Thin Source Pool: how to find where AI engines have no answer for your industrial product — and get there before your competitors

Ask any generative AI engine to explain a specific technical problem with differential pressure transmitters installed on superheated steam lines. The answer tends to be vague, generic and, in places, technically incorrect. To an experienced process engineer, the answer would be useless, and they would know it immediately.

Now imagine that engineer is specifying equipment for a new plant, is in a hurry and has no time to call three suppliers. They ask ChatGPT, Claude, Perplexity or Gemini, get an inadequate answer and conclude that nobody who sells to them can explain the problem properly via AI. They’ll look for the information elsewhere, probably a colleague or an old technical publication.

That’s a thin source pool: the condition in which AI doesn’t have enough sources to adequately answer a specific technical query. And for anyone managing marketing at an industrial company, that scarce pool isn’t a problem. It’s an opportunity.

This article goes deeper into the thin source pool concept introduced in the article on GEO for industrial marketing, which covers the broader context of generative engine optimization in the industrial sector. Here, the focus is strictly on how to identify, map and exploit your company’s thin source pool before competitors discover it exists.

Why generative AI engines can’t answer technical questions about your product

The problem starts with how LLMs are trained. Models such as ChatGPT, Gemini and Claude learn from large volumes of text collected from the web. The distribution of what the model knows reflects the distribution of what’s published in an indexable, machine-readable format. If your product, process or niche application isn’t documented in structured public sources, the model simply has no material to build a quality answer.

FactoryBench, a benchmark developed by Pebblous in 2026 to evaluate LLM performance on industrial tasks, documented that the best available models reach less than 50% accuracy on specific industrial tasks. An arXiv paper (2410.21418) found that LLMs lack sufficient exposure to industrial technical vocabulary and to the specific regulatory requirements of sectors such as manufacturing and continuous processing.

For those selling industrial automation, process instrumentation, control systems, or even niche technologies, this has a direct implication: generative AI engines aren’t failing to answer your customers’ technical questions properly because they’re incapable of learning. They’re failing because nobody has published the answers in a format the models can use as a source.

What the model knows — and what it makes up — when it lacks enough technical sources

When an LLM doesn’t have quality sources available for a technical query, it doesn’t simply declare ignorance. It builds an answer from the sources it does have, which are usually generic marketing texts, outdated product manuals or Wikipedia articles about the general category. The result is what researchers call contextual hallucination: the answer looks plausible, uses the right vocabulary, but is technically inaccurate in the details that matter to a specifying engineer.

For highly specific technical queries, such as “how to calibrate a capacitive pressure transmitter in an application with dielectric constant variation,” AI often produces answers a senior engineer would spot as incorrect in thirty seconds. But a less experienced field engineer, or an engineering student learning the process, may not catch the error, and will blame the manufacturer who couldn’t explain it.

Why automation, instrumentation and manufacturing are missing from AI citation reports

The largest AI citation analysis reports available, such as the Surfer SEO study of 46 million citations and the ALM Corp study of 30 million analyzed sources, don’t include a specific analysis of the industrial automation and process sector. The most represented sectors are software, healthcare, financial services and e-commerce, exactly the sectors with the highest volume of digital content in English.

This empirically confirms what any industrial marketing manager already knows intuitively: our sector publishes far less web-accessible technical content than SaaS and consumer sectors, and the little it publishes often sits in closed PDFs, partner portals behind logins or internal document management systems. AI can’t use what isn’t indexed.

What a thin source pool is — and why the concept matters for industrial marketing

A thin source pool is the condition in which LLMs have few quality sources available to answer a specific technical query. The concept was introduced in the article on GEO for industrial marketing as a strategic asset: in niches where the pool is scarce, the company that publishes structured, verifiable technical content will be cited far more consistently than it would need to be in markets with many competitors publishing on the same topic.

The term isn’t standardized in the GEO and SEO literature. Equivalent concepts include “knowledge desert” (in the academic literature on LLM hallucination), “sparse coverage” (in research on how industrial niches are represented in training data) and “AEO content gap” (in Answer Engine Optimization frameworks). The difference between thin source pool and these concepts is perspective: the academic literature uses equivalent terms to describe a systemic failure of the models; thin source pool treats the same condition as a competitive advantage for whoever decides to publish first.

