If you've already checked out our reviews you've seen their colorful graphs and quick-read insights into multiple aspects of a car. While our reviews will increasingly include proprietary and unique test results for ride quality, automated lane-centering, and adaptive cruise control - plus targeted commentaries by our authors - the primary text that surrounds these reviews is the consensus opinion of the worlds car reviewers. While I might happen to like a vehicle review I write... I know it's also just my singular opinion. Maybe you'd disagree with some of it. But here, in every car review you read, is the power of group think - harnessing the merged judgements that several reviewers - and throwing out the oddball opinions. However, this goes way (way) beyond a simple AI chatbot search. It's a complex and meticulous analytical process that brings considerable added value to offer you the best information about car out there. So, you might be wondering ....
How do we generate these?
Step 1: Vehicle Prompt Input
A vehicle's year, make, model, and trim seeds every downstream prompt and determines the vehicle lookup targets.
Step 2: Parallel Model Research
Three independent models, GPT-5.2, Gemini 3.1 Pro, and Grok 4.1, each receive the same 20-category research prompt and generate long-form reports in parallel. The prompt is auto-batched so each model answers the full category set sequentially.
Step 3: Web-Grounded Fact Verification
After the three initial drafts are complete, Grok, which has live web search, reviews each draft and lists factual errors and unverifiable claims. The other models do not have live web access, so their self-checks alone would not catch training-data errors.
Step 4: Correction Application
Each model receives its own draft plus the Grok-generated correction list. It rewrites only the objectively checkable facts while preserving subjective judgments, tone, and detail-level observations. The corrected draft replaces the original for downstream use.
Step 5: Democratic Merge
Gemini combines the three corrected drafts into a unified 20-section document. A statement is kept only if at least two of the three models agree on it. Detail-level praise and criticism are preserved, but overall summary verdicts are stripped out.
Step 6: Sources Consulted
The system extracts the names of publications, websites, and sources referenced by the models during research. This creates a transparent list of where the consensus information originated.
Step 7: Final Fact-Check Review
A final search-backed fact-checker, currently GPT-5.4, reviews the merged text and flags any remaining factual errors. The system also uses the earlier Grok verification evidence as supporting context for this final pass.
Step 8: Longform Synthesis Generation
The corrected merged text is expanded into a longform synthesis of roughly 350 words per category. This version is richer in detail and is included as a downloadable export.
Step 9: Shortform Synthesis Generation
A shortened version of roughly 140 words per category is produced for display on the website and for the sentiment analysis step. It is distilled from the same corrected merged text.
Step 10: Sentiment Analysis
Four models, GPT-5.2, Gemini 3.1 Pro, Grok 4.1, and Claude 3.5 Sonnet, each score the shortform text across eight dimensions over two rounds. The scores are validated, averaged, and rendered as charts and tables.