We are testing Pro search in API calls to create daily email briefs. It is returning better results overall however the in-line citations ([1]) seem to be one-off from the actual citations in the search results URLs.
Full report:
Pro Search API returns citation numbers that don’t match the actual source article in search_results array. When using domain-limited searches with Pro Search, the model sometimes attributes quotes to the wrong search_results index, causing incorrect citation URLs.
Steps to Reproduce
API Configuration:
-
Model:
sonar-pro -
Search mode:
auto(resolves to Pro Search for complex queries) -
Using
search_domain_filterparameter with specific Substack domains
Example Query:
What insights or analysis has been published about content clipping or viral video production? Structure response with: SECTION 1 - DIRECT QUOTES (extract verbatim quotes with publication names), SECTION 2 - STRATEGIC ANALYSIS
Domain filter:
“search_domain_filter”: [
“creatormag.blog”,
“thepublishpress.com”,
“contentgrip.com”,
“liahaberman.substack.com”
// (and others)
]
The Problem
In the API response, Perplexity cites [3] for this ContentGrip quote:
“Shared Collections expand TikTok’s existing save-and-organize functionality by allowing two users to jointly curate and manage saved videos.” [3]
But search_results[2] (citation [3]) is actually:
-
Title: “McDonald’s AI Christmas ad pulled - ContentGrip”
-
URL:
https://www.contentgrip.com/mcdonalds-pulls-ai-christmas-ad/ -
This article is about McDonald’s, not TikTok Shared Collections
The correct source is search_results[1] (citation [2]):
-
Title: “TikTok Shared Feeds and Collections - ContentGrip”
-
URL:
https://www.contentgrip.com/tiktok-shared-collections-and-feeds/ -
This is where the quote actually comes from
Pattern Observed
This happens consistently when:
-
Multiple search results are from the same domain (e.g., several ContentGrip articles)
-
Using Pro Search with domain filtering
-
Requesting verbatim quotes with citations
The model appears to synthesize across sources but mis-maps which specific article contained which quote.
Environment
-
API Endpoint:
/v1/messages(Perplexity Sonar API) -
Model:
sonar-pro -
Search Context:
medium -
Access: Via Superblocks workflow automation (programmatic API calls)
Supporting Data
Full API Response (excerpt):
{
“message”: {
“content”: “…ContentGrip\n\“Shared Collections expand TikTok’s existing save-and-organize functionality by allowing two users to jointly curate and manage saved videos.\” [3]…”
},
“search_results”: [
{
“title”: "A Prediction Market for Creator Videos? Bet. ",
“url”: “https://news.thepublishpress.com/...”
},
{
“title”: “TikTok Shared Feeds and Collections - ContentGrip”,
“url”: “https://www.contentgrip.com/tiktok-shared-collections-and-feeds/”
},
{
“title”: “McDonald’s AI Christmas ad pulled - ContentGrip”,
“url”: “https://www.contentgrip.com/mcdonalds-pulls-ai-christmas-ad/”
}
]
}
The quote cites [3] (McDonald’s article) but actually comes from [2] (TikTok article).
Expected Behavior
Citation numbers in the response content should correctly map to the search_results array index (1-indexed) where the information was sourced.
Impact
This breaks citation accuracy for automated research workflows that rely on Perplexity’s citation mapping. Users cannot trust that clicking a citation will take them to the correct source article.
I do not think that this was an issue for us when we were using the normal Fast search but popped up when we started testing Pro search (actually set to auto but for these queries it is using pro due to complexity)
Thanks in advance for any help!