I ran 100 B2B software search queries across five AI engines and recorded every source they cited. Five hundred data points later, the pattern is clear enough to change how you allocate budget. Review aggregators supplied 42% of all citations. Your own website supplied 35.4%. Which means the majority of what AI engines tell a prospect about your software comes from somewhere you do not control.
Key findings
- Review aggregators accounted for 42% of all AI citations. G2, Capterra, and Software Advice together supplied 210 of 500 cited sources.
- Vendor documentation accounted for 35.4% of citations. Company-owned docs, help centers, and product pages supplied 177 of 500.
- Community platforms accounted for 11.6%. Reddit, Quora, and industry forums supplied 58 of 500, concentrated in problem-aware and troubleshooting queries.
- Independent editorial accounted for 11.0%. Tech journalism, niche blogs, and expert reviews supplied 55 of 500, concentrated in comparative and trend queries.
- 88.6% of citations came from domains with authority above 80. That is 443 of 500. Low-authority domains were effectively filtered out regardless of content quality.
- Aggregators held the first-cited position in 54% of the responses they appeared in. Vendor docs held it in 28%. Aggregators validate. Vendor docs confirm.
The interpretation in one sentence: third-party authority now outweighs on-site marketing claims during the evaluation phase, and the gap is not close.
Introduction: The new dark social of search
For roughly two decades, B2B software evaluation ran through a predictable funnel. A buyer searched, clicked a few blue links, landed on vendor sites, compared, and converted. Every step of that journey was observable in analytics.
That funnel now has a segment you cannot see. A prospect asks ChatGPT which quoting tool handles commission splits for manufacturer’s reps. The engine returns three named products with a short justification for each. The prospect forms an opinion, eliminates two vendors, and only then visits a website. By the time they show up in your analytics, the comparison already happened somewhere you were not present and cannot audit.
This is the thesis of the study: visibility in answer engines is no longer deterministic. Ranking a landing page for “best quoting software” does not guarantee inclusion in the AI comparison table for that same query. Inclusion depends on third-party aggregators, entity associations across the knowledge graph, domain authority thresholds, and platform-specific retrieval behavior that varies more between engines than most practitioners assume.
I wanted to know what actually drives that inclusion. So I measured it.
Methodology: How I tracked 500 AI citations
- Sample: 100 B2B software search queries, each executed across 5 AI platforms, producing 500 total citation data points.
- Platforms tested: ChatGPT, Gemini, Perplexity, Copilot, and Claude.
- Software categories covered: quote automation, project management, billing, inventory control, and helpdesk systems.
- Query composition: high-intent commercial investigation queries (“best X for Y”), technical and integration queries (implementation questions, API and specification checks), troubleshooting queries, and comparative best-of queries.
- What was recorded per data point: the source cited, its category (aggregator, vendor documentation, community, editorial), its position in the citation order, the domain authority of the citing source, and the last-updated date of the cited content where available.
- Terminology used throughout: a citation is any source the engine surfaced in support of its answer. First-cited position refers to the source appearing first in the citation order, which is a separate measurement from whether a source was cited at all.
This kind of measurement is possible because of how these systems are built. Retrieval-augmented language models condition their output on retrieved documents and expose a source attribution mechanism as a byproduct of that architecture (Ram et al., 2023).
The citations are not decoration. They are a visible trace of the retrieval step, which means citation behavior can be observed systematically rather than inferred.
Finding #1: Review aggregators dominate AI citations
Review aggregators accounted for 42% of all citations: 210 of 500. G2, Capterra, and Software Advice were the three dominant sources. TrustRadius appeared but at lower frequency.
Vendor documentation was the second pillar at 35.4%: 177 of 500. Together, those two categories supplied 387 of 500 citations, or 77.4% of everything the engines surfaced.
