Significant Citation Discrepancy Between Perplexity Web UI and API

[mini]

Hello,

I would like to receive an official explanation regarding the citation quality differences between the Perplexity Web UI search and the Perplexity API responses.

{
    "domains": [
        "ft.com",
        "wsj.com",
        "barrons.com",
        "marketwatch.com",
        "seekingalpha.com",
        "reuters.com",
        "cnbc.com",
        "motleyfool.com",
        "bloomberg.com",
        "investing.com",
        "finviz.com",
        "tradingeconomics.com",
        "simplywallst.com",
        "fortune.com",
        "forbes.com",
        "investors.com",
        "investopedia.com",
        "finance.yahoo.com",
        "theblock.co",
        "nasdaq.com",
        "morningstar.com"
    ]
}

#payload
        payload = {
            "model": self.model,
            "search_domain_filter": domains,
            "messages": [
                {
                    "role": "system",
                    "content": "You are a concise and factual search agent. Search the web for the most relevant and recent information.",
                },
                {"role": "user", "content": query},
            ],
            "temperature": 0.1,
            "top_p": 0.9,
            "presence_penalty": 1,
            # "return_citations": True,
            # "web_search_options": {"search_context_size": "low"},
        }


#result.json 
{
  "test_date": "2025-12-08T10:04:10.380513",
  "test_type": "each",
  "model": "sonar-pro",
  "query": "NVIDIA news",
  "total_domains": 21,
  "success_count": 6,
  "fail_count": 15,
  "success": [
    {
      "domain": "investing.com",
      "status": "success",
      "citations": [
        "https://www.investing.com/news/company-news/nvidia-invests-2-billion-in-synopsys-as-part-of-strategic-partnership-93CH-4383769",
        "https://ca.investing.com/analysis/synopsys-nvidia-deal-highlights-undervalued-growth-after-a-weak-2025-200620326",
        "https://www.investing.com/news/stock-market-news/nvidia-invests-2-billion-in-chip-design-software-provider-synopsys-4383715",
        "https://ca.investing.com/news/insider-trading-news/seawell-a-brooke-nvidia-director-sells-23-million-in-nvda-stock-93CH-4349901",
        "https://www.investing.com/news/stock-market-news/nvidia-cfo-says-chipmaker-yet-to-finalize-100-billion-openai-deal-4386594",
        "https://www.investing.com/news/stock-market-news/analyst-shares-4-reasons-why-he-remains-bearish-on-nvidia-stock-into-2026-4383864",
        "https://ca.investing.com/analysis/nvidia-pops-postearnings-as-a-110-bn-threat-goes-unnoticed-200620030",
        "https://ca.investing.com/analysis/nvidia-earnings-3-red-flags-the-ai-trade-is-topping-200619943",
        "https://ca.investing.com/news/stock-market-news/nvidia-ceo-calls-china-a-bonus-opportunity-amid-surging-global-demand-93CH-4342318",
        "https://www.investing.com/news/stock-market-news/bipartisan-bill-aims-to-block-nvidias-advanced-chip-sales-to-china--ft-93CH-4391843"
      ],
      "content": "Nvidia’s most recent news centers on a major strategic investment, expanding AI partnerships, and rising political scrutiny over China sales.\n\n**Key corporate and deal news**\n\n- **$2 billion Synopsys stake and expanded partnership:**  \n  Nvidia invested **$2 billion** in chip‑design software leader **Synopsys**, buying common stock at **$414.79 per share** as part of a multi‑year, non‑exclusive strategic tie‑up to integrate Nvidia’s **CUDA** and AI/accelerated computing into Synopsys’ engineering and simulation tools.[1][3]  \n  The collaboration targets GPU‑accelerated chip design, p

At first, I used the Perplexity API’s search_domain_filter and called it with our domain list.
Among all 21 domains, the API generated citations for only 6 of them.

For comparison, I tested the same domains individually in the Web UI.
Using a repeated query like “{ ft.com } NVIDIA news.” (changing only the domain),
the Web UI cited 20 out of 21 domains, with theblock.co being the only exception.
This means there is a 14-domain difference in citation results between the UI and the AP

Next, I tried removing search_domain_filter entirely and asked the API using queries that directly included each domain, similar to how the Web UI processes searches.

