--- title: "Why AI Engines Don't Recommend Small B2B Companies: What Two 2026 Citation Benchmarks Show Together" date: 2026-09-04T21:11:36Z modified: 2026-09-04T21:20:48Z permalink: "https://spotlightonstartups.com/why-ai-engines-dont-recommend-small-b2b-companies-what-two-2026-citation-benchmarks-show-together/" type: post status: publish excerpt: "" wpid: 8294 categories: - Expert Insights tags: - Expert Insights - AI Citation Benchmark - AI Search - AI visibility - Answer Engine Optimization - B2B Marketing - B2B SaaS - ChatGPT - discovery blackout - Gadex Research - Gemini - Generative Engine Optimization - original research - Perplexity - Small Business featured_image: "https://spotlightonstartups.com/wp-content/uploads/2026/09/2026-B2B-AI-Citation-Benchmark-Comparison.png" featured_image_alt: 2026 B2B AI Citation Benchmark Comparison author: Gregg Kell timestamp: 2026-09-04T21:20:48Z --- **Two independent studies published two weeks apart measured opposite halves of the same system. Read together, they explain something neither could establish alone: the discovery blackout is not a recognition failure or a content failure. It is a consequence of what the recommendation layer is built out of.** **[Gregg Kell, Founder and Editor](https://www.linkedin.com/in/greggkell/)** · **Spotlight on Startups Analysis** · Published 4 September 2026 · Synthesizing fieldwork conducted 11 August and 17 to 19 August 2026 ## Answer first ![2026 B2B AI Citation Benchmark Comparison](https://spotlightonstartups.com/wp-content/uploads/2026/09/2026-B2B-AI-Citation-Benchmark-Comparison.png) In August 2026, Gadex Research tested 50 B2B software buying prompts across ChatGPT, Gemini and Perplexity and coded **[1,576 citations from 150 answers](https://gadex.ai/resources/b2b-ai-citation-benchmark-2026/)**. One week later, Spotlight on Startups tested 30 small and mid-sized B2B companies against the same three engines and coded **[720 observations from 504 runs](https://spotlightonstartups.com/wp-content/uploads/wp-mfa-exports/post/2026-b2b-ai-citation-benchmark.md)**. Gadex asked what kind of source an engine cites. We asked whether a particular kind of company ever gets named. Neither study was designed with the other in mind, and they were run by different teams using different methods on different dates. Put side by side, the Gadex source mix supplies a mechanism for our zero. The pages engines draw on when assembling a recommendation are, in the Gadex sample, overwhelmingly comparison and list content plus first-party material from large vendors. A 40-person managed service provider has no meaningful presence in either inventory. It is not being rejected at the moment of recommendation. It was never in the pool that the recommendation is assembled from. That reframing matters because it inverts the most common piece of advice in this field. ## The two studies | Gadex Research | Spotlight on Startups | | | --- | --- | --- | | Question | Which sources get cited in B2B buying answers? | Do engines recommend small and mid-sized B2B firms? | | Unit of analysis | Citation occurrences | Companies | | Scale | 1,576 citations, 1,211 URLs, 632 domains, 150 answers | 720 coded observations, 504 independent runs, 30 companies | | Prompts | 50 category-level buying prompts | 8 fixed prompts per company, split into discovery and named-company groups | | Engines | ChatGPT, Gemini, Perplexity via the DataForSEO LLM Responses API | The same three, run in their consumer interfaces | | Fieldwork | 11 August 2026 | 17 to 19 August 2026 | | Headline finding | 21.6% of citation occurrences came from a domain in Google’s organic top ten for the matching seed query | 0 of 54 discovery runs named any of the 30 companies tested | | Open data | **[Citation CSV](https://gadex.ai/data/b2b-ai-citation-benchmark-2026/gadex-b2b-ai-citations-2026.csv)**, prompt CSV, answer JSONL, **[methodology JSON](https://gadex.ai/data/b2b-ai-citation-benchmark-2026/methodology.json)** | **[720 coded observations and 168 permanent result URLs](https://spotlightonstartups.com/ai-citation-benchmark-dataset/)** | Both studies publish their raw data. Both disclose that they were produced by companies that sell services in this market. Readers should weigh both accordingly, and both datasets are open specifically so that they can. ## The mechanism, in three steps ### Step one: discovery answers are assembled from a narrow inventory Gadex classified every cited page by format. Across all three engines, comparison, list and review pages accounted for 56.9% of citation occurrences. In Perplexity the share was 78.3%. In ChatGPT the picture was different but equally concentrated: 83.3% of its citations went to domains their mapping identified as software vendor owned, and pricing pages alone made up 34% of its citations. The engines are not surveying the open web when asked who to hire. In this sample they are drawing on two things: third-party comparison inventory and first-party material from vendors that the comparison inventory already covers. ### Step two: that inventory is dominated by large vendors and a small set of rankers The most cited domains in the Gadex sample were YouTube, Reddit, G2, Forbes and Zapier, with monday.com, PCMag, ADP, HubSpot and Salesforce close behind. Their ChatGPT vendor citations went to HubSpot, Microsoft, Salesforce, Monday.com, Zoho, Asana, ClickUp, ADP and Tableau. Our own