Pool scarcity varies by query type

Not every technical niche has the same level of pool scarcity. At the scarcest end are equipment-specific troubleshooting queries — a differential pressure alarm on transmitter model X after preventive maintenance, for example. Each model has its own fault codes, operating conditions and installation quirks that are rarely documented outside internal technical manuals. If the manufacturer hasn’t published it, the AI doesn’t know it.

Calibration under specific process conditions comes right behind: the intersection of unique parameters (humidity above 90%, process temperature above 300°C, corrosive fluid) rarely finds adequate coverage in generic sources. Applications in regulated environments such as ATEX and SIL combine technical specificity with regulatory specificity — a high-scarcity pool by nature. Comparisons between niche technologies, such as guided wave radar versus ultrasonic for specific process conditions, depend on context that generic manufacturer articles rarely provide.

At the other end are general technical concepts: “what is process instrumentation” has a dense enough pool that AI almost always has a source available. Those aren’t the topics to work on first.

The examples in this article come from the automation and instrumentation segment because that’s the sector I work in. The same logic applies to any industrial sector with high technical specificity: mining, industrial pharma, food and beverage, steel. What defines a thin source pool isn’t the segment: it’s the distance between what exists as specialized technical knowledge and what’s publicly available in a format AI can use.

Troubleshooting has the scarcest pool: why it’s the biggest industrial GEO opportunity

Troubleshooting queries are the opposite of top-of-funnel queries about general concepts. Someone asking “how to fix an overtemperature fault on an industrial variable frequency drive in an environment above 45°C” has already bought the equipment, already installed it, and has a problem they need to solve now. Intent is high, specificity is maximum, and the thin source pool is almost empty.

For industrial manufacturers, integrators and distributors, this is a rare combination: high searcher intent, very little source competition and, in most cases, internal technical documentation available to be structured and published. The field engineer who calls technical support already has the material for a troubleshooting article. All that’s missing is structuring it in a format that’s highly readable by AI.

There’s an additional factor that makes the thin source pool even more valuable for industrial troubleshooting queries: the risk level of the instructions. According to lseo.com, which analyzes GEO specifically for the manufacturing sector, in industries with safety implications AI models are designed to favor content that demonstrates E-E-A-T. For queries where an incorrect instruction can cause equipment failure, process shutdown or risk to the facility, such as maintenance procedures on pressurized systems, configuring instrumentation in ATEX hazardous areas or adjusting safety parameters in SIL systems, models prefer verifiable sources from specialized manufacturers or distributors over generic blogs. That means the thin source pool in high technical-risk queries isn’t just a visibility opportunity: it’s where identified authorship with verifiable credentials carries the greatest differentiating weight.

The double gap in Portuguese: your technical niche has fewer sources in Portuguese than in English

A detail most English-language GEO guides don’t cover: for the same technical topic, the thin source pool in Brazilian Portuguese is structurally scarcer than in English. This happens for two reasons. First, most technical documentation from international automation and instrumentation manufacturers is produced first in English and translated incompletely, or not at all, into Portuguese. Second, the few technical sources in Portuguese tend to be adaptations of English content, not original content with Brazilian market specificity (ABNT standards, Brazilian regulations, local process conditions).

For companies in the sector in Brazil, this means the entry window into the technical thin source pool in Portuguese is wider than in English for the same niches. The direct competitor who has already published detailed technical content on calibrating transmitters for specific process conditions in English probably doesn’t have that content in Portuguese. The company that publishes first in Portuguese enters an even scarcer pool.

How to identify your company’s thin source pool: a four-step method

The pillar article introduces the exercise of running ten queries in ChatGPT, Gemini, Claude and Perplexity as a starting point. Here is the method developed with objective classification criteria that make the exercise reproducible and prioritizable.

