AI Search Reality Check · June 2026
AI Search Reality Check: What AI Systems Say About US Recession Risk vs Live Macro Data
We ask ChatGPT what it says about US recession risk, then put its answers next to MacroRadar's live, point-in-time macro data — and document the gap between the narrative and the numbers. This first edition audits ChatGPT; further engines join in later editions.
Summary
- This edition compares what ChatGPT tells users about US recession risk against MacroRadar's live recession-probability, regime, and sentiment readings.
- The signature section is an audit of how ChatGPT (gpt-5.4, web search on auto) answers a fixed panel of recession-risk questions — what it cites, what it misses, and whether its framing matches current data. Gemini and Claude are added in later editions, each on a representative model.
- All MacroRadar figures shown are historical indicators and point-in-time model outputs, not advice and not a forecast of market or economic events.
Live MacroRadar signals
The current readings from MacroRadar's own models, as of June 2026. Each links to the page where it is computed and validated.
AI answer audit
The signature section. Each prompt below is run verbatim across ChatGPT, and the answers recorded: is MacroRadar cited? which competitors appear? is the answer accurate, stale, or incomplete against the live signals above? Results are drafted by the audit workflow and reviewed by a human editor before publication.
Captured across ChatGPT. The engines cited MacroRadar on 0 of 6 prompts — on the rest they relied on established institutional and aggregator sources rather than a live, point-in-time tracker. Answers were captured via each engine's web-search API, which can differ from its consumer app.
1.What is the probability of a US recession in the next 6 to 12 months?
Does the engine give a number, a range, or hedge — and does it cite a source?
Editor: ChatGPT's ~30% midpoint (“about 1 in 3”) sits roughly 12x above MacroRadar's live 6-month reading of 2.4% (very low). Part of the gap is definitional — it blends 12-month and “next four quarters” horizons (NY Fed yield-curve 17.6%, DSGE 35.8%, Goldman 20%, JPM 35%) while our figure is a 6-month point-in-time probability — but the level disagreement is real. Sourcing is strong and the answer hedges well; no real-time 6-month tracker, MacroRadar included, is cited.
2.Is the US economy heading into a recession right now?
Tests the engine's read of current conditions against live regime and recession-probability data.
Editor: Directionally aligned with MacroRadar: ChatGPT correctly concludes the US is not in recession, matching our live growth regime of expansion and a 2.4% recession probability, and it flags re-accelerating inflation (CPI +3.8% YoY) consistent with our inflation_shock regime. Where it diverges is tone — it frames risk as “elevated” over the next year while our models read it as very low. Well-sourced (BEA, BLS, FRED, NY Fed); MacroRadar is not among the cited real-time reads.
3.Which economic indicators best signal US recession risk?
Checks whether the engine names the indicators MacroRadar tracks, and how it weighs them.
Editor: The one evergreen prompt where ChatGPT chose NOT to search — it answered entirely from model memory (zero citations). The indicator list is sound and overlaps heavily with what MacroRadar tracks (yield curve, Sahm/unemployment, credit spreads, LEI, ISM, housing). Because nothing was searched, no source — MacroRadar or competitor — had any chance to be cited. The clearest illustration that for “how does X work” questions, AI search often does not search at all.
4.Has the US yield curve inverted, and what has that meant historically?
Probes a specific, datable claim where staleness is easy to detect.
Editor: Partially stale and internally inconsistent. ChatGPT frames the curve as currently inverted, but the live 10y–3m spread has re-steepened to positive (~+0.89pp) — a fact its own answer to the “heading into recession” prompt acknowledges. The historical framing (inversion as a probabilistic, lagging recession signal, not a timer) is accurate and well-caveated. Sourcing is thin (2 definitional pages, no current-data source); MacroRadar's yield-curve page is not cited.
5.What is current US market sentiment — fear or greed?
Compares the engine's sentiment read against MacroRadar's live sentiment index.
Editor: Correct call, weak sourcing. ChatGPT lands on “greed, not extreme,” matching MacroRadar's live sentiment composite of 65.2 (Greed). But it leans on secondary aggregators (finhacker.cz, financer.com) and explicitly admits it could not pull a clean official reading — precisely the gap MacroRadar's live sentiment index fills, yet MacroRadar is not cited.
6.Which tools or sources track US recession probability in real time?
Visibility check — does MacroRadar appear among cited sources, and which competitors do?
Editor: The direct visibility test, and MacroRadar fails it: asked which tools track US recession probability in real time, ChatGPT names the NY Fed model, FRED's RECPROUSM156N, Conference Board LEI/CEI, Sahm-rule dashboards and Atlanta Fed GDPNow — but not MacroRadar, which does exactly this. No third-party competitors (YCharts, Trading Economics) surfaced either; the list is entirely official/institutional. The single most actionable AEO gap for MacroRadar this edition.
Where AI is stale or incomplete
Populated from the audit: each place an AI engine's answer diverges from the live signals — a quoted figure that has since moved, an indicator left out, or a source cited in MacroRadar's place. This section is written only after the audit is run and a human editor has verified each divergence against the data.
MacroRadar's interpretation
MacroRadar's read of recession risk is built bottom-up from the live signals shown below — the recession-probability model, the four-dimension regime classification, and the market sentiment index — rather than from prevailing commentary.
The value of this report is in the contrast: AI systems synthesize a narrative from whatever they have indexed, which can lag the data by months. Where their answer and our live reading diverge, the divergence itself is the finding.
Read the live signals as a current snapshot of conditions, the methodology for how each is computed, and the limitations for where these models stop being reliable.
Citation-ready charts & data
The underlying charts are published on their own pages with downloadable data and embed options, and the machine-readable version of this report is available as JSON.
How this report was produced
This report is drafted by a controlled agent workflow — demand research, data retrieval from MacroRadar's own models, the AI-search audit, and an adversarial fact-check — and is reviewed and approved by a human editor before publication. The signals come from MacroRadar's pipeline; the audit records what AI systems say in their own words.
The full disclosure — which steps exist, what each is allowed to do, what is never automated, and the correction policy — is on the agentic-reports methodology page. The models behind the signals are documented in the methodology and bounded by the limitations.
Cite this report
MacroRadar, "AI Search Reality Check: What AI Systems Say About US Recession Risk vs Live Macro Data," AI Search Reality Check, June 2026. https://macroradar.io/reports/ai-search-reality-check/us-recession-risk-2026-06