The practical reading is uncomfortable for most SaaS marketing teams. AI engines cross-reference your brand against industry aggregators rather than accepting your site copy at face value. Your product page can be well-optimized, fast, and perfectly written, and it will still lose the first citation slot to a G2 category page assembled from customer reviews you had no hand in writing.
| Source type | Citations | Share |
|---|---|---|
| Review aggregators | 210 | 42.0% |
| Vendor documentation | 177 | 35.4% |
| Community platforms | 58 | 11.6% |
| Independent editorial | 55 | 11.0% |
| Total | 500 | 100% |
Finding #2: How citations fracture across the long tail
The remaining 22.6% splits almost evenly between two source types that behave very differently.
Community platforms accounted for 11.6% of citations: 58 of 500. Reddit, Quora, and industry-specific forums. These citations concentrated in problem-aware and troubleshooting queries rather than category evaluation. When a query described a symptom rather than a category, community sources appeared. When a query asked for a recommendation, they mostly did not.
Independent editorial accounted for 11.0% of citations: 55 of 500. Tech journalism, niche industry blogs, and expert review sites. These surfaced in comparative best-of queries and trend questions, where the engine appeared to want a synthesizing voice rather than a data source.
Neither category is large enough to build a strategy around alone. Both are large enough that ignoring them costs you roughly a fifth of the available citation surface.
Finding #3: Every AI platform cites differently
This was the most surprising result, and the one with the most direct strategic consequence: the five engines do not behave like five instances of the same system.
Perplexity and Copilot cited review aggregators and vendor documentation almost exclusively. Community and editorial sources were rare in both. If your visibility strategy is aggregator-first, these two engines reward it most directly.
Gemini and Claude were the only engines that surfaced Reddit and community forums with any regularity. The other three effectively did not.
Claude was the only platform that heavily cited competitor-owned content marketing, with 24 instances across 100 responses. That is a distinct behavior no other engine displayed at anything close to that rate. Claude repeatedly cited a competitor’s blog post or comparison page as a source when answering a category question. If a rival publishes the definitive roundup in your category, Claude is disproportionately likely to hand them the citation.
| Source type | Perplexity | Copilot | Gemini | Claude | ChatGPT |
|---|---|---|---|---|---|
| Review aggregators | Heavy | Heavy | Moderate | Moderate | Heavy |
| Vendor documentation | Heavy | Heavy | Moderate | Moderate | Heavy |
| Community/forums | Rare | Rare | Present | Present | Rare |
| Competitor content marketing | Rare | Rare | Rare | Heavy (24) | Rare |
The takeaway is that “AI visibility” is not one problem. It is five problems that partially overlap, and a strategy optimized for Perplexity will underperform on Gemini.
Finding #4: Source choice depends on search intent
Query intent predicted source type more reliably than software category did.
On commercial investigation queries, aggregators supplied 68.0% of category-level citations: 136 of 200. These are the “best X for Y” queries. The engine appears to want consensus, and a review platform aggregating hundreds of verified user opinions is the closest available proxy for consensus. A vendor claiming to be the best is, structurally, one opinion.
On technical and integration queries, vendor documentation supplied 72.5% of citations: 116 of 160. Implementation questions, API behavior, specification checks. Here, the engine wants the authoritative primary source, and only the vendor has it. No aggregator can tell you the rate limit on an endpoint.
| Query intent | Data points | Dominant source | Share |
|---|---|---|---|
| Commercial investigation | 200 | Review aggregators | 68.0% (136) |
| Technical / integration | 160 | Vendor documentation | 72.5% (116) |
The split is clean enough to act on. Aggregators win evaluation. Vendor docs win execution. If your documentation is thin, you are invisible on exactly the queries where a buyer is closest to implementation.
Finding #5: Citation placement and who gets the first link
Being cited and being cited first are different outcomes with different value.
Aggregators held the first-cited position in 54% of the responses in which they appeared. While Vendor documentation held the first-cited position in only 28% of its appearances.
