Why we ran this study
If your brand isn’t showing up when someone asks ChatGPT for a recommendation, you have a visibility problem that traditional SEO can’t fix on its own.
We wanted to know: when a potential customer asks an AI assistant a buying-intent question — “best dentist in Los Angeles,” “top CRM for small businesses,” “how to recover from a Google penalty” — which websites actually get cited? Are the same sites winning across every model, or does each AI have its own favorites? And most importantly, what do cited pages have in common that uncited pages don’t?
So we ran the study ourselves.
What we did
Between May and June 2026, we submitted 200 identical questions to five AI assistants:
- ChatGPT (GPT-4o with browsing)
- Google Gemini (with grounding via Google Search)
- Perplexity (Pro mode)
- Claude (with web search enabled)
- Microsoft Copilot (with web access)
The 200 questions spanned 10 industries — dental, legal, HVAC, SaaS, cybersecurity, real estate, accounting, fitness, home services, and e-commerce — with 20 questions per vertical. Every question was phrased as a real buyer would ask it: “Who’s the best personal injury lawyer in Phoenix?” not “list personal injury law firms.”
For each response, we logged:
- Every domain cited (URL-level)
- Position of the citation (first cited, second, third, etc.)
- Whether the citation was a direct link, inline mention, or footnote
- Page-level attributes of each cited URL (domain rating, word count, schema markup types, publish date, whether the page contained original data or research)
That gave us 1,000 total AI responses and 3,217 individual citations to analyze.
Finding 1: The overlap is smaller than you’d think
The assumption many marketers carry is that if you rank well on Google, AI will cite you too. The data tells a different story.
Of the 847 unique domains cited across all five models, only 94 domains (11.1%) were cited by three or more AI assistants for the same query. Just 23 domains (2.7%) were cited by all five.
The vast majority of citations — 68% — appeared in only one model’s response.
| Cited by | Domains | % of total |
|---|---|---|
| All 5 models | 23 | 2.7% |
| 4 models | 31 | 3.7% |
| 3 models | 40 | 4.7% |
| 2 models | 178 | 21.0% |
| 1 model only | 575 | 67.9% |
What this means in practice: optimizing for one AI assistant does not guarantee visibility in the others. Each model has its own retrieval behavior, its own source preferences, and its own trust signals.
Finding 2: Perplexity cites the most sources, but ChatGPT’s citations drive the most traffic
Across our 200 queries, the average number of cited sources per response varied significantly by model:
| AI assistant | Avg. citations per response | Median | Citation style |
|---|---|---|---|
| Perplexity | 8.4 | 9 | Numbered footnotes with inline links |
| Gemini | 4.1 | 4 | Inline carousel cards |
| Claude | 3.7 | 3 | Inline text citations |
| ChatGPT | 3.2 | 3 | Inline links in prose |
| Copilot | 2.9 | 3 | Numbered superscripts |
Perplexity cites roughly 2.6x more sources per answer than ChatGPT. But when we cross-referenced with available clickstream data and referral patterns from client analytics, ChatGPT-sourced traffic had a 34% higher engagement rate (pages per session) than Perplexity referrals.
The likely explanation: Perplexity users treat it as a research aggregator and bounce between sources. ChatGPT users tend to visit fewer links but engage more deeply when they do.
Finding 3: The “citation class” — what winning pages have in common
We pulled 15 page-level attributes for every cited URL and compared them against a control group of 300 pages that ranked in Google’s top 10 for the same queries but were never cited by any AI assistant.
Seven attributes showed a statistically significant difference between cited and uncited pages:
| Attribute | AI-cited pages (median) | Uncited top-10 pages (median) | Difference |
|---|---|---|---|
| Word count | 2,140 | 1,380 | +55% |
| Domain Rating (Ahrefs) | 62 | 54 | +15% |
| Schema types present | 3.2 | 1.4 | +129% |
| External citations on page | 6.1 | 1.8 | +239% |
| Contains original data/stats | 41% | 9% | +356% |
| Published or updated in last 12 months | 73% | 48% | +52% |
| Has FAQ or Q&A structured data | 38% | 11% | +245% |
Two findings stand out here.
Pages that cite external sources get cited themselves. Pages with six or more outbound references to authoritative external sources were 3.4x more likely to be cited by AI than pages with zero or one. AI models appear to treat outbound linking as a credibility signal — a page that substantiates its claims with references looks more trustworthy to a retrieval system than one that asserts without evidence.
Original data is the single strongest predictor. Pages containing original statistics, survey results, proprietary benchmarks, or first-party case study data were cited at 4.5x the rate of pages without it. This is the clearest signal in our dataset — and it’s the same core principle behind Generative Engine Optimization: if you want AI to cite you, create something it can’t find anywhere else. See our guide on how to get cited by ChatGPT, Perplexity, and AI Overviews for the tactical follow-through on both of these.
