The "Manufacturing" filter in your B2B database returns 300,000+ companies in the US. Every one of them is classified under NAICS sector 31-33. That bucket covers a pharmaceutical plant making injectable drugs, a steel mill rolling structural beams, a bakery producing 3 million slices of bread per day, a semiconductor fab patterning wafers at the 3-nanometer node, a plastic injection molder stamping out food containers, and a metal-stamping job shop with 14 employees.
None of those plants share a buyer. None share a budget cycle. None share a buying process. None share a relevant certification regime. None share a decision-maker title. And yet, in every standard B2B database — ZoomInfo, Apollo, D&B Hoovers, Cognism — they are all one dropdown option: "Manufacturing."
If you are an industrial sales rep trying to build a prospect list, or a B2B marketer trying to build a campaign audience, the "Manufacturing" filter is the single most time-consuming false positive in your workflow. It returns everything. It discriminates nothing. You spend hours in Excel trying to re-filter a generic "manufacturing" export into a list that reflects the actual plants you can sell to.
This post explains exactly why the filter fails — it is a taxonomy-granularity problem, not a search-UI problem — walks through the specific absurdity with named plants across four different industries all stamped "Manufacturing," and shows what facility-level AI classification looks like instead. It is aimed primarily at industrial sales reps building prospect lists, and secondarily at B2B industrial marketers whose ABM lists dissolve the moment they try to segment beyond the top-level industry code.
The four-plant absurdity
Here are four real US manufacturing plants. Every one of them is classified under NAICS sector 31-33, Manufacturing. In most B2B databases, every one of them appears when you filter on "Manufacturing."
Plant 1: A Pfizer pharmaceutical manufacturing facility. Makes injectable drug products. Operates under FDA cGMP (current Good Manufacturing Practice) regulations. Runs ISO 7 and ISO 5 clean-room environments. Employs hundreds of people including pharmacists, quality-control microbiologists, process engineers, and sterile-fill operators. Buys from vendors in categories like: pharmaceutical-grade isolators, single-use bioprocess bags, HEPA filtration, cleanroom garments, pharmaceutical-grade solvents, validation services. The plant manager reports to a VP of Manufacturing at Pfizer. Capital budget is eight figures per year; MRO is seven figures per year.
Plant 2: A Nucor steel mill. Electric-arc-furnace steel production, producing hot-rolled coil and structural beams. Employs hundreds of people — melt-shop operators, rolling-mill operators, metallurgists, maintenance mechanics. Buys from vendors in categories like: refractories, scrap-metal feedstock, alloying additions, mill rolls, crane and hoist service, electrical contractors for high-voltage work, hydraulic systems. The plant manager reports to a regional VP at Nucor. Capital budget is substantial; scrap and alloys are the dominant input cost.
Plant 3: A Flowers Foods bakery. Produces packaged bread products at industrial scale. Employs mid-hundreds of people including bakers, line operators, sanitation, and maintenance. Buys from vendors in categories like: industrial ovens, dough-mixing equipment, conveyors, food-grade lubricants, sanitation chemicals, corrugated packaging, ingredients (flour, yeast, shortening) in bulk quantities. Operates under FDA food-safety regulations and SQF or similar food-safety certifications. Plant manager reports to an Operations VP.
Plant 4: A TSMC-Arizona or Intel-Arizona semiconductor fab. Patterns silicon wafers at leading-edge process nodes. Employs thousands of people including process engineers, equipment engineers, fab technicians, metrology engineers. Buys from vendors in categories like: lithography steppers, deposition and etch tools, ultra-pure gases, photoresist chemicals, wafer carriers, clean-room subfab utilities. Operates under SEMI industry standards and customer-specific qualification requirements. Capital budget is in the billions; a single tool purchase can be $200M.