The pattern that emerges is a two-tier trust structure. The AI engine leads with an aggregator to establish that a product is real, credible, and used by other people. It then reaches for vendor documentation to fill in specifics. Aggregators validate first. Vendor docs confirm details.
That ordering matters because first-cited sources carry more weight with a reader skimming an answer, and because in several interfaces the first citation is the one displayed most prominently. Winning a secondary citation slot is worth something. It is not worth what the primary slot is worth.
Finding #6: Domain authority and freshness operate as hard filters
Two variables behaved less like ranking factors and more like eligibility gates.
88.6% of all citations came from domains with authority above 80: 443 of 500. Only 57 citations across the entire dataset came from domains below that threshold. This was consistent across all five platforms and all five software categories. A well-written page on a domain with authority in the 40s was, in this dataset, not competitive for citation regardless of how directly it answered the query.
Content updated within the prior 12 months drove 74.2% of direct vendor citations. Freshness mattered asymmetrically. Vendor documentation was heavily filtered by recency. Aggregator pages were not, and older evergreen category pages on G2 and Capterra were cited routinely. The likely reason is that review volume and recency of individual reviews function as a freshness signal independent of the page’s last-modified date.
| Filter | Threshold | Effect |
|---|---|---|
| Domain authority | Above 80 | 88.6% of citations (443/500) |
| Content freshness (vendor docs) | Updated within 12 months | 74.2% of vendor citations |
| Content freshness (aggregators) | No strong effect | Older evergreen pages cited routinely |
Neither of these looks like a scoring factor you can offset with better content. They look like gates you either clear or you don’t.
Why this matters: The shift in the buyer journey
Consider a specific case. A construction manufacturer’s rep agency needs software that handles quoting, order tracking, and commission splits. That is a narrow requirement in a narrow vertical, and historically the buyer would have run three or four searches, opened a dozen tabs, and worked through vendor sites.
Now they ask an AI engine. The engine returns three products. Based on this dataset, roughly 42% of what it cites to justify those three will come from G2 or Capterra, and the source it leads with will be an aggregator more than half the time.
If your product is not on those aggregators with sufficient review volume, you are not in the answer. Not ranked lower in the answer. Absent from it.
That is the end of blue link dominance in a practical sense. The engines cross-reference brand presence against aggregators and trade media rather than accepting site copy. Because retrieval-augmented models ground their responses in retrieved documents and attribute those sources (Ram et al., 2023), what gets retrieved determines what gets said about you, and most of what gets retrieved is not yours.
The domain authority pattern points the same direction. Our finding that established, high-authority domains supplied 88.6% of citations aligns with research showing that retrieval systems produce more trustworthy output when they weight source reliability rather than relevance alone (Ni et al., 2025). What we observed empirically across five engines is consistent with those systems applying reliability priors during retrieval. If that reading is right, authority is not a tiebreaker in AI citation. It is closer to an entry requirement.
The funnel consequence is that SaaS companies need an off-site visibility strategy with the same rigor previously reserved for on-site SEO. Third-party authority now outweighs on-site marketing claims during evaluation. A budget allocation that puts 90% into owned content and 10% into review generation is inverted relative to where the citations actually come from.
The operator playbook: How to get cited by AI engines
Four moves, ordered by the size of the citation surface they address.
1. Build an aggregator review pipeline. With G2, Capterra, and Software Advice driving 42% of citations and holding the first-cited slot 54% of the time, review generation stops being a nice-to-have and becomes a visibility requirement. Set a monthly review target, build the ask into your customer success motion, and treat category page presence on the major aggregators as infrastructure rather than marketing.
2. Audit vendor documentation for machine readability. Vendor docs win 72.5% of technical queries, and those are the queries closest to implementation. Structure documentation with clear, crawlable hierarchies, answer specific workflow questions directly, and keep it current, since 74.2% of vendor citations came from content updated within 12 months. Language models recall facts more reliably when the underlying knowledge is explicit and structured (Yang et al., 2024), which means the way your documentation is organized affects whether your product gets described accurately.
