A two-wave public benchmark of how AI assistants describe the AI companies themselves — OpenAI, Anthropic, Google DeepMind, Meta AI, xAI, and 20 others — across ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews. Self-citation bias measured per engine.
Share of all citations surfaced by AI assistants when answering questions about AI companies themselves. January–May 2026, 32,200 prompts across two waves.
The AI industry is the most-covered story of 2026. The same AI assistants that drive that coverage — ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews — are also the systems consumers, founders, journalists, and investors increasingly use to make sense of who is leading the AI industry, which models are best, and which companies are worth tracking.
The AI Companies AI Visibility Index 2026 measures how often AI companies themselves are surfaced, cited, and recommended inside those answers. The Index analyzed 32,200 prompts across the five leading AI assistants, run in two independent waves — Wave 1 in January–February 2026, Wave 2 in April–May 2026. The category is more volatile than banking, venture capital, or credit cards because of frequent model releases and active news cycles; the two-wave structure reports only findings that held across both waves within ≤2.0 percentage-point tolerance.
Citation Share · The share of retrieved AI responses in which a company is named, recommended, or cited as a referenced source. Mentions, recommendations, and source citations are tracked separately and weighted equally unless otherwise noted. Self-Citation Lift · Within recommendation queries only, the ratio of how often an AI engine recommends its parent company's models compared to how other engines recommend the same models. A lift of 1.0x indicates no self-bias.
In the dataset, the most-cited AI company — OpenAI — accounts for 24.6% of cited responses across the five engines. Anthropic follows at 14.8%. Google DeepMind follows at 13.7%. Meta AI follows at 9.7%. xAI follows at 5.7%. The top five together account for 68.5% of all observed AI-company citations in the test set.
The structural observation that distinguishes this category from every other 5W has measured: code repositories and research papers — GitHub and ArXiv — together supply 31.2% of citations, second only to Wikipedia (24.3%). In banking, venture capital, and credit cards, the top citation sources are editorial publishers. In AI companies, technical content functions as primary source material.
The most consequential observation is behavioral. In recommendation queries specifically, every AI assistant recommended its parent company's models more often than other engines recommended the same models — except one. ChatGPT recommended OpenAI 2.0x more often than other engines did. Gemini recommended Google DeepMind 1.7x more often. Google AI Overviews recommended Google models 1.6x more often. Claude recommended Anthropic models 1.2x more often — the lowest self-citation lift of any major AI engine tested. The pattern held in both waves.
"The most uncomfortable finding in our entire AI Visibility Index series is in this dataset. Every major AI assistant favors its parent company's models in recommendation queries — at measurable, repeatable, two-wave-stable margins. The exception is Claude. That asymmetry is the story."
Ronn Torossian · Founder and Chairman, 5W
All numerical findings reflect the average of Wave 1 (Jan–Feb 2026) and Wave 2 (Apr–May 2026). Cross-wave deltas are reported where greater than 1.5 percentage points. Findings that did not hold across both waves are excluded from this Index. Note: AI companies showed higher cross-wave volatility than banking, venture capital, or credit cards — reflecting frequent model releases and active news cycles during the testing period.
OpenAI (24.6%), Anthropic (14.8%), Google DeepMind (13.7%), Meta AI (9.7%), and xAI (5.7%) together accounted for 68.5% of cited responses across both waves. The remaining 31.5% distributed across 20 other AI companies in the test set. Cross-wave delta on the top five: ≤1.8 pp. Concentration at the top is higher in AI than in any other category 5W has measured.
GitHub (17.8%) and ArXiv (13.4%) together account for 31.2% of AI-company-related citations — second only to Wikipedia (24.3%). This pattern does not appear in any other category 5W has measured. Banking, venture capital, and credit cards are described through editorial publishers; AI companies are described through their own code and their own papers. The implication: AI company communications strategy is fundamentally different from any other industry's.
Within recommendation queries: ChatGPT recommended OpenAI 2.0x more often than other engines did. Gemini recommended Google DeepMind 1.7x more often. Google AI Overviews recommended Google 1.6x more often. Perplexity surfaced Perplexity in adjacent search-tool queries at 2.4x baseline. Claude recommended Anthropic 1.2x more often — substantially lower than every other engine tested. The pattern held in both waves with cross-engine delta ≤0.4x.
Self-Citation Lift = ratio of how often an engine recommends its parent company's models compared to how other engines recommend the same models. Recommendation queries only. Average of both waves. Baseline of 1.0x = no self-bias.
For prompts asking about open-source models, fine-tuning, deployment, or technical comparisons, open-source labs surfaced at 2.4x their general-query rate. Mistral AI, Meta AI (Llama), Hugging Face, and DeepSeek each captured citation share inside technical query clusters meaningfully above their overall rankings. The pattern held in both waves and suggests AI assistants retrieve from GitHub and ArXiv differently when the prompt signals developer intent.
