Verified against Perplexity Pro · 2026-07-23
Trace a stat or quote back to its actual primary source, not the article repeating it
A citation-chain-walking prompt that instructs Perplexity to keep tracing backward past news aggregators and blog summaries until it reaches the original filing, dataset, speech, or statement a widely repeated figure actually came from.
The prompt
Ready to copy — highlighted parts are example details you can swap.
Find the primary source behind this — the original document, filing, dataset, speech, or press release — not the most convenient article that mentions it or the version of it that ranks highest in search. STATISTIC OR QUOTE "70% of AI projects fail to reach production" WHERE I FIRST SAW IT A conference keynote slide, no source cited on the slide itself WHAT I'M GUESSING IT'S ABOUT Possibly a Gartner or McKinsey survey on enterprise AI adoption PROCESS 1. Do not stop at the first news article or blog post that repeats this — treat that as a lead pointing somewhere, not as the answer itself. 2. Follow the citation chain backward: who did that article cite as its source, and who did that source cite, continuing until you reach something that is itself the original document rather than a summary or report of one. A press release from the organization that generated the data counts as primary; a news article summarizing that press release does not, even if it adds analysis. 3. Along the way, note every place the chain passed through, not just the endpoint — a short list of the intermediate sources helps confirm the chain is real and lets me verify it myself if needed. 4. If the trail dead-ends — the original source is paywalled, unlinked, no longer exists, or you genuinely cannot verify it exists at all — say exactly where the trail broke and what the last verifiable link was, rather than presenting that last secondary source as if it were primary just because the chain stopped there. 5. Once found, quote the exact relevant passage or number from the primary source itself, in its original wording, and note whether any secondary coverage along the way misquoted, rounded, changed the timeframe, or otherwise subtly altered it before it reached where I first saw it. CONFIDENCE State your confidence that the source you found is genuinely the original — high confidence if it's an explicit dataset, filing, or transcript; lower confidence if it's the earliest version you could find but you can't rule out an even earlier origin.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
A search-grounded answer engine's default behavior is to surface whichever page ranks best for the query, which is almost always a summary or aggregator article, not the original document — because ranking well and being first-to-publish the underlying data are entirely unrelated properties, and search relevance has no way to distinguish a primary filing from a well-optimized recap of it. Explicitly instructing the model to treat the first result as a lead rather than an answer, and to keep asking 'who did this cite' recursively, converts a single retrieval hop into a genuine citation-chain walk, which is the actual mechanism needed to reach a primary source instead of the most convenient secondary one a normal search would settle on. Requiring the intermediate sources to be listed, not just the endpoint, matters because a chain-walk that only reports its conclusion is unverifiable by the reader — showing each link lets you independently confirm the chain is real rather than trusting a single unsupported claim of 'I traced this back and here's the origin.' Requiring it to name exactly where the trail broke, rather than quietly presenting the last thing it found as primary, matters because an unverifiable claim that looks resolved is more dangerous than one that's honestly marked unresolved — the whole point of the exercise is knowing which one you actually have at the end, and a model under no pressure to distinguish the two will default to sounding resolved either way.
What you get back
Chain: conference slide (no source) -> a 2023 tech blog post citing "an industry report" with no link -> a 2022 news article citing "a recent Gartner survey" -> Gartner press release, October 2022, which actually states 85% of AI projects fail to reach production, not 70%. Confidence: high — the press release is an explicit, dated primary statement from Gartner itself. The primary source is that press release, and the widely repeated "70%" figure appears to be a rounding or misquote drift introduced somewhere in the secondary coverage chain, likely at the 2023 blog post.
Verified against
Perplexity Pro Sonar Pro · 2026-07-23
Changelog
- 2026-07-23 — Initial publish, verified against Perplexity Pro Sonar Pro search.
Need this built into your business?
If a prompt isn't enough — what Scult builds, built and maintained for you — that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