These four plants are in the same NAICS-33 bucket. A B2B database filtered on "Manufacturing" returns all four. A "manufacturing" sales list treats them as interchangeable. They are not. The vendor selling HEPA filtration to Pfizer has no product relevance to Nucor. The vendor selling bakery ingredients has no product relevance to TSMC. The vendor selling industrial lubricants has different products for each of the four — food-grade for Flowers, high-temperature for Nucor, pharmaceutical-grade for Pfizer, and ultra-pure low-particulate for TSMC.
The "Manufacturing" filter is not a filter. It is a label that conflates what should be 50 distinct market segments.
The taxonomy problem, stated clearly
The US government maintains NAICS — the North American Industry Classification System — and it has several levels of granularity:
- 2-digit sectors (like 31-33 Manufacturing): 20 total
- 3-digit subsectors (like 311 Food Manufacturing, 325 Chemical Manufacturing): ~100 total
- 4-digit industry groups (like 3254 Pharmaceutical and Medicine Manufacturing): ~300 total
- 5-digit industries (like 32541 Pharmaceutical and Medicine Manufacturing): ~700 total
- 6-digit industries (like 325412 Pharmaceutical Preparation Manufacturing): ~1,000 total
Even at the 6-digit level — the most granular NAICS provides — the resolution is coarse. NAICS 325412 covers pharmaceutical preparation manufacturing broadly. It does not distinguish between:
- Sterile injectable drug manufacturing (biologics, vaccines, oncology therapeutics)
- Solid-dose oral manufacturing (tablets, capsules)
- Topical drug manufacturing (ointments, creams)
- Controlled-substance manufacturing (schedule II–V, with different regulatory requirements)
A vendor selling sterile-fill isolators targets the first. A vendor selling tablet coaters targets the second. They do not share ICP. And yet both segments are lumped under NAICS 325412.
The problem is structural. NAICS was designed by the US government for economic-statistics collection, not for sales targeting. The taxonomy was fixed in broad strokes decades ago, updated every five years, and has structural resistance to getting more granular because getting more granular would break continuity in the government's time-series statistics. The system is doing what it was built to do — it was not built to power B2B sales lists.
The B2B databases that use NAICS as their industry filter inherit this limitation. They add a "Manufacturing" option to their dropdown because that is what NAICS provides. They sometimes let you drill to the 6-digit level. And at the 6-digit level, you still get a bucket of incompatible plants.
Why sub-industry filters still do not solve it
The most common workaround is to drill NAICS to the 6-digit level and hope for precision. It does not arrive.
Consider NAICS 311812, "Commercial Bakeries." That code covers:
- Industrial wholesale bakeries producing 3M+ units/day of packaged bread (Flowers Foods, Bimbo Bakeries, Grupo Bimbo subsidiaries) — hundreds of thousands to millions of pounds per day
- Mid-sized regional bakeries producing 500K units/day of buns and rolls for fast-food chains
- Artisan bread bakeries producing 50K units/day of retail and foodservice product
- Specialty cake and dessert bakeries producing 10K cakes/day
- Frozen dough manufacturers producing frozen raw dough for downstream bakeries
A vendor selling industrial-scale continuous ovens targets the first. A vendor selling retail packaging equipment targets the third. A vendor selling freezer tunnels targets the fifth. None of these vendors want a single 6-digit NAICS filter because that filter returns all five segments.
The pattern repeats across every 6-digit code in manufacturing. NAICS 326199, "Plastics Product Manufacturing (All Other)," covers plastic bottle manufacturers, plastic component molders, plastic film converters, foam-product makers, and acrylic-sheet producers — none of which share a buying committee or a vendor list.
NAICS 332710, "Machine Shops," covers precision machine shops serving aerospace tier-2, general contract machine shops, stamping operations, and specialty machining for medical devices. Four completely different ICPs, same NAICS code.
NAICS 33344 covers semiconductor machinery manufacturing broadly. It treats a clean-room-focused wafer-fab tool builder as similar to a printed-circuit-board equipment maker.
Drilling NAICS does not solve the granularity problem. It just moves the same problem down a level.
