Introduction
Citations are critically important to influencing AI search visibility. If your brand shows up in the sources of truth that LLMs refer to when conducting web searches to respond to a user’s query, it is more likely to be recommended in the LLM answer.
But "being cited by ChatGPT" is not one outcome. It is three, because the three ChatGPT access surfaces (free chat, paid chat, and the OpenAI API) cite materially different publisher mixes in response to the same prompt. New research from Petra Labs across 900 trials on a single model (GPT-5.2) shows the citation profile shifting in terms of which media types get credited depending on whether the prompt was asked in free chat, paid chat, or the OpenAI API. Read the full report here.¹
How Each Surface Decides What to Cite
Every ChatGPT response goes through two citation stages. First, the model consults sources during its research or “thinking” phase (interim citations). Then it selects a subset of those sources to include in its final response (final citations). The gap between the two is the model’s editorial filtering, and the three surfaces filter very differently.
In the Petra study, the paid chat surface cited an average of 39.4 sources during research and kept 6.3 in its final response, an 84% reduction. The free chat surface dropped from 25.3 to 6.7, a 73% reduction. The API dropped from 13.5 to 12.8, a 5% reduction. The API essentially does not filter. Nearly every source it reads becomes a citation in the final output. The chat surfaces, on the other hand, aggressively curate which URLs they actually show to the user. This filtering difference is the mechanism behind much of the publisher-level divergence that follows.
Earned vs Marketplace vs Social: How the Mix Shifts
Across 900 trials, every cited domain was classified into one of four media types: Earned (independent editorial), Marketplace (retailers and e-commerce), Social (community and user-generated platforms), or Owned (brand-operated domains). The distribution of these types in final citations differed sharply across surfaces.
Paid chat final citations: 66.9% Earned, 16.1% Marketplace, 8.5% Social. The most editorial-heavy mix of the three.
Free chat final citations: Earned-led, with a higher Social share (Reddit alone accounted for 13.8% of final citations).
API final citations: 46.8% Earned, 37.6% Marketplace. Combined Marketplace and Social citations (48.9%) nearly equaled Earned. The API’s final-citation mix was the most commerce-heavy of any surface.
These differences imply tradeoffs when developing GEO strategy. A PR team in the ski industry (the industry used for this experiment) placing coverage in earned editorial outlets is optimizing for the surface where 66.9% of citations are Earned (paid chat), and getting much less leverage on the surface where only 46.8% are (API). Conversely, a brand with strong retailer relationships is positioned well on the API and underweighted on the paid chat.
Why the API Is a Retail-Heavy Surface
The mechanism is visible in the search queries the API generates during its research phase. In the Petra study, the API averaged 4.08 search queries per trial with a 70% exact-duplicate rate. Its search phrasings followed rigid template constructions ("[Brand] [Model] [Year] review" for brand-specific queries, "best [category] [year] review" for broad ones). These templates retrieve retailer product pages and individual product review pages efficiently. They do not surface long roundup or editorial pieces the way longer natural-language queries do.
The chat surfaces, by contrast, averaged 14 to 15 word queries with high natural-language variety and low duplicate rates (5.6 to 9.5%). Their query shape retrieves roundup and editorial content much more effectively. The publisher mix difference follows from the search behavior difference, which follows from the access surface difference.
What This Means for PR and Earned Media Teams
Your citation strategy depends on which surface your audience uses. If you are pitching outlets and tracking coverage results, be explicit about which surface you are trying to influence. An earned media program that moves the needle on paid chat will not move the needle on the API in the same way.
Retailer and marketplace relationships are their own surface-specific lever. If a meaningful share of your category’s AI traffic comes through API-powered tools (agents, assistants, integrations), making sure your products appear correctly on major retail and marketplace sites is a separate citation strategy from earning editorial coverage.
Concentration levels differ by surface. The API’s top 10 domains accounted for 64.2% of all citations, the highest concentration of any surface. The paid chat was less top-heavy at 52.4%. A narrower set of publishers dominates API citations, which means the publisher tier that matters on that surface is smaller and harder to break into.
What Petra Does
Petra Labs tracks citation profiles across all three ChatGPT surfaces, plus Gemini, Claude, Google AI Overviews, and other LLMs, and separates earned, marketplace, social, and owned citations for each brand tracked. The 900-trial study that underpins this article documents why cross-surface citation tracking matters. If you want to see which publishers are driving AI citations for your brand, by surface, get in touch with us here.
Frequently Asked Questions
Q: Does ChatGPT cite the same publishers to everyone?
A: No. The free chat, paid chat, and OpenAI API surfaces of ChatGPT cite substantively different publisher mixes in response to the same prompt. Petra research found the paid chat’s final citations were 66.9% earned editorial, while the API’s were 46.8% earned and 37.6% retail and marketplace.¹
Q: Why does the API cite more retailer and marketplace sites than ChatGPT chat?
A: The API generates shorter, templated search queries ("[Brand] [Model] review") that retrieve product and retailer pages efficiently. The chat surfaces generate longer natural-language queries that retrieve roundup and editorial content. The difference in search behavior produces the difference in publisher mix.¹
Q: Do free ChatGPT and paid ChatGPT cite the same publishers?
A: They are closer to each other than either is to the API, but still not identical. In Petra research, the free chat surface had a higher Reddit share (13.8% of final citations) than the paid chat. The paid chat carried a slightly higher editorial-Earned concentration. Both are Earned-led, but with different specific publisher mixes within that.¹
Q: How many publishers does ChatGPT typically cite?
A: In the Petra study, the paid chat cited 6.3 publishers per final response on average, the free chat 6.7, and the API 12.8. Publishers that appear in the top 10 domains on a given surface account for roughly 52 to 64% of all citations on that surface.¹
Q: Should my PR strategy target specific AI surfaces?
A: Yes. If your category’s users predominantly query through the paid ChatGPT app, earned editorial coverage is the highest-leverage citation strategy. If your category is dominated by API-powered agents and integrations, retailer and marketplace presence matters more than editorial coverage. A credible AI-era PR plan treats the surfaces as distinct channels rather than as a single "ChatGPT" target.
Citations
1. Petra Labs. "How Access Surface Shapes LLM Search and Citation Behavior." April 2026. 900 trials across three GPT-5.2 access surfaces, conducted February 26, 2026.