When buyers ask AI for "best X", vendors get 2% of citations

Across 810 AI answers analyzed, vendors received only 2% of citations when buyers asked category discovery questions like "best X for Y," with the other 98% going to third parties such as listicles, YouTube and other vendors' blogs, according to the study's findings. Vendor share of citations rose to 43–57% on specific factual questions (feature checks 57%, how-tos 52%, pricing 43%), while Reddit accounted for 22 of 6,375 citations (0.35%) and all 22 came from Gemini, with the ChatGPT and Perplexity APIs citing Reddit zero times. The data covers 27 companies including Stripe, Notion, Figma, HubSpot, Slack and Canva, and shows each engine draws on a distinct source mix — 92% of TechRadar citations came from ChatGPT and 65% of YouTube citations came from Perplexity.

What the data says 1. When a buyer asks "best X for Y", vendors get 2% of citations. The other 98% goes to third parties — mostly listicles, YouTube, and other vendors' blogs. 2. When a user asks a specific factual question, vendors get 43–57%. The questions you can win are the ones asked by people who already found you. 3. Reddit was 0.35% of all citations — and every single one came from Gemini. The ChatGPT and Perplexity APIs cited Reddit zero times in 810 answers. G2, Capterra and TrustRadius combined: 0–3%. 4. The 1 non-vendor source is YouTube. The 2 is Zapier's blog — cited 93 times in answers about Zapier's competitors. 5. Each engine has its own source diet. 92% of TechRadar citations came from ChatGPT. 65% of YouTube citations came from Perplexity. Which engine your buyers use changes which sources matter. 6. AI engines don't trash established brands. 742 brand mentions, zero warnings. What varies — from 10% to 80% — is whether you're framed as the answer or just an answer. 01Two kinds of questions, two different games Every question a user asks an AI assistant about software falls into one of two shapes, and the engines treat them almost as different products. Discovery questions — "best CRM for a small agency", "Notion vs Asana" — ask the engine to rank or choose. Technical questions — "does HubSpot have email sequences", "how do I set up Klaviyo flows", "is Typeform free" — ask it to verify a fact about a product the user can already name. Here is who gets cited, across 810 answers: | Question type | Vendor cited at least once | Vendor share of citations | Who wins instead | |---|---|---|---| | Category "best X for Y" | 39% | 2% | Listicles, YouTube, cross-vendor blogs | | Comparison "X vs Y" | 89% | 11% | Listicles and comparison posts | | Feature check "does X have Y?" | 96% | 57% | — | | How-to "how do I X" | 96% | 52% | — | | Pricing | 96% | 43% | — | Vendor share of cited sources, by question type Share of all citation URLs resolving to the vendor's own domain including docs. , help. , developer. The split is not subtle, and it is not about content quality. These are 27 companies with excellent websites — Stripe, Notion, Figma, HubSpot, Slack, Canva. When the question has a precise factual answer, engines go to the source: the vendor's own pages carry 43–57% of citations, and on how-to questions roughly a fifth of those vendor citations land on docs. and help. subdomains specifically. The mechanism behind the gap is more mundane than it looks, and worth stating plainly: an AI assistant answering a question runs a search, reads the top handful of results, and writes from those. When someone names your product, the search is your brand name — your site ranks first, gets read, gets cited. When someone asks for "the best CRM for agencies", the search is that phrase, and your product page does not rank for it. Round-ups do. The engine never sees your site, so it cannot cite it. There is no judgment being made about vendor bias; you simply aren't in the reading pile. The questions your website can win are the ones asked by people who already know you exist. The questions that create new customers are decided on pages you don't control. 02Three things everyone believes that the citations don't support Reddit accounted for 22 of 6,375 citations — 0.35% . More striking than the size is the distribution: all 22 came from Gemini. The ChatGPT and Perplexity APIs cited Reddit zero times across all 810 answers. One caveat we'll make for you before you make it: consumer ChatGPT, which has a content deal with Reddit, visibly surfaces Reddit threads in its app. Our probes ran against the APIs. If your buyers use the consumer apps, Reddit may matter more than this shows — but in the retrieval layer we could measure, it is close to absent, and two of three engines never touched it. G2, Capterra and TrustRadius combined accounted for 0–3% of citations per question type . For a category of sites whose entire pitch is being the trusted aggregation layer, they are strikingly absent from the answer layer. Part of this may be self-inflicted. When we ran our own crawler over these sites, g2.com's homepage was blocked by bot protection — the same class of defence that stops AI fetchers reading a page. Review aggregators may be under-cited partly because they've made themselves hard to read. Their scores may still matter indirectly, as inputs to the listicles that do get cited. True only for the questions where you're named. On category and comparison questions, roughly nine out of ten citations go to third parties. Your site's job on those questions is not to be the answer — it's to be accurately represented in the pages that are. 03Who actually wins the discovery questions If vendors get 2%, who gets the rest? We tallied every host cited on category and comparison questions. The leaderboard is not what the marketing playbooks predict. Most-cited non-vendor hosts — category + comparison questions Number of citations across all three engines. Green = another SaaS vendor's content marketing. Three things stand out. YouTube is the single biggest citation target. Video reviews and walkthrough content out-cite every text publication. Almost no SaaS content strategy treats YouTube as an AI-visibility channel. The second-biggest source is a competitor. Zapier's blog was cited 93 times — overwhelmingly in answers about other companies' categories. Monday.com, HubSpot and Salesforce show the same pattern at smaller scale. Zapier obviously publishes that content deliberately; what almost nobody does is treat it as an AI-citation strategy and measure it as one. The rest is the listicle economy — TechRadar, Forbes, and mid-size specialist reviewers like The Digital Project Manager and EmailToolTester. If