Across U.S. consumer-facing queries, the major Chinese labs (DeepSeek, Alibaba Qwen, Moonshot Kimi, Baidu ERNIE, Zhipu) registered less than 2% combined Citation Share. In technical and open-source-model queries specifically, DeepSeek surfaced in 18.4% of responses — third only to Meta and Mistral. The retrieval gap between consumer and technical queries is the largest of any company segment in the test set.
Despite Elon Musk's X platform reach and xAI's status as one of the most-covered AI companies in 2025–2026, xAI captured only 5.7% Citation Share — below Meta AI (9.7%) and well below the company's share of voice in tech and business press. X/Twitter activity did not translate proportionally to AI citation in this dataset. The mechanism appears to be that social-media coverage does not generate the GitHub, ArXiv, and Wikipedia source material that drives most AI-company retrieval.
For prompts specifically about AI safety, alignment, interpretability, constitutional AI, or responsible AI deployment, Anthropic surfaced at 31.2% — substantially above its overall 14.8% share. OpenAI surfaced at 19.8% in the same query set. Google DeepMind at 14.6%. The category specialization is the clearest brand-positioning signal in the dataset and held in both waves.
Application-layer AI companies — Glean, Harvey, Cursor, Replit, Runway, Adept, Character.AI — surfaced in less than 1.5% combined Citation Share across foundation-model queries. They surfaced more frequently in vertical-specific prompts (legal AI, coding AI, video AI), where each captured 8–22% of relevant queries. The implication: the application-layer category is functionally invisible inside "best AI company" prompts but well-positioned inside "best AI for [task]" prompts.
Citation Share by engine for the top 10 AI companies in the test set. Darker cells = higher share. The diagonal pattern of elevated self-citation is visible across ChatGPT/OpenAI, Gemini/Google DeepMind, and Perplexity/Perplexity — with Claude/Anthropic showing the most balanced distribution.
Mean cross-wave delta for top-10 AI companies was 1.7 pp — roughly 2x the volatility observed in Banking, Venture Capital, or Credit Cards. Three companies showed Δ > 2.5pp: Mistral AI (Wave 2 +3.1pp on new model release coverage), xAI (Wave 2 -2.7pp), DeepSeek (Wave 2 +3.4pp on open-source release cycles). The Index excludes findings that did not hold within tolerance across both waves; this category demands more frequent re-measurement than slower-moving industries.
Wikipedia surfaced in 22.1–26.8% of retrieved responses across all five engines — the most uniform citation source in the dataset. The variance lay in the runners-up: ChatGPT weighted ArXiv more heavily than other engines; Gemini weighted technical Google domains (Google Research, DeepMind papers); Perplexity weighted recent news (TechCrunch, The Information); Claude weighted ArXiv and direct company blog posts more evenly than competitors. Wikipedia presence is necessary but not sufficient; the second-tier citation pattern is where engine differentiation emerges.
| Rank | Company | Citation Share |
|---|---|---|
| 01 | OpenAI | 24.6% |
| 02 | Anthropic | 14.8% |
| 03 | Google DeepMind | 13.7% |
| 04 | Meta AI / FAIR | 9.7% |
| 05 | xAI | 5.7% |
| 06 | Mistral AI | 4.5% |
| 07 | Perplexity | 4.1% |
| 08 | Hugging Face | 4.0% |
| 09 | DeepSeek | 3.8% |
| 10 | Cohere | 2.2% |
| 11 | Stability AI | 2.0% |
| 12 | Inflection AI | 1.6% |
| 13 | Character.AI | 1.5% |
| 14 | AI21 Labs | 1.3% |
| 15 | Alibaba (Qwen) | 1.2% |
| 16 | Moonshot AI (Kimi) | 1.0% |
| 17 | Adept | 0.8% |
| 18 | Runway | 0.8% |
| 19 | Reka AI | 0.7% |
| 20 | Zhipu AI | 0.6% |
| 21 | EleutherAI | 0.5% |
| 22 | Allen Institute for AI (AI2) | 0.5% |
| 23 | Together AI | 0.4% |
| 24 | Baidu (ERNIE) | 0.4% |
| 25 | ByteDance (Doubao) | 0.4% |
| Engine | Parent Company | Lift vs Other Engines |
|---|---|---|
| Perplexity | Perplexity (narrow application — search queries) | 2.4x |
| ChatGPT | OpenAI | 2.0x |
| Gemini | Google DeepMind | 1.7x |
| Google AI Overviews | 1.6x | |
| Claude | Anthropic (lowest lift of major engines) | 1.2x |
Lift = engine's parent-company citation rate in recommendation queries ÷ average citation rate across other engines for the same parent. Baseline of 1.0x = no self-bias. A lift of 2.0x means the engine recommends its parent company twice as often as other engines do in identical prompts.
