PulsCar analysis of NHTSA public data, September 2026 Article: https://www.pulscar.io/blog/car-smells-like-gas License: CC BY 4.0 (attribute "PulsCar analysis of NHTSA public data, September 2026") SOURCE FILES - NHTSA vehicle owner complaints flat files (COMPLAINTS_RECEIVED_2015-2019, 2020-2024, 2025-2026), field dictionary https://static.nhtsa.gov/odi/ffdd/cmpl/CMPL.txt. Loaded into SQLite; 777,040 unique reports received 1 Jan 2015 - 8 Sep 2026, de-duplicated by ODI number. - NHTSA recalls flat file FLAT_RCL_POST_2010 (https://static.nhtsa.gov/odi/ffdd/rcl/RCL.txt): one row per make/model/year per campaign; campaigns counted DISTINCT. - NHTSA manufacturer communications (TSB) flat files TSBS_RECEIVED_2015-2019, 2020-2024, 2025-2026 (https://static.nhtsa.gov/odi/ffdd/tsbs/TSBS.txt): one row per make/model/year; documents counted by DISTINCT NHTSA id. - NHTSA investigations flat file FLAT_INV (https://static.nhtsa.gov/odi/ffdd/inv/INV.txt). Data snapshot built 14 September 2026; queries run 17 September 2026. COMPLAINT QUERY (gas_smell_complaints_*.csv) Full-text (FTS5) match on the owner's narrative for any of these phrases: "smell of gas", "gas smell", "smell gas", "smelled gas", "smells like gas", "smell of gasoline", "gasoline smell", "smell gasoline", "smelled gasoline", "smells like gasoline", "smell of fuel", "fuel smell", "fuel odor", "gasoline odor", "gas odor", "odor of fuel", "odor of gasoline", "smell fuel", "smelled fuel", "fuel fumes", "gas fumes", "gasoline fumes", "smelling gas", "smelling fuel", "smelling gasoline", "raw fuel odor", "raw fuel smell", "smell of raw fuel", "smell raw fuel". Result: 7,747 complaints (1.0% of 777,040). No filter on vehicle type: the file includes a small number of motorcycles and three-wheelers (Can-Am Spyder, BMW R 1200 GS appeared in our reading samples). No filter on diesel; one diesel Audi Q7 appeared in a reading sample. QUERY TIGHTENING AND FALSE-POSITIVE READING First draft of the query also included "strong gas", "raw gas" and "raw fuel". A random sample of 30 matches (seed 1) contained 2 clear false positives caused by those terms ("hit the gas", "add gas") and 1 unclear truncated report; the three terms were dropped. Second random sample of 30 (seed 7) with the final query: 0 clear false positives; 3 reports were about a gasoline smell in the engine oil (fuel dilution), which we count as a genuine fuel-smell complaint but flag separately; 1 report was truncated and unclear. Third random sample of 100 (seed 11) with the final query, read in full by a person and categorized: - Not a gasoline smell from the fuel system: 4 (1 diesel; 3 exhaust/sulfur smell described as "gas fumes"). Estimated false-positive rate across the 130 texts read: about 5%. - Not a passenger car: 3 (two Can-Am Spyder three-wheelers, one BMW motorcycle). - Owner or shop named a cause (55 of 100): fuel pump / pump module / tank flange 14; fuel tank body, seam or seal 10; fuel lines or hoses (incl. corrosion and rodent damage) 12; gasoline in the engine oil 4; injector, rail or under-intake leak 3; exhaust or other non-fuel cause (O2 sensor threads, head gasket, valve cover, intake backfire) 7; shop could not find or duplicate 4; EVAP or cap 1. - No cause stated: 45. Reading note: the complaint file over-represents serious, unresolved and recall-related cases. Owners rarely file a federal complaint about a loose cap. Counts are therefore NOT the frequency of causes in the driving population and are not failure rates. KEYWORD GROUPS (gas_smell_complaints_keywords.csv) Regular expressions applied to the lower-cased narrative of the 7,747 matches; a complaint can match several groups. The CSV lists each regex and its count. MILEAGE Failure mileage as reported by the owner; excluded blank, 0 and >= 400,000 (5,179 of 7,747 usable). Median 80,000; quartiles 38,000 and 120,000. Baseline for all 777,040 complaints with the same exclusions: median 70,000; quartiles 33,000 and 110,000. CRASH / FIRE / TOW SHARES Complaint flags as filed by the owner. Gas-smell set: fire 180 (2.3%), towed 224 (2.9%), crash 33 (0.4%), injuries reported in 99, deaths in 1. All complaints: fire 14,121 (1.8%), towed 39,739 (5.1%), crash 33,940 (4.4%). MAKE AND MODEL COUNTS (gas_smell_complaints_by_make.csv, _by_model.csv, _by_model_year.csv) Make and model as entered in the file; "MERCEDES-BENZ" and "MERCEDES BENZ" are two spellings of one make and are combined in the article (939 complaints). Model counts are the top 60. Share of the make's complaints in the whole file is given as a baseline; it is not a normalization by vehicles on the road. RECALLS (fuel_leak_recalls_2015_2026.csv) Campaigns with component beginning "FUEL SYSTEM, GASOLINE", report date on or after 1 Jan 2015, and the word "leak" in the defect or consequence text. 264 distinct campaigns, 9,226,822 potentially affected vehicles (max potaff per campaign). remedy_type is a keyword classification of the remedy text (software_or_update / inspect_and_replace_if_needed / replace / other); it was not hand-checked for every row. BULLETINS (fuel_odor_bulletins.csv) Manufacturer communications whose NHTSA summary matches "fuel odor", "gasoline odor", "fuel smell", "gas smell", "smell of fuel", "odor of fuel", "raw fuel odor", "fuel fumes", "gas odor" or "gasoline smell": 166 documents (153 distinct bulletin numbers), 2015 - Aug 2026. 6 are engine-computer software flashes that list a fuel smell among many symptoms. Only documents whose PDFs we opened are quoted in the article; the rest are cited from NHTSA's summary text. LIMITATIONS Self-reported narratives; duplicate filings by the same owner are possible; keyword matching misses misspellings ("smel", "oder") and counts a smell mentioned in passing. Cause categories come from what the owner wrote, not from an inspection. Nothing here measures how often a part fails. ------------------------------------------------------------------------------ CASE SENSITIVITY — SECTION ADDED 19 SEPTEMBER 2026 ------------------------------------------------------------------------------ NHTSA stores complaint narratives in ALL CAPS for records filed through roughly 2020 and in mixed case for records filed from 2021 on. A case-SENSITIVE search of the narrative field loses about 92% of the complaints filed after 2020, so this is the single largest way an analysis of this database can go wrong while still looking finished. Every query behind this file is case-insensitive. SQLite's LIKE and its FTS5 full-text index both ignore case by default, and no query in this work uses GLOB or any case-sensitive comparison. This was not taken on trust. On 19 September 2026 every published folder was re-checked against the same database, comparing the figures here with what a case-sensitive query would have returned. The result: the complaint set published here (7,747) is LARGER than the absolute ceiling a case-sensitive search of the same terms could have returned (729 narratives). A case-sensitive query cannot produce a bigger set than itself, so the bug is ruled out for this folder by arithmetic, not by assurance.