your brand isn't in the round-ups these engines keep citing, you are not a candidate answer on discovery questions, regardless of how good your product page is. Except "the engines" is the wrong unit Aggregating three engines into one leaderboard hides the most practical finding in the dataset. Split the same citations by which engine produced them: Same hosts, split by engine Category + comparison citations. The aggregate leaderboard is mostly an average of three very different habits. "YouTube is the top source" is really Perplexity's top source — 65% of those citations are Perplexity's. The listicle economy is largely ChatGPT's house style: 92% of TechRadar citations and 74% of Forbes citations came from ChatGPT alone. And every Reddit citation in the study came from Gemini. Which means the practical advice depends on which engine your buyers use. If they're in Perplexity, video coverage is the lever. If they're in ChatGPT, it's the big tech round-ups. The one source that pays across all three is other vendors' blogs — Zapier's citations split almost evenly between Perplexity and Gemini, and it's the only host in the top five that isn't dominated by a single engine. 04The engines are not talking trash about anyone We also classified the tone of all 742 brand mentions the engines produced: is the brand framed as the leader, mentioned neutrally, hedged with a caveat, or warned against? | Tone | Mentions | Share | |---|---|---| | Leader — framed as a top choice | 340 | 45.8% | | Neutral listing | 369 | 49.7% | | Caveat "good, but…" | 28 | 3.8% | | Incidental | 5 | 0.7% | | Warning | 0 | 0% | Zero warnings in 742 mentions. For established SaaS, the fear that AI assistants are out there bad-mouthing your product is unsupported — the worst that happens is a mild caveat, and even the most-hedged brand in the sample Figma, at 17% caveats is hedged less than one time in five. What varies enormously is the leader framing . Typeform was presented as a top choice in 80% of its mentions; Figma in 10%. That spread doesn't track website quality, size, or product excellence — Figma's product is not worse than Typeform's. It tracks how tightly the training corpus binds the brand to a specific use case. "Typeform" means elegant forms in the corpus; "Figma" shares its category with a decade of design-tool discourse. That association was built over years of third-party writing, and no on-site change moves it quickly. 05The strangest result: famous, discussed, never cited One brand produced the study's cleanest anomaly. Notion was mentioned in 100% of its answers — and its website was cited in 0% of them. The engines know Notion the way people know Coca-Cola: as common knowledge that needs no reference. When they discuss it, they cite listicles about Notion rather than notion.so itself. Mention-without-citation matters commercially: the engine describes your product from third-party summaries and training memory, not from your pages. You don't control the description, and you don't get the visit. Being famous, it turns out, is not the same as being the source of record for yourself. What this study can and can't tell you - Sample: 27 tier-1, English-language SaaS brands Stripe, Notion, HubSpot, Figma, Slack, Canva, and 21 more , 10 queries each across 5 question types, against the ChatGPT, Gemini and Perplexity APIs. 810 answers, 6,375 cited URLs, 742 tone-classified mentions. - Retrieval was forced. Every answer here came from a search-grounded API call, so every answer had sources to cite. Consumer assistants often answer "best X for Y" straight from training memory with no search at all — in which case discovery is even more memory-driven than these citation shares suggest, not less. - APIs, not apps: the engines' APIs are documented to differ from their consumer apps in retrieval behaviour — the Reddit split above is a live example. Cross-engine, cross-query patterns are the reliable signal; any single cell is not. - A snapshot: retrieval pipelines change; treat these as measured patterns, not permanent laws. - Famous brands only: everything here describes products the engines already know. What happens to brands they've never heard of is a different — and harsher — story, which is the subject of our next report. 06What to do with this If you sell software, split your effort by question type. Your site can win feature, how-to and pricing questions — make those pages precise, current and quotable, including your docs and help subdomains, which carry a fifth of vendor citations on how-to answers. Your site cannot win "best X for Y", because it doesn't rank for those phrases and the engine therefore never reads it. For those, the work is off-site: the round-ups the engines already cite, video coverage if your buyers use Perplexity, and — the most under-measured lever in the dataset — authoritative content about your own category, the way Zapier's blog is cited into its competitors' conversations 93 times. And check where the sources you're chasing actually get cited. The data gives no support for Reddit-seeding as a citation strategy for two of the three engines, or for review-aggregator campaigns on any of them. They may have other value; being what AI quotes is not reliably it. Next report: the brands AI has never heard of Everything above describes the winners' bracket — brands every engine already knows. We re-ran the same style of probe on 51 obscure SaaS products fresh off Product Hunt, and the results are stranger: half were never mentioned by any engine on any question, despite being technically better-built than the visible half — and zero brand recognition meant zero mentions, not few. That report — including the visibility staircase from unknown to household name — is next in this series. Method in brief: each brand was probed with 10 queries spanning category discovery, feature verification, how-to, pricing and comparison intents, across the ChatGPT, Gemini and Perplexity search APIs 810 query-engine answers; per-type cells range n=27–81 . Every citation URL returned by an engine was resolved and classified by host; vendor citations include the brand's subdomains. Per-engine splits are computed over the same citation set. Tone classification of 742 brand mentions was performed by a separate model pass and spot-checked by hand. AIVerdict measures AI visibility for websites; this series publishes what we learn from the measurement infrastructure. Data tables available on request — \ email protected\ https://www.aiverdict.io/cdn-cgi/l/email-protection 42212d2c3623213602232b342730262b21366c2b2d .