  TEST_QUERY = "ft.com NVIDIA News"
        
        payload = {
            "model": self.model,
            "messages": [
                {
                    "role": "system",
                    "content": "You are a concise and factual search agent. Search the web for the most relevant and recent information.",
                },
                {"role": "user", "content": query},
            ],
            "temperature": 0.1,
            "top_p": 0.9,
            "presence_penalty": 1,
        
        }



#result.json
{
  "test_date": "2025-12-08T11:17:59.886648",
  "test_type": "no_filter",
  "model": "sonar-pro",
  "query": "ft.com NVIDIA News",
  "result": {
    "status": "success",
    "citations": [],
    "content": "Here are some of the most recent and relevant NVIDIA‑related pieces from the Financial Times (ft.com). I can’t link directly, but you can find them by searching FT with the quoted headlines:\n\n1. **Nvidia and the AI chip boom**\n   - FT has been running regular coverage on how Nvidia dominates the AI accelerator market (H100, B100/Blackwell, and next‑gen architectures) and how that drives data‑centre spending by Microsoft, Amazon, Google and Meta.\n   - Articles focus on:\n     - Whether hyperscalers can reduce reliance on Nvidia via in‑house chips (e.g., AWS Trainium, Google TPU, Microsoft Maia).\n     - Capacity constraints, long lead times, and how much of Nvidia’s growth is sustainable versus “bubble‑like” AI capex.\n\n2. **US export controls and China**\n   - FT has covered repeated rounds of US restrictions on advanced GPU exports to China and the Middle East and Nvidia’s responses (A800/H800, then other “compliant” variants).\n   - Key themes:\n     - Impact on Nvidia’s China data‑centre revenue.\n     - How much lost China demand is offset by US and other markets.\n     - The geopolitical risk premium now embedded in Nvidia’s valuation.\n\n3. **Valuation, earnings and market concentration**\n   - Frequent FT markets and Lex columns analyse:\n     - Nvidia’s quarterly earnings beats, guidance, and dependence on a small number of big cloud customers.\n     - Whether Nvidia’s market cap and revenue growth imply an AI “mania” or are justified by margins and order visibility.\n     - Comparison with prior tech bubbles and with other “Magnificent Seven” stocks.\n\n4. **Competition and ecosystem control**\n   - Coverage of:\n     - AMD’s MI300/MI325 and upcoming MI400 accelerators as the main rival in data‑centre GPUs.\n     - Intel’s efforts (Gaudi line, foundry strategy) and how far behind it is.\n     - Nvidia’s software moat (CUDA, cuDNN, networking/Mellanox, NVLink) and FT commentary on potential regulatory scrutiny over ecosystem lock‑in.\n\n5. **Nvidia’s role in broader AI and semiconductor supply chains**\n   - FT pieces on:\n     - How Nvidia’s demand shapes TSMC’s capex and advanced node roadmaps.\n     - The impact on memory suppliers (HBM from SK Hynix, Samsung, Micron).\n     - Power and infrastructure constraints in data centres driven by GPU clusters.\n\nTo see the latest, go to ft.com and search for:\n- “Nvidia FT Lex”\n- “Nvidia AI chips Financial Times”\n- “Nvidia US export controls China site:ft.com”\n\nThat will surface the most up‑to‑date news and opinion pieces on Nvidia from the FT.",
    "search_results": [],

Running the tests again across all 21 domains,
I observed that this approach produced 1–2 more citations compared to using search_domain_filter.
However, the citation quality still showed a significant gap compared to Web UI search.

The Perplexity official FAQ clearly states the following:

Given this, I would like to ask:

• Why does this citation quality difference occur between the Web UI and the API?

• Does the API truly use the same search system as the Web UI?

• Is this discrepancy intentional by design?

• Is there a roadmap to bring API citation quality to the same level as the Web UI?

Please let me know.

Thank you.