What plant-level classification needs to look like
The right taxonomy for industrial sales has two properties that NAICS does not:
Property 1: Classification at the plant, not the company. A parent like Berry Global operates plants that span plastics bottle manufacturing, flexible film production, nonwoven fabric production, and engineered packaging. NAICS assigns Berry a code at the company level — typically the primary revenue-generating sector — which masks the diversity of what each plant actually makes. A plant-level classification assigns the Lawrence, KS flexible film plant a "flexible film" label, the Evansville plastic bottle plant a "plastic bottle" label, and the Monticello nonwoven plant a "nonwoven fabric" label. Each plant gets the classification that matches what that plant actually produces.
Property 2: Classification based on products made, not on a pre-built taxonomy. The taxonomy expands as the data expands. A plant that makes something NAICS never imagined — carbon-fiber pressure vessels for hydrogen storage, lithium-iron-phosphate battery packs for industrial vehicles, vertical-farm growth pods — gets classified by what it makes, using a taxonomy that grows to include those products. NAICS cannot do this. A fixed 6-digit system cannot represent 35,000 distinct product-industry combinations.
Facilities Finder is built around both properties. Our AI ingests billions of public signals — satellite imagery, map providers, company websites, EPA filings, permit records, trade publications — and extracts what actually matters at each facility: products, capabilities, employees, certifications. The output is 35,000+ AI-generated industry classifications and 7 million+ products indexed per facility, all drawn from what plants actually produce.
That is the structural fix. The four-plant absurdity from the top of this post — Pfizer, Nucor, Flowers, TSMC — resolves when each plant gets a real plant-level classification:
- The Pfizer plant classifies as "sterile injectable pharmaceutical manufacturing" with products including "injectable drug products" and "sterile-fill operations," plus relevant certifications (FDA cGMP).
- The Nucor plant classifies as "electric-arc-furnace steel production" with products including "hot-rolled coil" and "structural steel beams," plus relevant certifications (ISO 9001, AISC).
- The Flowers plant classifies as "industrial wholesale bakery — packaged bread" with products including "packaged bread products" and "bun and roll production," plus relevant certifications (SQF, FDA-registered).
- The TSMC-Arizona plant classifies as "semiconductor wafer fabrication — leading-edge logic" with products and capabilities at the process-node level.
Each plant now sits in a different searchable bucket. A vendor selling to pharmaceutical sterile-fill operations filters to plants in the first bucket. A vendor selling to steel melt shops filters to the second. The four-plant absurdity disappears because the plants are no longer in the same bucket.
Semantic search replaces the filter tree
The second move is a UX one. Once the classifications are fine-grained, the filter dropdown stops working. No human can navigate a 35,000-option dropdown.
The right interface is natural-language search. The user types what they are looking for — "pharmaceutical plants with sterile-fill operations," "electric-arc-furnace steel mills in the Midwest," "industrial bakeries over 300 employees" — and the AI extracts the intent, embeds the query, searches the full database semantically (not by keyword), and ranks results by how well each plant actually matches the query.
This is what "Type what you're looking for. No filters to learn" means in practice. The Facilities Finder search does not make you guess the right NAICS code. You describe the plant in plain English. The AI handles the translation to the plant-level classifications it has already assigned.
The difference matters most on niche queries. "Injectable drug manufacturing with 400+ employees within 150 miles of Indianapolis" is a specific ICP. It cannot be built from NAICS at all — 325412 returns too many non-injectable pharma plants, and the employee filter cannot be combined with the sub-product granularity because NAICS has no sub-product layer. A semantic search on the same query returns the handful of plants that match all three criteria.