discovery runs, using entirely different prompts against a different company set, produced a matching cast. Asked about staffing, the engines named Randstad, Robert Half, ADP, Adecco, ManpowerGroup, Insperity and TriNet. Asked about consulting, they named McKinsey, BCG, Bain, Deloitte, Accenture and EY-Parthenon. They leaned repeatedly on Forbes, Vault, G2 and ClearlyRated. Two studies, different designs, same gatekeepers. That is the strongest form of corroboration available when neither study can be a replication of the other. ### Step three: the blackout follows from steps one and two Our thirty companies were cited in 97.5% of observations when we handed the engine the name. The engines returned founding years, head office cities, named founders and accurate service lines for firms with a few dozen employees. The information exists in the models. But a discovery answer is not built by consulting what the model knows. On the Gadex evidence it is built by retrieving a small set of comparison pages and large-vendor pages and summarizing them. Our companies appear in neither. So they are absent, unanimously, on all three engines, with no partial credit and no category in which one broke through. The zero stops being surprising once you know what the answer is made of. ## The finding this reverses The single most common prescription in AI visibility work is to publish better first-party pages. The Gadex ChatGPT number appears at first to support it: 83.3% of citations went to vendor-owned domains, which reads like proof that owned content earns inclusion. Our data suggests the causal arrow runs the other way, and Gadex is careful to say their own result is not proof of causation. All thirty of our companies have websites, service pages and in most cases pricing or packaging information. None of it put them into a discovery answer. Meanwhile the vendor pages that did get cited belong to companies that were already in the recommendation set. The reading consistent with both datasets is that first-party citation is a downstream consequence of selection, not a route into it. Once an engine has decided to recommend HubSpot, it cites HubSpot’s pricing page to support the recommendation. Publishing a better pricing page does not appear to be what caused HubSpot to be selected. Neither study can prove that. Neither intervened on anything. But it is the interpretation that survives contact with both sets of numbers, and it is a materially different problem statement from the one most published advice addresses. ## The second bind: Google is not the back door Gadex found that only 21.6% of citation occurrences came from a domain that also appeared in Google’s organic top ten for the matching seed query, with engine level overlap ranging from 20.6% to 25.8%. This is a narrower comparison than it first appears, and Gadex says so: it is a same-day, domain-level check against a shortened seed query rather than the full conversational prompt. A cited page could rank well for something else entirely. Even read conservatively, it closes off an obvious escape route. A small firm looking at our blackout might reasonably conclude that the fix is conventional SEO, on the theory that ranking on Google eventually feeds the engines. The Gadex overlap number says a Google page one snapshot and an AI citation set are substantially different objects. Ranking is not a reliable path in, and it is not sufficient once you are there. So the combined position for a small B2B firm is: absent from discovery answers, and unable to buy a way in through the ranking channel that most marketing budgets are pointed at. ## Where the two studies point the same way on categories Both studies found that category explains more than company-level attributes do, though they measured different things and the convergence should be treated as suggestive rather than established. Gadex found Google overlap varying from 34.2% in finance operations down to 12.5% in customer support, and found AI and automation the most fragmented category at 103 distinct domains across 155 citations. We found something structurally similar from the opposite direction. In B2B project management software, six vendors took 56% of every recommendation slot regardless of how the question was phrased. In management consulting, eight firms took 55%. In financial services, staffing, IT services and PR, no firm at all survived all three rephrasings. Both datasets say the same thing at a high level: the shape of the answer is a property of the category, not of the vendor. What neither says is which category conditions favor entry. Our own first guess was wrong on this. We assumed the categories with the most entrenched incumbents would be hardest to enter, and the data did not support it. ## Where the two studies cannot be merged It would be easy to blend these numbers into one figure. They should not be blended, for four reasons. **Different model tiers.** Gadex accessed the engines through an API layer and reported the models returned as `gpt-5.4-mini-2026-03-17`, `gemini-3.5-flash-lite` and Perplexity `sonar`. Those are the small and fast variants. We ran the consumer interfaces, which is what a buyer actually uses. Interface behavior and API behavior can differ, and Gadex flags this scope limit directly. **Different units.