Perplexity Fingerprint

Technical description

F5 scores a document on two components and normalises the sum to the 0 to 5 scale. The lexical component sums weight × occurrences over a per-language dictionary of Perplexity-associated source-attribution phrases weighted 1 to 5 (studies indicate, evidence suggests, research has shown at 2; according to research, as reported by, as documented in at 3; sources indicate, sources suggest, according to multiple sources, multiple sources confirm, the consensus among sources at 4; based on the search results at 5); a matched phrase that also belongs to the shared cross-model generic set is multiplied by 0.25 first. The structural component targets the answer-engine citation signature: inline numeric citations of the form [N] with no accompanying bibliography (+5), residual bare URLs (+4), minimal formatting where four or more paragraphs carry no headings and no bold (+2), three or more generic source phrases (+2), and a dense block of five or more uniform inline citations (+2), the last of which also sets recommend_citation_verification in the metadata. The raw total R maps to the reported score as min(5.0, R / 15 × 5). Twelve language dictionaries are available; the document language selects one. F5 reports the pattern at Layer 1 and hands the authenticity question to the citation-verification indicators L3, G1 and G4.

How it works

The implementation is deterministic and runs at Layer 1 over compiled regular expressions.

Lexical scoring. The active per-language dictionary maps each phrase to an integer weight from 1 to 5, rising with how distinctive the phrase is of Perplexity output. The weight concentrates on the source-attribution register the answer engine uses to frame retrieved material: based on the search results at weight 5, the source-aggregation formulas sources indicate, according to multiple sources, multiple sources confirm, the consensus among sources at weight 4, the attributive connectives according to research, as reported by, as documented in at weight 3, and the generic evidential hedges studies indicate, evidence suggests, research has shown at weight 2. Each case-insensitive match adds its weight times its occurrence count, and a phrase held in the shared cross-model set has its weight multiplied by 0.25 first, leaving the Perplexity-specific residue. Matches of weight 4 or more are reported at warning severity, lighter matches at informational severity.

Structural scoring. Five signatures contribute. Inline numeric citation markers [N] that appear with no bibliography section add 5, the strongest single signal, because the answer engine inserts numbered references inline while the source list lives in a separate panel that is lost on copy. Residual bare URLs add 4. Minimal formatting, a body of four or more paragraphs with no headings and no bold spans, adds 2 and captures the flat answer layout. Three or more generic source phrases add 2, marking the even, hierarchy-free way the engine attributes every claim. A dense block of five or more uniform inline citations adds 2 and sets the metadata flag recommend_citation_verification, with a finding that cites the 2025 Columbia Tow Center audit, which found AI search engines wrong on more than 60 percent of queries and Perplexity on 37 percent, frequently citing real URLs with fabricated content.

Aggregation and handoff. The lexical sum and the five structural contributions are added into a raw score R, reported as min(5.0, R / 15 × 5). The metadata returns the inline-citation count and the recommend_citation_verification flag, which is set when the document carries at least three inline citations or any bare URL. F5 identifies the citation signature at the surface and routes the authenticity question downstream: L3 verifies each reference against external databases, G1 screens for fabricated citations internally, and G4 checks whether a real cited paper supports its claim.

Score thresholds

Score	Meaning
0 to 1	Sparse attribution vocabulary, citations backed by a bibliography, ordinary formatting.
2 to 3	A concentration of source-attribution phrases, or one structural signature such as bare URLs, flat formatting, or many even source attributions.
4 to 5	The attribution register and the answer-engine citation signature co-occur: inline numbered markers without a reference list, bare URLs, and a dense uniform citation block. Strongly consistent with text pasted from a Perplexity answer, and a prompt to verify every reference.

Why this matters

Perplexity is an answer engine: it retrieves sources and threads numbered citations through a synthesised reply. That mechanism leaves a distinctive trace. The reply attributes nearly every claim, in an even, hierarchy-free way, with phrases like according to multiple sources and based on the search results, and it carries inline markers [1], [2] whose corresponding source list sits in a separate interface panel that does not travel when the text is copied. A manuscript that contains dense inline numbers with no reference section is therefore showing the seam of an answer-engine paste. The stakes are high because these citations are often wrong: a 2025 Columbia Tow Center audit found AI search engines incorrect on more than 60 percent of queries and Perplexity on 37 percent, frequently attributing claims to real URLs whose content does not support them. F5 detects the surface signature deterministically at Layer 1 and routes the question of whether each reference is real and supportive to the dedicated verification indicators, so the fingerprint and the verdict stay separate.

Limitations

The inline-marker signature overlaps with legitimate numbered-citation styles such as Vancouver referencing; the no-bibliography condition is what separates an answer-engine paste from a properly referenced manuscript, so a document whose reference list was retained will not trigger the strongest signal. The attribution vocabulary thresholds were calibrated against 2024-2025 output and require periodic recalibration. The lexical signal yields to paraphrase of the attribution phrases, and the citation signature is removed by attaching a real bibliography. The dictionaries are most developed for English and thinner across the other eleven languages; the citation, URL and formatting checks are language-independent. F5 reports the citation pattern and its associated verification flag; the authenticity of any cited reference is established only by the Layer 3 citation stack, not by this indicator.

Theoretical background

F5 combines lexical fingerprinting with a structural detector of retrieval-augmented output. The lexical weights follow the excess-vocabulary logic shared across the F-series, restricted to the source-attribution register characteristic of an answer engine and separated from the evidential hedging common to all assistants by the cross-model generic discount. The citation strand draws on the emerging literature documenting that retrieval-augmented answer engines produce confident inline attributions that frequently fail verification, with the Tow Center audit providing the headline rates; this motivates treating a dense, bibliography-free citation block as a flag for verification rather than as evidence of scholarship. The division of labour, surface fingerprint at Layer 1, database verification at Layer 3, mirrors the wider citation-hallucination literature, in which detecting the pattern of a fabricated reference and confirming its non-existence are distinct steps.

References

1. Jaźwińska K, Chandrasekar A. AI search has a citation problem: we compared eight AI search engines and they are all bad at citing news. Tow Center for Digital Journalism, Columbia Journalism Review. 2025. https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php
2. Kobak D, González-Márquez R, Horvát EÁ, Lause J. Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Science Advances. 2025. https://arxiv.org/abs/2406.07016
3. Liang W, Zhang Y, Wu Z, Lepp H, Ji W, Zhao X, Cao H, Liu S, He S, Huang Z, Yang D, Potts C, Manning CD, Zou J. Quantifying large language model usage in scientific papers. Nature Human Behaviour. 2025. DOI: 10.1038/s41562-025-02273-8 https://www.nature.com/articles/s41562-025-02273-8
4. Ansari MS. Compound deception in elite peer review: a failure mode taxonomy of 100 fabricated citations at NeurIPS 2025. arXiv preprint arXiv:2602.05930. 2026. https://arxiv.org/abs/2602.05930