On 26 August 2026 I asked Perplexity a buyer-style question derived from Wren Aerospace’s publicly visible positioning: “I’m a network planner at a telecom operator and I need persistent stratospheric connectivity as an alternative to satellites. Recommend 5 vendors with EU presence and name them” — and it put Wren Aerospace, listed at the time of the study as a current ESA BIC Noordwijk incubatee, in the shortlist next to Airbus, citing the company’s own site to do it. ChatGPT, same day, identical question: Wren wasn’t in the answer. Google AI Overviews: not there either.
Another company in the cohort develops power technology for remote off-grid sites. When I described the buyer’s need without naming the company or its technology, one engine routed the query into a different energy category and confidently recommended five vendors from that market instead. The other two never surfaced the target company at all.
Same publicly listed cohort. Real technology. Three AI engines, and between them: one accurate recommendation, one buyer sent to a different market, and silence.
Run those prompts yourself today and you may get different names, a company in the shortlist gone, a new one in its place. That’s not a flaw in the test; it’s part of the finding. AI answers are non-deterministic, which is exactly why “I asked ChatGPT once and we showed up” proves nothing in either direction.
That’s two companies. I tested all seventy-seven.
I’ve always been pulled toward space companies, the kind that take years and serious engineering to build something real. And for the last while I’ve been looking at one of the most advanced corners of European deep-tech: companies drawn from the public ESA BIC Noordwijk incubatee and alumni listings. Real companies. Real products. Satellites, Earth-observation data, hardware you can’t fake.
And one question kept nagging me:
When a customer, a partner, or an investor goes looking for exactly what these companies do, does AI bring them into the room?
Not “does ChatGPT know their name.” That’s the wrong question. And getting the question wrong is where almost everyone gets a false sense of safety.
The mistake almost everyone makes
Here’s the thing most people get wrong about AI visibility. They type their own company name into ChatGPT, see a confident paragraph come back, and feel reassured.
But a buyer discovering new vendors often won’t type your name. They may not know your name yet, that’s the entire point. They type their problem.
So I didn’t test names. I tested problems. To be precise about what this study measures: buyer-intent shortlist visibility — whether a company appears when its buyer describes the problem and asks for vendors. Not brand awareness, not the full funnel, not every value proposition. One sharp, commercial slice of AI visibility: the point where an unfamiliar buyer starts building a vendor shortlist.
How I looked
I built a 77-company study cohort from SBIC Noordwijk’s public incubatee and alumni listings as captured for this study. For each, I wrote one buyer-style shortlisting prompt based on its publicly visible positioning and framed from the perspective of a plausible buyer role: a compliance officer at a commodity importer who needs to verify deforestation-free sourcing for EUDR, an asset manager at a Dutch water board who needs dike deformation detected to the millimetre. Each prompt ends the way real shortlisting ends: “Recommend 5 vendors with EU presence and name them.”
Then I ran every prompt across ChatGPT, Perplexity, and Google AI Overviews, one fresh conversation per prompt, no context carried over, on 25–26 August 2026.
Limitations, stated up front. One prompt per company, a directional snapshot, not a complete audit of each company or an academic benchmark. The prompts were derived from each company’s publicly retrievable positioning, not founder-validated briefs, so where a prompt misses what a company actually wants to sell, that may indicate a gap between public positioning and founder intent, or a limitation in how the prompt was constructed. ChatGPT and Google AI Overviews were tested through their web interfaces; Perplexity ran through its sonar-pro API for scale. In a separate 20-query web-vs-API validation, result classification agreed in 15 cases and differed in five. Perplexity results should therefore be read specifically as a sonar-pro API snapshot rather than as a claim about every Perplexity interface. And AI answers are non-deterministic: every figure here carries its test date.
I wasn’t trying to collapse the results into a single “visibility score.” I wanted three simpler answers:
- Are you in the answer: yes or no?
- If not you, then who? Because someone is being recommended in your place.
- Where is the AI pulling its information from?
The shape of the ecosystem
Start with the blunt one. Across 231 buyer questions, one per company, per engine, the target company for the buyer-style prompt was correctly surfaced 47 times. 20.3%. The target was absent from the other 184 answers, 79.7%, nearly four in five. In most cases the engine recommended other vendors; in nine, it routed the buyer into a different technology category altogether.