| Rank | Source | Share of Citations |
|---|---|---|
| 01 | Wikipedia | 24.3% |
| 02 | GitHub (code repositories) | 17.8% |
| 03 | ArXiv (research papers) | 13.4% |
| 04 | AI company-owned domains (combined) | 8.6% |
| 05 | TechCrunch | 6.7% |
| 06 | The Verge | 5.2% |
| 07 | The Information | 4.8% |
| 08 | Wired | 3.9% |
| 09 | The New York Times | 3.4% |
| 10 | All other publishers | 11.9% |
"Across two waves, GitHub and ArXiv accounted for 31% of citations about AI companies — second only to Wikipedia. No other industry 5W has measured surfaces code and research papers as primary sources. The implication: AI company communications strategy is unlike any other industry's, and the firms that publish frequently into those surfaces compound advantage."
Ronn Torossian · Founder and Chairman, 5W
The AI Companies AI Visibility Index 2026 analyzed 32,200 prompts across ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews, run in two independent waves to test for stability across retrieval drift, model updates, and the high cross-wave volatility characteristic of this category.
Wave 2 used the same prompt set as Wave 1 with no modification. Only findings stable across both waves within reporting tolerance (≤2.0 percentage-point delta on company-level Citation Share; ≤2.5 pp on source-level share) are published here. This category showed higher cross-wave volatility than banking, venture capital, or credit cards — driven by frequent model releases, active news cycles, and ongoing competitive dynamics during the testing period. The wider tolerance band reflects that reality without compromising stability requirements.
Queries simulated real founder, developer, journalist, investor, and policymaker research behavior. Prompts included branded company queries ("What does OpenAI do?"), non-branded category queries ("Best AI company in 2026"), comparison queries ("OpenAI vs Anthropic"), capability queries ("Best AI for coding", "Best AI for safety research"), recommendation queries ("Which AI should I use for my startup?"), and technical queries ("Best open-source LLM"). Prompts were distributed evenly across the five engines so that each engine received the same prompt mix per category.
Self-Citation Lift was calculated within recommendation queries only — prompts that explicitly ask the AI assistant to suggest or recommend a model, company, or tool. For each engine with a parent company in the test set, the engine's citation rate for its parent's products was compared against the average citation rate for the same parent across the other four engines in identical prompts. A lift of 1.0x indicates no self-bias. A lift of 2.0x indicates the engine cites its parent twice as often as other engines do. Lift was averaged across both waves.
Each prompt was issued three times per engine within a wave, with responses sampled at varied time-of-day windows to reduce within-engine retrieval drift. Reported Citation Share values represent the average of all retrieved responses for a prompt across both waves.
Three distinct response signals were tracked per retrieved answer:
For the headline metric reported as Citation Share, all three signals were weighted equally per retrieved response. Source-level analyses (the publisher rankings in Figure 01) used source citations only.
Each engine was weighted equally in aggregate figures, regardless of differences in response length or default-citation frequency per engine. This was a deliberate choice: weighting by raw citation volume would have over-weighted Perplexity and Google AI Overviews, both of which return more citations per response than ChatGPT, Claude, or Gemini by default. Where per-engine results diverge, those differences are reported separately (Figures 02 and 03).
This benchmark does not separately attribute citations to retrieval (live web search at query time) versus training data (pre-trained associations). The two are increasingly entangled in production AI assistants and not externally observable. The Self-Citation Lift finding (Figure 02) cannot fully distinguish between retrieval bias and training-data bias — both contribute. Where engine-level behavior differs from a pure retrieval-only baseline, that variance is interpreted as a combination of retrieval architecture, training-data composition, and engine-specific ranking heuristics.
Results reflect sampled outputs during a defined testing window. AI models, training data, retrieval indexes, and ranking systems evolve continuously; results may shift outside the test period. The category measured here — AI companies themselves — shows higher cross-wave volatility than slower-moving industries and demands more frequent re-measurement. The Index is best read as a structured snapshot of observed system behavior across two waves, not as a continuous live measurement. The full prompt set, per-engine response logs, and category-level datasets are available on request for replication.
Disclosure: 5W has commercial relationships in the AI industry and operates as a senior advisor to Curium.io, the team that coined Generative Engine Optimization. Companies cited in this Index were selected based on observed Citation Share in the test set; inclusion does not imply a commercial relationship. 5W does not accept compensation for ranking placement.