Step 1 — Build the query bank

Before opening any AI platform, build a bank of thirty to fifty queries that represent the real questions customers, specifying engineers and field technicians ask about your portfolio. The sources for this are closer than they seem:

The company’s technical support has a record, formal or informal, of the most recurring questions. Questions that reach the application engineering team by email, WhatsApp or phone are direct candidates. The sales team knows which technical objections come up most in visits. The after-sales ticket history has the troubleshooting queries.

To cover the types of thin source pool completely, structure the bank by category: at least eight equipment-specific troubleshooting queries, eight on specification by process condition, five comparing technologies, five on regulatory compliance and five on applications in specific industrial sectors. Avoid top-of-funnel queries about general concepts, which have denser pools and don’t represent the structural advantage of the industrial B2B marketing technical niche.

A practical way to build this bank is to run internal interviews with the technical team before testing any AI platform. Simple questions like “what technical question do you get at least once a week?” or “what installation mistake do you correct most often in field service?” produce highly specific queries that rarely appear in generic SEO or GEO guides. This internal exercise usually reveals that the most valuable questions for the thin source pool are the ones the technical team considers too obvious to publish, precisely because AI can’t find them anywhere either.

Step 2 — Platform audit distinguishing base LLM vs. RAG

Test each query in two different configurations. The first is the base LLM without browsing enabled: ChatGPT without active search, Gemini in offline mode, Claude without search tools. This configuration reflects the model’s static training: what it knows up to its knowledge cutoff. The second is the LLM with real-time RAG enabled: Perplexity (which always uses search), ChatGPT with browsing, Gemini with web access.

The distinction matters for timeline planning. For base LLMs, content published today starts showing up in results in six to eighteen months, when models are updated with new training data. For real-time RAG, content can appear days or weeks after indexing. If the goal is a more immediate result, prioritize queries where Perplexity is already returning inadequate answers: well-structured content shows up faster there. If the goal is building long-term authority in base LLMs, every query in the bank is relevant.

Step 3 — Classification by scarcity pattern

For each query and each platform, classify the answer you get into one of these categories:

Absent: the AI explicitly states that it doesn’t have enough information or that the query is too specific to answer
Vague: the answer is generic, uses the right vocabulary for the field but doesn’t address the query’s specificity (what “pressure transmitter calibration” is instead of how to calibrate that specific model under that condition)
Incorrect: contains verifiably wrong technical data: wrong parameter, incorrect procedure, outdated standard
Partially useful: ranges from answers that are technically correct for the general concept but don’t apply to the query’s specific context (“what pressure calibration is,” not “how to calibrate that sensor under that condition”), to answers that cover part of the specificity but leave critical dimensions open. The distinction between these two sub-cases rarely changes prioritization: both indicate that the pool exists but has room for content that goes deeper. What’s worth recording is the degree of incompleteness — 80% coverage and 20% coverage carry different weights in the next step.

Categories 1, 2 and 3 identify high-priority thin source pools. Category 4 identifies an opportunity to differentiate through specificity: a pool that exists but has room.

Step 4 — Prioritization by impact and effort

With the pattern applied, cross three dimensions to prioritize which thin source pool to exploit first. The first dimension is pool scarcity: queries classified as Absent or Incorrect have higher priority because any quality technical content will fill an empty space or correct an error. The second is query intent: troubleshooting and specification queries have high commercial or technical intent: the engineer asking that question is close to a decision. The third is the availability of internal content: if the company already has datasheets, technical bulletins or application notes on the topic, the cost of structured publication is radically lower than creating from scratch.

The ideal combination to start with is: a query classified as Absent or Incorrect + high technical intent + internal technical documentation available. That combination guarantees the greatest return for the smallest investment in content production.

One detail that prioritization makes obvious: most companies in the sector don’t need to create GEO content for every product in the portfolio at the same time. It’s enough to start with the product with the highest commercial value in the portfolio and exhaustively map the technical queries for that specific product. Once that niche is well covered, the topical authority built on that specific product starts to influence how models perceive the company’s overall technical authority. Going deep vertically on one product creates a base for horizontal expansion into adjacent products with less effort spent building credibility.