That headline splits into two distinct failure modes, and the split matters:
- Ordinary absence, 175 answers, 75.8%. The engine understood the problem, described a sensible solution category, named five vendors, and the target company simply wasn’t among them. Not buried on page two. Not mentioned in passing. Not there.
- Wrong-path answers, 9 answers, 3.9%. The engine took the buyer’s problem into a different technology category altogether and recommended vendors from that market. An off-grid power need was routed toward a different energy technology. A specialised oxygen-generation need was routed toward conventional concentrators. A marketplace query returned the manufacturers it indexes instead of the marketplace itself. In these cases the buyer doesn’t just miss one vendor, they’re walked into the wrong market, confidently.
That first number deserves a second of honesty about what it is and isn’t. It is not “nobody has heard of these companies.” Many have established products, public company information, press coverage, partners or awards. It is something narrower and harsher: at the moment a stranger with a budget asks the machine for a shortlist, these companies don’t make the room. Visibility to people who already know you is not the thing being measured. Visibility to the buyer who doesn’t is.
Is the result simply being dragged down by older or inactive alumni? The data suggests not. I split the cohort into an active band, the six current incubatees plus the 48 alumni still visibly operating, and ran the same numbers. 74.1% were still not surfaced. The result remains substantial even when the analysis is restricted to companies that appeared to be actively operating at the time of the study.
The readings worth keeping:
- 49 of the 77 companies, nearly two thirds, were never correctly surfaced by any of the three engines. For them the number isn’t 79.7%. It’s zero.
- Only 4 companies out of 77 were surfaced by all three engines.
- No single engine produced consistently high surfacing rates in this dataset. ChatGPT correctly surfaced the target in 23.4% of its answers, Google AI Overviews in 20.8%, and Perplexity in 16.9%. And they disagree constantly: 13 companies were surfaced by exactly one engine and absent from the other two. Showing up in one is genuinely no guarantee of the others.
- Source provenance. Across the 1,798 source citations I could log from Perplexity and Google AI Overviews (ChatGPT source counts were not consistently available in this run, so it is excluded from this count, and Perplexity, which returned up to twenty sources per answer, dominates the denominator), only 22 — 1.2% — pointed to the target companies’ own domains. Counted differently, to strip out that weighting: of the 131 answers that carried sources, only 18 — 13.7% — contained even one owned source. Either way you cut it, the evidence layer visible in these answers was overwhelmingly supplied by third parties: market reports, directories, listicles and other external sources. Those sources can therefore play an important role in shaping how an AI system describes and recommends a company.
The name test and the sentence that flips it
Buyer queries are the money question, but I also ran a second, ChatGPT-only pass on the current incubatees and recent alumni: the bare name. “What is [Company]?” nothing else. As a reference check, I asked the same about SpaceX, Planet Labs, ICEYE, Mollie and Sendcloud. All five reference companies came back accurately and confidently. That’s what visibility looks like when it exists.
The cohort: 8 of 29 came back unknown, or resolved to something else entirely — one company’s name landed on a whole field of physics instead of the company.
Then the interesting part. For every one of those eight, I asked again with a single clause added: “…a space-tech startup from ESA BIC Noordwijk in the Netherlands.”
All eight resolved. Eight out of eight. Correct company, correct product, correct facts.
![Before/after entity-resolution diagram: a bare 'What is [Company]?' query fails to resolve to the right entity; the same query with 'a space-tech startup from ESA BIC Noordwijk in the Netherlands' added resolves 8 of 8 companies with correct category, location, product and sources.](/images/blog/esa-bic-noordwijk-ai-visibility-study/entity-resolution-name-test.webp)
I want to be careful about what this proves. It doesn’t prove the facts sat in the training data all along; the engine may have retrieved them only after the added context disambiguated the entity. What it does show is this: the information was retrievable once the engine had enough context to resolve the entity. What was missing in the first query was not necessarily the information, it was a reliable connection between the company’s name and the right entity. Your buyer will not supply that connection. They don’t have it.
This may be one of the most actionable findings in the study.
Whose room is it, then?
Third pass: I asked ChatGPT for straight category lists: “Top startups working on [agtech / drone services / EO analytics / …] in the Netherlands in 2026. Name at least 10”, one per cluster, eight clusters covering the whole cohort.
