AI Citation Sources Study Q3 2026: ChatGPT and Perplexity
We ran 200 buyer prompts through ChatGPT and Perplexity in August and September 2026: the domains they cited, UGC vs brand vs publisher share, and shifts.
ChatGPT
Claude
Perplexity
GrokOn this page
- How we ran the study, and its limits
- The domains ChatGPT and Perplexity cite most
- Cited sources by category
- How much comes from UGC, brands and publishers
- What changed between August and September
- How this compares with other published figures
- What to do with it
- Build your own map now and look again next month
This AI citation sources study ran the same 200 buyer prompts through ChatGPT and Perplexity on August 9 and September 26, 2026. Publishers earned the largest share of citations on both engines in both months. Between the two runs, Reddit fell from 13.5% of Perplexity answers to 1.5%, and ChatGPT stopped citing Forbes in our sample.
Those shifts are the reason for the index. Advice such as "get on Reddit" assumes the source mix is stable. In our data it differs by engine, by category and by month, and a plan built on last quarter's winner can miss this quarter's.
How we ran the study, and its limits
We sent 200 buyer-style prompts, once each, through ChatGPT and Perplexity with web search enabled, over their APIs, in August and again in September, and counted the sources each answer cited in its text.
| Design element | Detail |
|---|---|
| Prompts | 200 buyer questions across 20 product categories, 10 each: "best X", "X vs Y", alternatives, budget and use-case questions |
| Categories | 11 B2B software, 4 consumer software and services, 5 consumer products. AI visibility tools excluded, since we sell one |
| Engines | ChatGPT and Perplexity, each through its API with web search enabled |
| Runs | One run per prompt per engine per month: 400 answers in August, 400 in September |
| Dates | August 9 and September 26, 2026, identical prompts and settings |
| Cited source | ChatGPT's inline citations; for Perplexity, sources referenced by a numbered marker in the answer text |
| Domain types | We hand-labeled the 300 most-cited domains, which cover 82.5% of cited sources; the rest are reported as unclassified |
Read the results with five limits in mind:
- The runs went through each engine's API, not its consumer app. Answers in the app, with memory and location, can differ.
- One run per prompt per month supports statements about the whole corpus, not about any single brand or prompt. Use the numbers as context for category-level planning; to learn how often a specific brand or page gets cited, track your own prompts with repeated runs.
- ChatGPT cited at least one source in only 39% (August) and 40% (September) of its answers, so its domain figures rest on about 300 citations a month.
- Google AI Overviews, AI Mode, Gemini, Copilot and Claude are outside this study. We chose ChatGPT and Perplexity because both return their cited sources through their APIs with web search on, which lets us re-run an identical prompt set each month and compare like with like; Google's AI surfaces need a different collection method.
- Domain types are our judgement, documented below. Another analyst could label a marginal site differently.
Two definitions matter. Perplexity's API returns more sources than its answer text uses. We count only sources the text cites, and we report the "returned" figure separately where it helps. Our August write-up counted returned sources, which is why the stats page reports Reddit in 77% of Perplexity answers while this index reports 13.5% for the same run.
The domains ChatGPT and Perplexity cite most
Review and technology publishers lead on both engines: TechRadar, Tom's Guide and G2 appear near the top of each, and Forbes and PCMag dominate Perplexity. The two lists still share little, as each engine favors a different set of sites.