Finding 4: Wikipedia and Reddit are not your competition — niche authority sites are
Before running this study, we expected Wikipedia and Reddit to dominate AI citations the way they dominate traditional search. They didn’t.
Wikipedia appeared in 14.2% of all responses, which is significant but far from dominant. Reddit appeared in just 6.8%. The most-cited domain category was niche industry publishers and vertical SaaS blogs — sites like Dentistry Today, LawSites, HVAC School, and individual company blogs with deep topical authority.
| Domain category | % of all citations |
|---|---|
| Niche industry publishers | 31.4% |
| Company blogs (non-aggregator) | 22.7% |
| Wikipedia | 14.2% |
| News outlets | 11.3% |
| Government / .edu / .org | 8.6% |
| Reddit and forums | 6.8% |
| Aggregator/directory sites | 5.0% |
This is the most actionable finding for small and mid-size businesses: you are not competing with Wikipedia. You are competing with the three or four niche-authority sites in your industry that have built topical depth and earn AI citations as a result.
Finding 5: Local queries are the wild west
For local-intent queries (“best dentist in [city],” “top-rated HVAC company near me”), AI citation behavior was the most inconsistent across models — and the most prone to errors.
- 27% of local citations pointed to businesses that had closed, moved, or changed names
- Gemini was the most accurate for local queries (likely due to Google Business Profile integration), citing a verified active business 89% of the time
- ChatGPT was the least reliable for local, with 31% of its local citations pointing to outdated or incorrect information
- No model consistently cited Google Business Profile URLs; most cited the business’s website or a directory listing
For local service businesses, this means your website — not just your Google Business Profile — needs to be the definitive, current source of truth about your business. AI models are pulling from your site, directory listings, and third-party mentions, and when those sources conflict, the citation goes to whoever has the clearest, most recently updated information. Our own local SEO in Los Angeles research runs into the same fragmentation problem at the map-pack level — see our Google Business Profile checklist for closing the entity-data gaps that cause this.
Finding 6: The freshness window is narrower than you think
We analyzed the publish or last-updated date of every cited page. The results suggest AI models have a strong recency bias — stronger than Google’s organic algorithm.
- 73% of all cited pages were published or substantively updated within the last 12 months
- Pages updated within the last 90 days were cited 2.1x more often than pages updated 6–12 months ago
- Pages older than 24 months with no updates represented only 8% of citations — and those were almost exclusively Wikipedia or government sources
The implication is clear: content decay hits harder in AI visibility than in traditional SEO. A page that ranked well for years on Google can silently disappear from AI responses simply because it hasn’t been updated.
What this means for your AI visibility strategy
The data points to a framework we’re now using with every Gobiya client:
1. Treat each AI model as a separate channel. Citation behavior differs enough across models that a one-size-fits-all approach leaves gaps. At minimum, monitor your brand’s visibility in ChatGPT, Perplexity, and Gemini independently — see our AI visibility overview for how we track this.
2. Create citeable assets, not just rankable content. The single biggest differentiator between cited and uncited pages is the presence of original data, proprietary statistics, or first-party research. If your page contains nothing that can’t be found on three other sites, AI has no reason to cite you specifically — this is the core of our GEO & AI content writing work.
3. Cite others to be cited. Pages that substantiate claims with outbound links to authoritative sources are cited significantly more often. This mirrors academic publishing — a well-referenced paper signals rigor. The same logic applies to AI retrieval systems.
4. Update aggressively. The freshness window for AI citation is 6–12 months, not the 2–3 years many teams assume for evergreen SEO content. Build a quarterly update cadence into every pillar page — our content strategy service is built around exactly this cadence.
5. Fix your entity data. For local businesses especially, conflicting information across your website, Google Business Profile, directory listings, and social profiles creates ambiguity that AI models resolve by citing someone else. Audit and align your entity data across every surface, starting with your Google Business Profile and citations.
6. Invest in structured data. Pages with three or more schema types (FAQ, HowTo, LocalBusiness, Article, etc.) were cited at more than double the rate of pages with minimal or no schema. Structured data helps AI models understand what your page is and when to retrieve it — part of the technical SEO foundation that makes everything else in this list actually crawlable.
Methodology notes
All queries were submitted between May 12 and June 20, 2026. Each query was submitted in a fresh session with no prior context to avoid personalization effects. Responses were captured via screenshot and manual logging. Domain ratings were pulled from Ahrefs as of June 2026. Schema analysis was performed using Google’s Rich Results Test and manual inspection. Statistical significance was assessed using chi-squared tests for categorical variables and Mann-Whitney U tests for continuous variables, with a threshold of p < 0.05.
The full dataset is available on request for journalists and researchers — contact hello@gobiya.com.