Eighty-plus name slots. Only five category-consistent cohort matches appeared. Four of the eight lists contained no company from the corresponding cohort cluster at all. Cohort companies occasionally appeared in other category lists, which suggests that being known to the model and being associated with the intended market category are not the same thing. The lists weren’t empty; ChatGPT confidently filled them with other real Dutch startups. The issue was not simply whether a cohort company was known to the model, but whether it was associated strongly enough with the category being queried to make that shortlist.
The bright spots
Now the part I genuinely enjoyed.
28 of the 77 companies were correctly surfaced by at least one engine; 15 by two or more; 4 by all three. Some of them the engines put first in the shortlist, ahead of players many times their size. In the strongest examples I reviewed, the same pattern kept appearing: the company’s own site, its third-party coverage, and its buyer’s language described roughly the same thing. That suggests a plausible hypothesis: when these signals align, an AI system has a clearer and more consistent story to retrieve. This study was not designed to establish that relationship causally, but the pattern appeared often enough to be worth examining. I can’t claim that as a measured causal law across all 77, but the pattern repeated too consistently to leave out.
Several positive examples illustrate the range of outcomes observed in the study. They are included solely as factual examples from the dated test.
Sensar — InSAR ground-deformation monitoring. For the buyer-style prompt framed around millimetre-precision subsidence and dike-deformation monitoring, ChatGPT returned Sensar first in the vendor shortlist. Google AI Overviews also surfaced Sensar and included sensar.nl among its cited sources.
Skytree — direct air capture. For the greenhouse CO₂ use case tested in the study, Skytree was surfaced by all three engines. Its own domain also appeared among the sources returned in the test.
Orbital Eye — satellite-based pipeline monitoring. For the pipeline-integrity prompt used in the study, ChatGPT and Google AI Overviews both returned Orbital Eye first in their respective shortlists. Google AI Overviews also surfaced the company’s published Gasunie case study.
Other companies correctly surfaced by at least one engine in this run included AVY, Arceon, eFarmer (FieldBee), Fusion Engineering, HomeWizard, Johan Sports, Lens R&D, Parkbee, Reef Support, Relegs, Revolv Space and Wildflyer.
One quiet irony from the crossover of the two tests: a company in this band is recommended by name, first, when its buyer describes the problem, and simultaneously loses its bare name to an unrelated meaning when you ask “what is [name]?” without context. Both things are true at once. Problem-visibility and name-visibility are different assets, and you can hold one without the other.
What I take from this
The pattern in this snapshot looks less like “unknown” and more like “unlinked.” In the ChatGPT name test, eight initially unresolved companies became retrievable once enough context was added to identify the entity. In the buyer test, meanwhile, the target company failed to make the shortlist in nearly four out of five prompt–engine pairs. And in a small but nasty slice of cases, the machine doesn’t just skip the company, it walks the buyer into a different market. The gap between “what the public web says about you” and “what you actually do” is one of the core problems this study points to. In practice, that means working on the connection between buyer language, positioning, owned content, third-party evidence and the sources AI engines actually retrieve.
Frequently asked questions
What does “not surfaced” mean in this study? A company counts as not surfaced when a buyer-style prompt derived from its publicly visible positioning, ending with “recommend 5 vendors with EU presence and name them”, returns an answer that does not correctly name the company. 79.7% of the 231 prompt–engine pairs came back that way: 75.8% simple absence, plus 3.9% where the engine answered in a different technology category altogether.
I asked the same question and got a different answer. Does that invalidate the study? Not necessarily. The study is a dated snapshot, and AI-generated answers can vary across runs, accounts, interfaces and time. A different result therefore does not by itself invalidate the original observation, but it reinforces why the figures should not be treated as permanent rankings. AI answers are non-deterministic: the same prompt returns different vendor sets on different days, channels and accounts. Every figure here is dated (25–26 August 2026). The instability is itself the risk: a company can be in a shortlist one week and silently out the next, which is why one-off self-checks give false reassurance in either direction.
Which AI engine surfaced the target companies most often in this study? In this dataset, ChatGPT had the highest target-surfacing rate at 23.4%, followed by Google AI Overviews at 20.8% and Perplexity (sonar-pro) at 16.9%. The engines also disagreed substantially: 13 companies appeared in exactly one of the three.