The technical documentation you already have is the most underrated thin source pool

Most GEO and AI content guides treat the work as creating new content. For manufacturers and distributors in the industrial sector, the biggest opportunity lies elsewhere: in the technical documentation that already exists but that AI can’t use because it’s in the wrong format.

Windmill Strategy documents that B2B content that lives “behind forms, in PDFs without proper metadata, or on recently launched domains” faces exclusion from LLM training data. For companies in the sector, that description covers most existing technical documentation. Datasheets in PDF without proper metadata. Installation manuals on partner portals with authentication. Field technical notes in internal repositories. Application bulletins in printed catalogs that were never digitized with semantic structure.

That’s the paradox: the company in the sector has the most specific and verifiable technical content that exists about its product, exactly the kind of material LLMs need to answer technical questions, and that content is in a format invisible to the models.

What makes a datasheet citable by AI — and what keeps 90% of them invisible

A typical industrial equipment datasheet has: precise technical specifications, operating ranges, installation conditions, performance data, certifications, materials of construction and configuration codes. It’s exactly the kind of dense information specifying engineers need, and that LLMs would need to answer specification queries.

The problem isn’t the content. It’s the format. A PDF with the complete technical specification of the most advanced pressure transmitter in the portfolio is invisible to most base LLMs, which don’t index PDFs directly. Even when indexed, a PDF without descriptive metadata, without semantic HTML structure and without schema markup gives the model a blob of text with no hierarchy, hard to extract as a structured answer to a specific query.

GEO work on existing technical documentation follows a logic of translation, not creation. Take the specification that already exists, structure it in HTML with TechArticle or Product markup in JSON-LD, make sure the H1 answers the most likely technical query (“specifications of transmitter X for application Y”), and publish on your own domain with identified technical authorship. The content is the same. The readability for AI changes completely.

This reminds me of a conversation I had with an engineering director at an industrial valve company. He was clearly uncomfortable when I suggested publishing the design specifications of one of the products on the website with full technical granularity. “Our competitors will use this,” he said. I spent about fifteen minutes trying to explain that competitors already had access to that information through the catalog and that the real problem was the engineer who had never heard of the company. I don’t know if I convinced him. But the conversation made me realize that resistance to technical publishing isn’t irrational: it’s a perception of real risk that doesn’t account for the opposite risk, which is invisibility.

This resistance is common in the sector: the technical team often sees publishing detailed specifications as a competitive risk.

“If we publish the complete technical parameters, competitors will copy them.” The concern is understandable but inverts the logic of the thin source pool. Competitors who need to copy technical specifications from your website aren’t the real risk. The real risk is that the specifying engineer deciding between two manufacturers for a BRL 800,000 project doesn’t find enough technical information about your product on AI platforms, and makes the decision based on what they found from the competitor who published more. The strategic value of technical documentation published in a structured way far outweighs the risk of a competitor copying parameters they could find in the printed catalog anyway.

Field technical note vs. blog article: why the real use case has a scarcer pool

A blog article on “how to choose a pressure transmitter” competes with dozens of generic articles from manufacturers and distributors. A field technical note on “how to fix unstable readings on a differential pressure transmitter installed on a line with process pulsation above 2 Hz” competes with essentially no other public source. The first has a reasonably dense pool. The second has an almost empty pool.

The difference is in the origin: the field technical note documents real project experience, with verifiable context. That’s the material LLMs prefer as a source for highly specific technical queries, not because the models “know” the content is more reliable, but because the most specific, verifiable content appears far less often in the training corpus, making it more distinct and more memorable to the model.

What wipes out the advantage before you publish

Identifying your company’s thin source pool is the easy part. Two common situations eliminate the advantage before you even publish the first article.

Ranking on Google doesn’t guarantee AI citation. A Writesonic study found that only 40.58% of AI citations come from Google’s Top 10. A company in the sector that ranks first in organic results for a specific technical query can still be completely invisible in AI engines for the same query. Organic SEO and GEO are kinds of visibility that need to be built separately — and in a thin source pool this matters more than in generic markets, because the company that assumes Google has it covered and postpones GEO work is giving a competitor time to occupy the pool first.