3. Optimize for entity density rather than keyword density. Make sure AI models can contextually link your software to the core concepts in your category across the knowledge graph. This recommendation reflects established research showing that knowledge graphs supply the structured factual grounding language models rely on for accurate inference (Pan et al., 2024). Practically: name the concepts, the integrations, the use cases, and the adjacent categories explicitly rather than assuming the connection is obvious. Learn more about entity SEO optimization.
4. Tailor tactics to platform archetypes. The data supports different plays for different engines. For Claude, industry listicles and competitor roundups matter disproportionately, since it cited competitor content marketing 24 times across 100 responses. For Gemini, community forum presence and trade publication coverage carry weight the other engines do not reward. For Perplexity and Copilot, the aggregator and documentation strategy is close to the whole game.
Methodology, limitations, and how to cite this study
Full methodology
One hundred B2B software search queries were executed across ChatGPT, Gemini, Perplexity, Copilot, and Claude, producing 500 citation data points. Queries spanned five software categories: quote automation, project management, billing, inventory control, and helpdesk systems. Query intent types included commercial investigation, technical and integration, troubleshooting, and comparative best-of.
For each response, the following were recorded: every source cited, its category classification, its position in the citation order, the domain authority of the cited domain, and the last-updated date of the cited content where that was available.
Limitations
I want to be clear about what this study does and does not establish.
- Sample size. Five hundred data points across 100 queries is enough to identify strong patterns. It is not enough to make precise claims about small differences between platforms, and I have avoided doing so.
- Category scope. All queries were B2B software. These findings should not be extrapolated to consumer categories, local services, healthcare, finance, or any vertical where the available sources differ structurally. There is no G2 for orthodontists.
- Point-in-time snapshot. AI retrieval behavior changes without announcement. This dataset reflects the platforms as they behaved during the study period, not a stable property of these systems.
- Intent classification. The commercial and technical intent buckets together cover 360 of the 500 data points. The remaining 140 spanned troubleshooting and comparative queries and were not large enough per bucket to report reliable percentages.
- Domain authority as a proxy. Domain authority is a third-party metric, not a signal the engines expose. The correlation observed here is consistent with reliability weighting during retrieval, but this study cannot establish causation.
Now It’s Your Turn
I hope this gave you something concrete to work with, whether that is a reason to finally build the review pipeline you have been postponing, an argument for documentation budget, or a platform-specific play you had not considered.
So I have to ask, which finding changes what you do next? Is it the 42% coming from aggregators, the fact that domain authority operates as a hard filter at 88.6%, or the discovery that Claude is handing citations to your competitors’ content marketing?
References
Ni, B., Liu, Z., Wang, J., Deng, Y., Liu, X., Wang, H., Yang, Y., Zhou, X., Liu, X., Chen, W., Yao, H., & Guo, Q. (2025). Towards trustworthy retrieval-augmented generation for large language models: A survey. ACM Computing Surveys. https://doi.org/10.1145/3837074
Pan, S., Luo, L., Wang, Y., Chen, C., Wang, J., & Wu, X. (2024). Unifying large language models and knowledge graphs: A roadmap. IEEE Transactions on Knowledge and Data Engineering, 36(7), 3580–3599. https://doi.org/10.1109/TKDE.2024.3352100
Ram, O., Levine, Y., Dalmedigos, I., Muhlgay, D., Shashua, A., Leyton-Brown, K., & Shoham, Y. (2023). In-context retrieval-augmented language models. Transactions of the Association for Computational Linguistics, 11, 1316–1331. https://doi.org/10.1162/tacl_a_00605
Yang, L., Chen, H., Li, Z., Ding, X., & Wu, X. (2024). Give us the facts: Enhancing large language models with knowledge graphs for fact-aware language modeling. IEEE Transactions on Knowledge and Data Engineering, 36(7), 3091–3110. https://doi.org/10.1109/TKDE.2024.3360454