ChatGPT, share of answers citing each domain:
| Domain | August | September |
|---|---|---|
| techradar.com | 12.0% | 10.0% |
| wired.com | 2.5% | 7.0% |
| g2.com | 6.5% | 5.0% |
| tomsguide.com | 6.0% | 5.0% |
| softwareadvice.com | 1.5% | 2.5% |
| capterra.com | 0.5% | 2.5% |
| zapier.com | 4.5% | 2.5% |
| techrepublic.com | 2.0% | 2.5% |
| wikipedia.org | 1.5% | 2.0% |
| forbes.com | 11.5% | 0.0% |
Perplexity, share of answers citing each domain:
| Domain | August | September |
|---|---|---|
| forbes.com | 27.5% | 47.0% |
| pcmag.com | 26.0% | 41.5% |
| techradar.com | 22.0% | 37.5% |
| g2.com | 10.5% | 26.0% |
| zapier.com | 21.0% | 25.5% |
| nytimes.com | 9.5% | 19.5% |
| tomsguide.com | 5.5% | 17.5% |
| cnet.com | 8.0% | 14.0% |
| reddit.com | 13.5% | 1.5% |
| youtube.com | 20.0% | 0.0% |
Perplexity's September answers cited more sources (median 10 per answer, against 6 in August) at the same length of about 250 words, so most domains' answer shares rose with it. Compare Perplexity's domains within a month, or use the type shares below, which account for the change.
On the same prompt, the two engines agreed rarely. ChatGPT and Perplexity cited no domain in common on 79.5% of prompts in August and 74.0% in September.
Cited sources by category
Consumer product questions lean hardest on publishers, while B2B software questions draw more on brand-owned sites and review platforms. The top domains change from one group to the next.
| Category group | Engine | Top domains, September (share of answers) | Publisher share | Brand-owned share | Review platforms |
|---|---|---|---|---|---|
| B2B software (110 prompts) | ChatGPT | techradar.com 12.7%, g2.com 9.1%, softwareadvice.com 4.5% | 26.9% | 29.1% | 10.8% |
| B2B software | Perplexity | forbes.com 66.4%, pcmag.com 46.4%, g2.com 41.8% | 41.5% | 36.7% | 8.7% |
| Consumer software (40 prompts) | ChatGPT | wired.com 12.5%, tomsguide.com 12.5%, techradar.com 7.5% | 53.1% | 10.2% | 2.0% |
| Consumer software | Perplexity | pcmag.com 75.0%, techradar.com 57.5%, cyberinsider.com 42.5% | 72.2% | 13.0% | 2.4% |
| Consumer products (50 prompts) | ChatGPT | wired.com 16.0%, goodhousekeeping.com 6.0%, bonappetit.com 6.0% | 73.0% | 1.6% | 0.0% |
| Consumer products | Perplexity | nytimes.com 46.0%, businessinsider.com 38.0%, forbes.com 34.0% | 67.2% | 3.5% | 0.4% |
Shares are of all cited sources in that group and month. The rest of each row is UGC, reference sites and unclassified domains; ChatGPT's small consumer groups (under 65 citations each) carry the widest error.
How much comes from UGC, brands and publishers
Publishers took 41% to 54% of citations, brand-owned sites about a fifth to a quarter, and user-generated content almost nothing in September. Review platforms such as G2 and Capterra added 3% to 7%.
| Source type | ChatGPT Aug | ChatGPT Sep | Perplexity Aug | Perplexity Sep |
|---|---|---|---|---|
| Publishers and independent review sites | 45.9% | 40.9% | 50.9% | 53.9% |
| Brand-owned: vendors in the prompt's category | 14.3% | 11.7% | 4.5% | 6.1% |
| Brand-owned: other companies | 12.2% | 8.4% | 18.3% | 18.0% |
| Review platforms | 6.8% | 7.0% | 2.8% | 5.5% |
| UGC (Reddit, YouTube, LinkedIn and similar) | 0.0% | 1.0% | 6.8% | 0.2% |
| Reference (Wikipedia, .gov, .edu) | 1.7% | 1.7% | 0.2% | 0.1% |
| Unclassified long tail | 19.0% | 29.2% | 16.3% | 16.2% |
| Cited sources counted | 294 | 298 | 1,200 | 2,127 |
Two rows deserve a closer look. "Other companies" is mostly content marketing from businesses outside the category, such as Zapier's software roundups, and it takes about 18% of Perplexity's citations. Category vendors' own sites earn more share in ChatGPT (12% to 14%) than in Perplexity (5% to 6%), so a vendor's own pages count for more in one engine than the other.