Content without verifiable authorship loses the E-E-A-T edge. Short for Experience, Expertise, Authoritativeness and Trustworthiness, it’s the central criterion LLMs use to assess the credibility of a technical source. For the industrial context, the “Experience” element is the hardest differentiator to replicate: any content agency can publish a technically correct article about variable frequency drives. Only an applications engineer or technical manager with documented projects can publish with a verifiable credential of field experience. It’s also worth remembering that text dense with technical jargon and no context in natural language can be ignored by the model even when the authorship is strong — the main technical claim needs to appear in the first sixty to one hundred and twenty characters of each section for the LLM to be able to extract it as an answer.

With few competitors publishing on the same technical niche, a single competitor with well-structured content and verifiable authorship can occupy the available space in the pool before you do.

How to exploit the thin source pool competitively: from the first article to topical authority

Identifying the thin source pool and publishing a well-structured technical article is the beginning. The sustainable advantage comes from a progressive coverage strategy that makes each new publication harder for a competitor to displace.

The conceptual starting point is understanding that a thin source pool isn’t a fixed position. When a company publishes structured technical content on calibrating pressure transmitters under specific process conditions, it doesn’t just occupy that space in the pool: it changes the pool’s composition for future queries on the topic, making it harder for a competitor to enter with equivalent content later. AI cites sources that already have a history of being cited more consistently. Whoever gets to a technical thin source pool first builds a reinforcement loop that progressively raises the barrier to entry for competitors.

The logic is vertical depth, not horizontal breadth. For a specific technical niche such as pressure transmitters for applications in environments with H2S, the right strategy isn’t publishing articles about every type of transmitter, but exhaustively covering the H2S niche: specification, installation, calibration, required certifications (ATEX, SIL), common troubleshooting, maintenance and applicable standards. Each additional article in that specific niche reinforces the company’s topical authority in that scarce pool and creates a progressive barrier to entry for competitors.

Share of model is the generative-engine analogue of share of voice: the percentage of answers in which your company is cited within a set of relevant queries. To measure it, take the thirty to fifty queries from the bank created in the identification step and query ChatGPT, Gemini, Perplexity and Claude periodically. Record in what percentage of queries your company appears as a source, mention or reference.

The 2026 Machine Relations benchmark found that 85% or more of AI citations come from editorial mentions: coverage in third-party publications, not content published on the company’s own website. For Brazilian industrial companies, this points directly to the trade press and technical publications from industry associations. A technical article published on one of these platforms with the author’s name and a verifiable credential has a significantly higher probability of generative citation than the same article published only on the company website.

There’s a practical reason for this that goes beyond domain authority. Trade publications have editorial processes that signal to models that the content has been reviewed. An article on instrumentation published in the trade press with the byline of an engineer whose title and company are identified carries E-E-A-T signals that an article on the company’s own blog, even with content of equal quality, may not have yet. Combining publication on your own site for incremental topical authority with publication in the trade press for editorial mentions covers both mechanisms of generative citation in a complementary way.

Why an editorial mention in the trade press is worth more than ten articles on your own site

The reason is structural. LLMs were trained on data collected from the web and include implicit curation that favors publications with greater established domain authority. A trade publication with fifteen years of editorial history, inbound links from industry associations and citations in academic work has domain authority that a recent company website doesn’t, regardless of the quality of the technical content published.

The most efficient strategy for building generative visibility in a technical thin source pool combines publishing on your own site (for incremental topical authority and real-time RAG) with publishing in the trade press (for editorial mentions with established domain authority). Both contribute to the goal, but through different paths.

What to do when a competitor publishes in the same niche: depth as a barrier to entry

When a competitor identifies the same thin source pool and starts publishing, the most effective response isn’t trying to publish more. It’s publishing with greater technical depth. An article on calibrating differential pressure transmitters can be displaced by a competitor who publishes a similar article. An article on calibration on a superheated steam line with process pulsation above 2 Hz, with a step-by-step procedure, a reference to the ISA-RP3.2 standard and testimony from a field engineer, isn’t easily displaced, because the cost of producing equivalent content is high for a competitor and the pool already has a quality source.