GitHub supplied 17.8% of all AI-company citations in the dataset — second only to Wikipedia. For an AI company, the GitHub organization page, repository READMEs, model cards, and release notes function as primary brand content the AI assistants retrieve directly. Most communications teams do not treat GitHub as a brand surface. The AI does. Audit the organization page, top repositories, and recent release notes for clarity, recency, and citation-readiness.
ArXiv supplied 13.4% of citations. Companies with sustained ArXiv publishing cadences (OpenAI, Anthropic, Google DeepMind, Meta, Mistral) over-index in technical queries by 2.4x. Companies with strong commercial traction but limited ArXiv presence under-index regardless of revenue. ArXiv is not just for research — it is the most efficient citation surface for technical AI authority.
Wikipedia surfaced in 22.1–26.8% of retrieved responses across all five engines — the most uniform source in the dataset. For AI companies, founder Wikipedia entries are as critical as the company entry. Several of the top 25 AI companies in this Index have outdated, thin, or missing Wikipedia entries — including for high-profile founders. Wikipedia is the single highest-leverage GEO surface in this category.
AI-company-owned domains supplied only 8.6% of citations in aggregate — a low number, but with significant variance by company. Companies that publish detailed model cards, structured benchmark results, capability documentation, and changelog content on their own domains were cited at meaningfully higher rates than companies relying on landing pages and feature copy. Structured technical content is citation-eligible. Marketing-style content is not.
The companies with the highest Citation Share in this dataset — OpenAI, Anthropic, Google DeepMind — all operate research blogs that publish at a measurable cadence, with content that bridges technical depth and editorial accessibility. The cadence (not the volume of any single post) is the variable that correlates most strongly with sustained citation growth. Companies publishing irregularly or only on product launches under-index relative to peers.
The AI Companies AI Visibility Index 2026 is the first public two-wave benchmark measuring how often AI companies themselves — OpenAI, Anthropic, Google DeepMind, Meta AI, xAI, and 20 others — are surfaced, cited, and recommended inside ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews. It was produced by 5W, the AI Communications Firm.
Citation Share is the share of AI responses in which a company is named, recommended, or cited as a referenced source. Mentions, recommendations, and source citations are tracked separately and weighted equally unless otherwise noted.
OpenAI captures 24.6% of AI Citation Share across the five major AI assistants, followed by Anthropic at 14.8%, Google DeepMind at 13.7%, Meta AI at 9.7%, and xAI at 5.7%. The top five together account for 68.5% of cited responses in the dataset.
Anthropic. For prompts specifically about AI safety, alignment, interpretability, or responsible AI deployment, Anthropic surfaces at 31.2% — substantially above its overall 14.8% share. OpenAI surfaces at 19.8% in the same query set, Google DeepMind at 14.6%. Category specialization is the clearest brand-positioning signal in the dataset.
Self-citation bias measures how much more often an AI assistant recommends its parent company's models compared to how other AI assistants recommend the same models. In the dataset, ChatGPT recommended OpenAI 2.0x more often than other engines did. Gemini recommended Google DeepMind 1.7x more often. Claude recommended Anthropic only 1.2x more often — the lowest self-citation lift of any major AI engine tested.
Wikipedia (24.3%), GitHub (17.8%), and ArXiv (13.4%) together supply 55.5% of all AI-company-related AI citations. TechCrunch, The Verge, The Information, Wired, and the New York Times follow. AI-company-owned domains supply 8.6% in aggregate.
AI companies are uniquely cited through their code repositories and research papers. GitHub and ArXiv together account for 31.2% of citations in this category — a pattern that does not appear in banking, venture capital, or credit card categories 5W has measured. Technical content functions as primary source material for AI assistants describing AI companies.
Anthropic. Claude recommended Anthropic models only 1.2x more often than other engines recommended Anthropic — substantially lower than ChatGPT's 2.0x lift toward OpenAI or Gemini's 1.7x lift toward Google DeepMind.
Chinese AI labs (DeepSeek, Alibaba Qwen, Moonshot Kimi, Baidu ERNIE, Zhipu) register less than 2% combined Citation Share in U.S. consumer-facing queries. In technical and open-source-model queries specifically, DeepSeek surfaces in 18.4% of responses — third only to Meta and Mistral. The retrieval gap between consumer and technical queries is the largest of any company segment in the test set.
GEO — Generative Engine Optimization — is the practice of building brand authority and content infrastructure that AI assistants surface, cite, and recommend. It is the discipline replacing SEO in an AI-mediated discovery layer.
The Index tested 25 leading AI companies — including U.S. foundation model labs, specialized AI companies, open-source projects, and major Chinese AI labs — across 32,200 prompts in two independent waves between January and May 2026.
5W builds the citation infrastructure that puts AI companies inside the answers their buyers, recruits, partners, and regulators see first. Earned media. GEO. GitHub and ArXiv strategy. AI visibility measurement. One firm.