The human audit of early 2023 found 74.5 percent citation precision on average (63.6 to 89.5 by engine) and the automatic ALCE baseline about 50 percent; the LLM-judged audit of generative search engines in August 2025 found 39.8 to 68.3 percent citation accuracy. The only repeated design, the BBC rounds of December 2024 and mid 2025, shows significant sourcing issues falling to 10 to 15 percent for three assistants and unchanged near 47 percent for Gemini. Whether support improved between 2023 and 2026 cannot be read from the cross-study numbers, and the one within-design comparison shows improvement for three of four assistants on news questions.

Liu et al. contribute the human-rated baseline and its spread; ALCE the automatic baseline on research pipelines; DeepTRACE the 2025 reading on a partly overlapping set of products (Perplexity, Bing/Copilot, You.com) with a different rater and query set. These three differ in rater, queries and systems, so the step declines to compute a trend from them. The BBC comparison contributes the only pair with one design, with the report's own caveats (product tiers changed from paid to free, definitions adjusted, samples of 362 and 237).

A replication of the 2023 human audit protocol on 2025 or 2026 engines; it would replace the cross-study comparison with a trend. Query: any paper citing arXiv 2304.09848 that reuses its annotation protocol on current systems.

The models have got much better, so citation problems are a 2023 story. No published measurement holds the method fixed across those years except one news audit over six months.