** A Gadex citation occurrence is one normalized URL inside one answer. One of our observations is one company coded against one prompt on one engine. The denominators are not compatible. **Different dates.** One week apart, in a system that drifts between model versions. **Different classification frames.** The Gadex residual class, labeled publisher, agency or unclassified, holds 58.9% of all their citations and may contain unmapped vendors. Their vendor-owned share is what their version 1.0 mapping identified, not a census. Our coding rule is a three-way judgment applied by a single assessor with no second rater and therefore no inter-rater reliability figure. Every one of these is disclosed by the study that carries it. That is what makes the synthesis possible at all, and it is why the correct output here is a shared mechanism rather than a merged statistic. ## What neither study establishes Neither is an intervention. Neither can tell you that changing anything produces a citation. Our study set out to test whether AI readiness scores predict citation and could not answer it, because on discovery prompts every company scored zero and a correlation needs two quantities that vary. On the one prompt band with real variation the relationship came out unresolved at r = +0.33, p = 0.072, with a confidence interval running from −0.03 to +0.62. Around 69 companies would have been needed for adequate power. Gadex is explicit that their experiment identifies patterns and not causal ranking factors, cannot give the probability that any particular page will be cited, and cannot speak to persistence over time or to non-commercial query types. Combining two observational snapshots does not produce causal evidence. It produces a better-specified hypothesis. The hypothesis is that inclusion in AI recommendations is gated by presence in third-party comparative inventory rather than by the quality or completeness of a firm’s own pages, and that this gate operates upstream of everything a company publishes on its own domain. That is testable. It is not yet tested. ## What would settle it The experiment that resolves this is an intervention study: take a set of companies absent from discovery answers, place a subset into credible third-party comparative coverage in their categories, hold the rest as a control, re-run a frozen prompt set on a schedule and see whether the treated group enters the recommendation set. That design would separate the two competing explanations. If third-party comparative presence is the gate, treated companies should appear and controls should not. If the gate is something else, size, revenue, brand search volume, age, nothing will move. We intend to re-run our benchmark annually against the same categories and prompt structure with a sample sized for the power this edition lacked. Gadex has published version 1.0 with a stated commitment to release future waves as new dated versions rather than silently replacing the current one. Two independent longitudinal series measuring different halves of the same system is a better position for this field than either study alone. We would change our minds about the central finding if a repeat run produced discovery citations at any meaningful rate, if a larger sample showed readiness scores predicting citation, or if the pattern turned out to be an artifact of category selection. ## Sources 1. Gadex Research (2026). **[The B2B AI Citation Benchmark 2026](https://gadex.ai/resources/b2b-ai-citation-benchmark-2026/)**. 150 API answers across ChatGPT, Gemini and Perplexity, 1,576 citation occurrences. Fieldwork 11 August 2026. Open data: **[citation CSV](https://gadex.ai/data/b2b-ai-citation-benchmark-2026/gadex-b2b-ai-citations-2026.csv)**, **[prompt CSV](https://gadex.ai/data/b2b-ai-citation-benchmark-2026/gadex-b2b-ai-prompts-2026.csv)**, **[answer JSONL](https://gadex.ai/data/b2b-ai-citation-benchmark-2026/gadex-b2b-ai-answers-2026.jsonl)**, **[methodology JSON](https://gadex.ai/data/b2b-ai-citation-benchmark-2026/methodology.json)**. Gadex research methodology: **[gadex.ai/methodology](https://gadex.ai/methodology/)**. 2. Spotlight on Startups (2026). **[The 2026 B2B AI Citation Benchmark](https://spotlightonstartups.com/wp-content/uploads/wp-mfa-exports/post/2026-b2b-ai-citation-benchmark.md)**. 30 companies, 8 prompts, 720 coded observations from 504 independent runs. Data collected 17 to 19 August 2026. Full dataset: **[720 coded observations with permanent result URLs](https://spotlightonstartups.com/ai-citation-benchmark-dataset/)**. **[PDF edition](https://spotlightonstartups.com/wp-content/uploads/2026/08/2026-B2B-AI-Citation-Benchmark.pdf)**. ## Citing this analysis Spotlight on Startups (2026). Why AI Engines Don’t Recommend Small B2B Companies: What Two 2026 Citation Benchmarks Show Together. A synthesis of Gadex Research’s B2B AI Citation Benchmark 2026 and the Spotlight on Startups 2026 B2B AI Citation Benchmark. Journalists and researchers may reproduce any figure with attribution. ## Disclosure Spotlight on Startups is a d/b/a of Kell Web Solutions, Inc., which sells services related to AI visibility. This analysis reaches no conclusion that a purchasable service resolves the pattern it describes, and states plainly that the central mechanism remains an untested hypothesis. Gadex AI, operated by ALPHAOSCAR EURL, also sells services in this market. Neither company commissioned, reviewed or funded the other’s study, and there is no commercial relationship between them.