What managing industrial marketing reveals about the thin source pool

I run marketing for an industrial technology company in the automation and instrumentation segment. One of the most striking features of this work is the distance between what the technical team knows and what’s publicly accessible about the company’s products and processes.

That distance is the thin source pool in real life.

The technical knowledge nobody published

There’s a specific situation that comes to mind when I talk about this. A technician on our applications team spent months solving a recurring reading drift problem in a specific type of sensor in an environment with cyclical temperature variation. He developed his own procedure, tested it in the field, and it worked. That procedure is in a text file on his computer. It isn’t anywhere AI could find it. It probably never will be, unless someone decides to turn it into publishable content. Multiply that by every technician at every industrial company in Brazil.

At any company with an established technical portfolio, accumulated knowledge is distributed unevenly: a lot inside the organization, little in a public, indexable format. The application engineering team knows in detail the process conditions that cause unstable measurements in certain sensors. Technical support knows which configurations fix the most common field failures. The commissioning team has records of how to adapt installations to conditions the manual doesn’t cover.

That knowledge is rarely in a format AI engines can use. It’s in emails, in team memory, in internal technical reports, in maintenance procedures that never left the corporate server. The consequence is that an engineer at another company who asks an AI about a problem your specialists would solve in minutes gets a generic or incorrect answer, not because the answer doesn’t exist, but because it was never published where models could find it.

Why the internal specialist sees the thin source pool better than any external tool

GEO audit tools show where AI has gaps by querying the platforms from the outside. Those managing marketing at an industrial company already have access to a more precise diagnosis: they know which questions frequently reach the technical team, which objections come up in sales visits, which problems after-sales solves repeatedly. Each of these recurring situations is a thin source pool confirmed by the real behavior of customers and users, not by an audit.

The practical difference is that the internal specialist doesn’t need to discover whether the thin source pool exists. They already know it exists because they live the gap every day. The query audit exercise serves to confirm and prioritize, not to reveal what people inside the sector already know intuitively.

What changes when this perspective informs the content strategy

The most significant change in how I think about content after working with this concept is the selection criterion. The question stopped being “what should we publish to rank on Google” and became “on which of our customers’ technical questions can AI not give an adequate answer, and what can we publish that would fill that gap with verifiable content.”

That criterion produces a radically different list of topics. Generic content about the product category competes with everything that already exists. Content about how to solve a specific problem, under a specific process condition, with the specific type of equipment the company supplies, competes with almost nothing.

Where to start

If you haven’t read the article on GEO for industrial marketing yet, it’s worth starting there: it provides the full picture of how generative engine optimization works in the industrial sector and puts into context why the thin source pool matters within a broader strategy.

The most useful conversation to start with isn’t with a GEO tool. It’s with whoever runs technical support — ask which three questions come in every week. Troubleshooting queries for the main portfolio have the scarcest pool, the highest technical intent and, at most companies, internal documentation that already exists and just needs structuring. While that survey is under way, take the datasheet or manual for the most consulted product and work on its HTML structure with schema markup: the content already exists, it’s verifiable and it’s exactly what specifying engineers need when they ask an AI. The biggest return per hour invested rarely lies in creating new content — it lies in making what already exists readable.

Publishing in the trade press amplifies what is being built internally. An article signed by an engineer with an identified title and company contributes editorial mentions and domain authority that your own site takes a long time to build on its own.

The thin source pool won’t last forever. As more companies in the sector discover GEO and start publishing structured technical content, the pool will become denser. The Brazilian industrial sector is systematically underrepresented in the training data of the language models available today — that’s documentable, not speculative. The entry window into the technical thin source pool in Portuguese is wider than in English for the same topics, and the company that establishes topical authority in that space before the market gets denser will build a citation advantage that becomes progressively harder to reverse. You don’t need to do everything at once. Starting with the right product, the right query and the right format is enough.