How we labeled types: publishers include media, magazines, independent review and affiliate sites and analyst firms; brand-owned means a company's own site that sells something other than content, including retailers, agencies and SaaS blogs, split by whether the company sells in the prompt's category; review platforms are user-review marketplaces and software directories; UGC means user-published platforms and forums.
What changed between August and September
Perplexity's API stopped citing Reddit and YouTube almost entirely, doubled the sources it returned, and replaced half of its top 20 domains. ChatGPT kept 13 of its top 20 and dropped Forbes.
- Reddit on Perplexity: cited in 13.5% of answers in August and 1.5% in September. Among returned sources, 77.0% of answers included reddit.com in August and 9.5% in September.
- YouTube on Perplexity: 20.0% of answers cited it in August, none in September. It also disappeared from the returned sources, from 67.5% of answers to zero.
- Sources per Perplexity answer: the API returned a median of 20 in September against 10 in August, and the answer text cited a median of 10 against 6.
- Top-20 turnover: Perplexity kept 10 of its August top 20 domains. New entrants included TechRepublic, Tech.co, Capterra, Business Insider and HubSpot. ChatGPT kept 13.
- Same prompt, next month: on Perplexity, the cited domains for a given prompt overlapped by 24% on average between the two months (Jaccard similarity). On ChatGPT, among the 101 prompts with citations in either month, 63% shared no cited domain at all.
- Forbes on ChatGPT: 11.5% of answers in August, 0% in September.
We did not change prompts, engines or settings between runs, so the shifts came from the engines or from the web they search, and our data cannot say which. It also cannot say whether the consumer apps changed in the same way, since we queried the APIs.
How this compares with other published figures
Published studies disagree on Reddit and YouTube, and they measure different things: different engines and interfaces, prompt sets, counting units and dates.
Adweek reported in January, citing six months of data from the analytics firm Bluefish, that YouTube was cited in 16% of LLM answers and Reddit in 10%. Larger multi-engine studies from AI visibility vendors, which include Google's AI surfaces, have ranked Reddit as the most-cited domain overall. Our September API sample found Reddit in 1.5% of Perplexity answers and 1.0% of ChatGPT answers, and YouTube in none.
These results can all be accurate at once. A corpus that includes Google's surfaces measures engines our study does not. Brand-monitoring prompts differ from our category questions, and a window that closed in January or August describes a different moment from a run on September 26. Any single "most cited source" ranking describes one engine set, one prompt style and one date.
What to do with it
Map your own sources, per engine and per category, and re-check them monthly before committing budget to any one type of site.
- Run your priority prompts through each engine you care about and export every cited URL.
- Label the domains by type, the same way we did, so you can see whether your answers lean on publishers, review platforms, brand sites or UGC.
- List the publishers and review platforms that cite competitors but not you. Those are your outreach and PR targets, in priority order.
- Fix the facts on your own pages first where the brand-owned share is high. In our data that share runs to 29% of ChatGPT's citations for B2B software prompts.
- Treat UGC as volatile. Our Perplexity data shows Reddit's share moving sharply within weeks.
- Re-measure monthly with the same prompts, and compare type shares rather than raw counts when an engine changes how many sources it cites.
The guides to ChatGPT's retrieval layer and tracking Perplexity cover each engine in more depth, and our note on answer variance explains the run-to-run variation behind step 6.
Build your own map now and look again next month
Publishers carry most AI citations for buyer questions, the mix differs by engine and category, and it moved within six weeks. Citlyze's citation tracking stores the cited domains and URLs of every answer it collects for your prompts, in the chat engines and in AI Overviews and AI Mode, which is the raw material for this map. Plans start at $29 a month with a 14-day trial and no card.