What this means for prospect lists
The practical change for a sales rep or a marketer is how the prospect list gets built. Under the NAICS-filter model, the workflow is:
- Select "Manufacturing" (sector 31-33) or drill to the best-available 6-digit code
- Export to CSV — thousands of rows, 60–80% of which are not your ICP
- Spend hours in Excel trying to filter on company-name keywords, website text, or LinkedIn descriptions to eliminate the false positives
- Accept the remaining 20–40% false positive rate because further filtering is not worth the time
- Run outbound against a list that includes irrelevant plants; write off the waste as "low-touch discovery"
Under a plant-level AI-classification model with semantic search, the workflow is:
- Describe the ICP in natural language
- Review the returned plants — precision is high because the classification is granular and the search ranks by match quality
- Apply secondary filters at the plant level — employee range, geography, certification, parent context
- Export or push directly into the built-in CRM for outbound
The first workflow takes hours and produces a polluted list. The second workflow takes minutes and produces a tight list. The difference is the underlying taxonomy granularity, not the interface polish.
FAQ
Isn't NAICS the "standard" for B2B data?
It is the standard for US government economic statistics. It is the standard for SBA small-business classification. It is the standard for many state procurement systems. It is not an effective standard for B2B sales targeting because it was not designed for that purpose. Using NAICS for sales targeting is analogous to using ZIP codes for territory planning — the ZIP is a postal-routing label, not a sales-geography label, and it will let you down at the boundaries.
Can't the B2B databases just add sub-NAICS classifications?
Technically, yes. In practice, the underlying data ingest pipelines — corporate registrations, web crawls of "About" pages, LinkedIn company descriptions — do not produce the granularity needed. A company's LinkedIn page might say "Manufacturing company in Ohio." It does not say "electric-arc-furnace steel mill producing hot-rolled structural beams." The classification can only get as granular as the source data, and the source data most databases use is not granular at the plant level.
What about SIC codes? Aren't those more granular?
SIC (Standard Industrial Classification) has similar resolution issues. It was the predecessor to NAICS and was retired for US statistical purposes in 1997. Some databases still carry SIC codes alongside NAICS, but the underlying taxonomy granularity is similar — four digits of SIC ≈ 6 digits of NAICS. Neither reaches the product-level granularity industrial sales requires.
How does Facilities Finder handle NAICS at all then?
We do not use NAICS as filter input. Facilities Finder's AI classifies each plant by what it actually makes, from public signals — 35,000+ AI-generated industry classifications and 7 million+ products. NAICS as subject matter is fine — we can tell you what a NAICS code means when a user is researching it — but we do not make the user type a NAICS code to filter our data. The user types natural language; the AI handles the translation.
What about for segmentation inside a marketing platform?
If your marketing-automation platform or ABM platform exposes an "industry" field that is NAICS-derived, the field will have the same granularity problem. The right fix is to sync plant-level classification data into the marketing platform as a custom field, so segments can be built on the fine-grained classification rather than the NAICS 2-digit bucket. A plant-level database makes this integration straightforward because the classification is already at the right level.
Start the ICP rebuild with facility-level classification
If your prospect list is currently polluted with 40% irrelevant plants, the cause is almost certainly that the industry filter you used was NAICS-grade. "Manufacturing" is not a segment. NAICS 325412 is not a segment. NAICS 332710 is not a segment. Real segments are at the product, facility-type, and capability level — and those segments do not fit inside the NAICS taxonomy because NAICS was not designed to hold them.
Facilities Finder indexes every US industrial facility with an AI-generated classification at the plant level. 35,000+ industry classifications and 7 million+ products — all derived from what plants actually produce, not from a pre-built government taxonomy. Type what you're looking for in plain English — "sterile-fill pharmaceutical plants" or "electric-arc-furnace steel mills" or "industrial bakeries over 300 employees" — and the AI extracts intent, searches across 600,000+ plants semantically, and ranks by match quality. Every plant record includes the plant's own industry classification, its own product list, its own employee count, and its own decision-maker contacts at that specific site.
25 million+ decision-maker contacts, all keyed to the physical facility where they work.
Rebuild your ICP with plant-level classifications — get access at Facilities Finder.
See also: Why Your CRM Shows 1 Record for a Company That Runs 87 Plants · How to Run ABM for Industrial: Target Plants, Not Logos · What Is a Tier 2 Supplier? A Plain-English Guide for Sales Teams