Ask a generative engine a narrow operational question — who handles sour-gas treating on a compressed turnaround, who has run managed-pressure drilling in a specific basin, who recycles produced water on pads with almost no surface area — and the answer often names a company with forty people instead of the company with four thousand. The larger firm has more revenue, more projects, more press coverage, and a much older domain. It still loses the answer.
That looks like a bug in the model. It is not. It is the predictable result of how generative engines assemble an answer, meeting how large energy companies produce content. The big company loses the narrow question because its publishing process removes the specifics a model needs to quote, and the specialist wins because it never had a process capable of removing them.
Why does a smaller competitor show up in the AI answer instead of us?
Because the engine is not choosing a company. It is choosing a passage.
On a broad prompt — the largest oilfield service companies, the major midstream operators — scale does win, because prominence is the whole signal and the model has seen the big names described that way thousands of times. But buyers close to awarding work do not ask broad questions. They ask constrained ones, with conditions attached: this fluid, this basin, this pressure regime, this schedule, this regulatory wrinkle. A constrained question can only be answered from text that carries those same constraints. If nothing you have published states the conditions you actually work under, there is no passage to retrieve, and your size does not enter the calculation.
You are not being outranked. You are absent from the shortlist the engine draws from, on the questions closest to a purchase.
What is a generative engine actually looking for?
The mechanics have been studied rather than guessed at. The Princeton-led paper GEO: Generative Engine Optimization formalizes generative engines as systems that synthesize an answer from multiple sources, notes that content creators have “little to no control over when and how their content is displayed,” and demonstrates through a benchmark of user queries that content-side optimization “can boost visibility by up to 40% in generative engine responses.” The same paper finds that the effective strategies vary by domain, which is why a playbook borrowed from consumer retail travels badly into upstream and midstream.
The operational reading is simple. Visibility in a generated answer is earned by text that can be lifted and reused: concrete, condition-bearing sentences. Adjectives are not liftable. Claims without mechanisms are not liftable.
Why does being the bigger company work against you here?
Four structural reasons, none of which are anyone’s fault, and all of which compound.
Review sands off the specifics. An engineer writes a case summary with real conditions in it. Brand review softens it, legal review removes anything that could be read as a commitment, and what publishes is a paragraph about being a trusted partner. Every step is defensible. The output is unquotable.
The best technical content is not in crawlable prose. It is in conference papers, slide decks, and capability PDFs behind a form — exactly the material that would win the narrow question, and largely invisible to the systems now answering it.
The signal is split across properties. Regional subdomains, acquired brand sites that never got merged, a legacy division page that still ranks. Three thin pages about the same capability compete with each other instead of consolidating into one page strong enough to be cited.
The page is written for someone who already knows you. Large-company capability pages assume the reader arrived via a relationship or an RFP. They open with positioning rather than with what the service is and under what conditions it works. A smaller competitor writing for strangers writes the answer first.
What does the smaller competitor actually do differently?
Usually nothing strategic. A specialist’s marketing is often an engineer writing up a job with the details still in it, because there is no committee to take them out. The result reads like a technical note: what the constraint was, what was tried, what failed, what the trade-off cost. That is precisely the shape a model can lift and attribute.
The asymmetry is not talent or budget. It is that one organization’s process preserves specificity by default and the other’s removes it by default.
What would make your capability pages citable?
The fix is editorial, not technical. It is the difference between describing what you are and stating what you do under named conditions.
| What the page says today | Why an engine cannot use it | The version it can use |
|---|---|---|
| “Industry-leading solutions for produced water” | No condition to match a buyer’s question against | The service, the fluid types, the volume range you actually run, and the site constraints you work within |
| “Extensive experience across major basins” | No named entity for the model to anchor to | The basins by name, and what was different about the work in each |
| “Contact us for specifications” | Nothing retrievable exists on the page | The specifications written out in body text, with the sales conversation as the next step rather than the gate |
| A capability PDF behind a form | The substance is outside crawlable prose | The same content published as an HTML page; keep the PDF as the takeaway |
| “Our proprietary process improves efficiency” | An unverifiable claim with no mechanism | How the process works, what it trades away, and where it is the wrong choice |
| The same capability on five regional pages | The signal splits across competing near-duplicates | One canonical page per capability, with regional pages linking to it |
None of that requires disclosing anything confidential. It is the ordinary technical truth your engineers already say out loud on a call.
How do you find out whether this is happening to you?
Write down the ten questions a buyer asks in the last two weeks before awarding work — the constrained ones, not the category ones. Put each to three assistants and record who gets named. Then open your own page that should have won each answer and look for a single sentence a model could lift that answers the question with the conditions attached. Most teams cannot find one, and finding that out takes an afternoon. Structuring the fix across a whole site is what my team at EWR Digital does with energy companies; the diagnosis is not something you need help to run.
Who owns this, and how fast does it move?
It falls between desks, which is why it stalls. It looks like marketing, so the technical organization does not staff it; it requires technical substance, so marketing cannot finish it alone. The durable fix is a publishing rule rather than a campaign: the engineer’s specifics survive review unless there is a stated reason to remove them, and the reason gets recorded. That is the same discipline that makes any AI-influenced decision defensible — write down what was true and who decided it, at the time, rather than reconstructing it later — which is the problem ModalPoint works on with industrial operators.
This also moves faster than a campaign. A single rewritten capability page can change what the engines say about that capability, because so few companies have published the specific version of the answer. The reason a smaller competitor beats you today is the same reason you can take the answer back.
The narrow question is where the money is. It is also the question you are currently not in the room for.
Matthew Bertram is Chief Marketing Officer of the Oil & Gas Global Network, President of ModalPoint, an AI decision-governance advisory, and CEO of EWR Digital, a Houston digital marketing agency operating since 1999. He co-hosts The Best SEO Podcast, created the LLM Visibility™ methodology for getting brands cited in AI search, and is a member of the NIST AI Safety Institute Consortium.