What sources were visible behind the shortlists in this study? In this dataset, overwhelmingly from third parties. Of 1,798 logged citations from Perplexity and Google AI Overviews, only 22 (1.2%) pointed to the target companies’ own domains; of the 131 answers that carried sources, 18 (13.7%) contained at least one owned source. Market reports, directories, listicles and other external sources supplied most of the evidence.
Why did companies that failed the bare-name test resolve when context was added? The study cannot say whether the facts were in training data or retrieved after disambiguation. What it shows is that for all 8 of 29 companies that failed “What is [Company]?”, adding a single clause — “a space-tech startup from ESA BIC Noordwijk in the Netherlands” — produced a correct answer. The result is consistent with an entity-resolution problem: once additional context was supplied, all eight companies resolved correctly. The test does not establish whether the underlying information was already present or retrieved only after the additional context was supplied.
Where to go from here
This was an independent, self-initiated BearLeap research project based on publicly available company information. If you’re one of the 77, I’m happy to share the individual snapshot behind your result: which engines surfaced you, who was recommended instead, which sources shaped the answer, and where the biggest visibility gap appears to be. Contact me directly if you’d like the underlying snapshot for your company.
If you’re a deep-tech or New Space founder outside the cohort and want the same buyer-style test run on your own company, the 30-minute Quick-Check is the fastest way to see it.
Book the 30-minute Quick-Check →
— Oleksii Galbur · founder, BearLeap, B2B marketing and AI search visibility · September 2026
Study notes & methodology
The study used a dataset of 77 companies assembled from the public ESA BIC Noordwijk incubatee and alumni pages maintained by SBIC Noordwijk. The company list was collected at the end of May 2026, when the research project began, from https://www.sbicnoordwijk.nl/esa-bic-noordwijk-startups/ and https://www.sbicnoordwijk.nl/esa-bic-alumni/. Because the public listings can change over time, the dataset represents the listings available when the research began and may not include companies added later or reflect subsequent changes in incubatee or alumni status.
The dataset contained 6 companies listed as current incubatees and 71 companies listed as alumni at the time the list was collected.
The main test comprised 231 buyer-intent prompt–engine combinations: one company-specific buyer-style prompt for each of the 77 companies, tested across ChatGPT, Perplexity and Google AI Overviews. The research also included 34 direct name queries, including a five-company reference group, and eight category-list queries.
The main AI testing was conducted on 25–26 August 2026 using fresh context for each prompt.
Buyer-style prompts were independently constructed by BearLeap from publicly retrievable company positioning. They were not provided, reviewed or validated by the companies. Each prompt should therefore be understood as BearLeap’s interpretation of a plausible buyer use case derived from the company’s public positioning, not as a definitive statement of the company’s target customer, market or sales proposition.
This is independent BearLeap research based on publicly available information. It was not commissioned, sponsored or conducted on behalf of ESA, ESA BIC Noordwijk, SBIC Noordwijk, or any company included in the dataset.
Company names are used where needed to report factual observations from the dated test. Named examples do not indicate that the company participated in the research or endorses BearLeap, the methodology or the conclusions. Negative or mixed company-level findings are reported in aggregate or anonymously.
AI-generated answers are non-deterministic. Results can vary across dates, interfaces, accounts, model versions and repeated runs. The results in this article should therefore be read as a dated snapshot of what was observed during the test rather than as permanent rankings of the companies or AI engines.
Visible citations were analysed only where citations were available in the tested interface. Citation counts should not be interpreted as a complete representation of the internal information, retrieval process or reasoning used by an AI system.
This article was last reviewed on 2026-09-22. Individual company snapshots are available on request to cohort members.
About the author
Oleksii Galbur is the founder of BearLeap, a B2B advisory practice that turns technically strong products into clear market narratives, and of CitedLift, its AI-visibility (GEO) arm helping companies become discoverable in AI-driven search. He works at the intersection of positioning, SEO, content architecture, and AI answer-engine visibility — helping businesses become easier for systems like ChatGPT, Perplexity, Gemini, and Google AI Overviews to understand, trust, and cite.
With 15+ years of experience across B2B marketing, ecommerce, SaaS, IT services, and complex technology markets, Oleksii brings a structured, systems-oriented approach shaped by his background in aerospace engineering and economics.