Citation share by source type
Share of all citations by the type of site cited.
| Source type | Four AI engines (of 2,594 citations) | Google AI Overviews (of 129 citations) |
| Other / unclassified | 999 (38.5%) | 54 (41.9%) |
| Company / brand sites | 687 (26.5%) | 13 (10.1%) |
| Research / data providers | 198 (7.6%) | 3 (2.3%) |
| General news / business press | 170 (6.6%) | 7 (5.4%) |
| Community / social | 131 (5.1%) | 41 (31.8%) |
| Trade / specialist press | 125 (4.8%) | 4 (3.1%) |
| Directories / rankings | 80 (3.1%) | 0 (0%) |
| Official / government | 79 (3%) | 6 (4.7%) |
| Reference | 61 (2.4%) | 1 (0.8%) |
| Review / comparison publishers | 47 (1.8%) | 0 (0%) |
| Institutions / trade bodies | 17 (0.7%) | 0 (0%) |
How can AI infrastructure companies improve their AI visibility?
These steps follow from the measured results above. They describe what the engines cited in this study and are not a guarantee of results.
| Step | What this study measured |
| Put your facts on your own domain. | Amazon's own domains were cited in 9.4% of grounded answers, Microsoft's in 8.3% and Oracle's in 6.9%. |
| Keep your facts accurate on the sites engines cite. | The most-cited sites other than the tracked names' own were runpod.io (in 14.2% of grounded answers), northflank.com (12.8%), nvidia.com (11.1%), en.wikipedia.org (10.4%) and gmicloud.ai (10.1%). Research and data providers supplied 7.6% of citations and general news and business press 6.6%. |
| Check each engine separately. | ChatGPT named Microsoft in 81.9% of answers against 56.9% on Perplexity. Perplexity named Nebius in 25.0% of answers against 1.4% on ChatGPT. |
| Watch names outside the tracked list. | NVIDIA Corporation (31 answers), Vast.ai (24 answers), RunPod (23 answers), Vantage Data Centers (19 answers), OVHcloud (18 answers) were named often but were not tracked. |
| Re-run the same questions every month. | Compare counts and intervals, not ranks. The 24 questions are in the appendix. |
What do these results mean for AI infrastructure companies?
The results describe two different contests. The first is being named. Google, Amazon and Microsoft each appeared in roughly 70% of answers, CoreWeave and Oracle in about four in ten, and every company after them appeared in fewer than three in ten. Because the top three intervals overlap, the honest reading is a leading group of three, not a single winner.
The second contest is being cited. Naming and citing do not move together. Meta was named in 26.7% of answers, yet its own domains were cited in 1 of 288 grounded answers. Google was named in 71.5% of answers, with its own domains cited in 6.3%. Amazon had the highest own-domain rate at 9.4%. A company can be well known to an engine and still have little of its own material used as a source.
Outside sources also carry real weight. The most-cited sites in the four-engine set were runpod.io, northflank.com, nvidia.com, en.wikipedia.org and gmicloud.ai, and company and brand sites made up 26.5% of citations. In Google AI Overviews, community and social sources supplied 31.8% of citations, with youtube.com the most-cited source.
For monthly tracking, treat a change in rank as noise unless the count moves outside the earlier interval. The limits section below lists what this study did not score, including frequently named companies that were not on the tracked list.