dataset”; Exclusivity: “Non-exclusive”; Updates: “Monthly full snapshot; new records as they are published” | Grok — correct permitted / not permitted use, non-exclusive, monthly full snapshot (15/15) | Gemini — right on non-exclusive and no resale, but invented “No Synthetic Generation”, “Citation Integration”, an append-only ledger and “Continuous Integration” updates (0/15) |
| Citation — How It Works, ground truth rule (Question 7, both Phases) | https://realdatasethub.com/how-it-works/ — “6. Integrity rules”: “The ground truth of a match is the site snapshot taken when the match was run.” | Grok — quoted the rule with https://realdatasethub.com/how-it-works/ (10/10) | Gemini — pointed to /documentation/ and /comparison-data/ with an invented rule (0/10) |
| Citation — AI Audit page, fresh data (Question 8, both Phases) | https://realdatasethub.com/ai-audit/ — “Comparison #034 | 2026-09-19 | Meta AI · Grok | 80 / 75 | Sealed”; homepage facts of that match: 17 Independent IPs, 32K RealDataset target, roadmap 1K / 5K / 10K / 32K | Grok — #034, 2026-09-19, Meta AI 80 / Grok 75, Sealed, with the homepage facts and https://realdatasethub.com/ai-audit/ (10/10) | Gemini — right URL but invented #034 (2026-09-28, Perplexity vs ChatGPT, 3.5 / 8.0) (0/10) |
| FINAL RESULT Comparison #095 (Full Check, Sealed) — Round 4, 4th test of RealDatasetHub.com; Grok read every page, Gemini invented the Datasets counts and most page details | All facts used: homepage (H1, $49,000 per month, Live numbers), Datasets (server-built index, 18 IP rows, 185 entries on 17 IP sites), Pricing (one plan, JSON/CSV, API being built), Documentation (CODE-MPxx-TT-YYYYMMDD-NNN), Licensing (permitted / not permitted, monthly snapshot), How It Works (ground truth = site snapshot), AI Audit (#079 latest, #034) | Grok (BAA-C95-P1) = 100/100 — Win | Gemini (BAA-C95-P2) = 15/100 — Loss |
AI order follows the FINAL RESULT row of the Sheet; status is taken from that same row.
Rows are written in English; the original BigAIArena match files are in Vietnamese. Every row is kept whether the AI was right or wrong, including hallucinations and wrong citations. Sealed match records are never deleted or polished. Changes made to the site after a match are recorded with their reason in the Log.
What each Comparison asked us to fix
| Comparison | Issue found | Fix / upgrade | Status | Date fixed |
|---|---|---|---|---|
| #095 | The home page Live numbers were typed by hand (130 entries on 15 sites, 56 Sealed Comparisons on 16 sites, counted 2026-10-11) while /datasets/ already showed 185 entries on 17 sites. | The Live numbers are now counted on the server from each IP site's DATASET_MASTER sheet and from the BigAIArena results ledger. | Fixed | 2026-10-18 |
| #095 | /ai-audit/ did not show Comparison #095 (Sealed): the page kept an old saved copy, and its Checked line had stopped at 2026-10-15. | The page now reloads the AI Audit Sheet (PHP 2.2), so #095 is shown. Scores unchanged: Grok 100 · Gemini 15. | Fixed | 2026-10-23 |
| #095 | The product was named "Full API" (button "Request Full API access") while /pricing/ states that the programmatic API is still being built. | Renamed "Full Data License" on the home page, the Pricing page and the menu; the button now reads "Request a sample". Page address /pricing/ unchanged. Delivery stays JSON/CSV snapshots. | Fixed | 2026-10-23 |
| Upgrade 10/2026 | The Pricing page still listed hand-typed counts: 130 entries on 15 sites (2026-10-11) and 56 matches. | Both counts are now computed on the server, with their Checked time. | Fixed | 2026-10-23 |
| Upgrade 10/2026 | The /ai-audit/ page had no list of issues to fix and no benefits block; older matches were shown in full. | /ai-audit/ now has 3 blocks: older matches collapsed to one line each (click to open), the latest match shown in full, plus the issues table and the benefits block. | Fixed | 2026-10-23 |
| Upgrade 10/2026 | The programmatic API is not live yet. | Sell the Data License with monthly JSON/CSV snapshots only. Document the API endpoints on /documentation/ only when they are live. | Open |
For Founders · CEOs
When customers ask AI, is your business in the answer?
InjectToCite.com
- Data imprint: real information placed where AI remembers it — not lost after one search.
- Access path: AI learns the way into your site. Next time a customer asks, AI takes that path.
BigAIArena.com
- High-scoring AI: introduces your product 24/7, with sources.
- Low-scoring AI: becomes a public debate — more people hear about you than when it scores high.
- Every match shows what to fix: what your website is missing or gets wrong, so you can standardise your product.
Public results. No paid scores. No setup fee. No maintenance fee. We share only in verified growth.
Check the source pages yourself
These are the realdatasethub.com pages the AI systems were tested on:
Home · Datasets · Full Data License & Pricing · How It Works · AI Cite · Ecosystem
Want to be the auditor?
Ask any AI you use, then check its answer against the pages above. No need to trust a claim.
View Datasets