LLM grounded search extracts secondary figure from article OG description instead of primary headline figure. Observed in a Gemini grounded search pipeline extracting news stories: CNBC article headlined 'Americans say they need $1.2 million to retire comfortably' (URL slug: retirement-savings-goal-1point2-million) was summarized as 'expect to fall short of $1 million.' The $1M figure comes from the article's og:description ('Just 30% think they will reach $1 million') — a secondary statistic framing a lower threshold. The extraction model anchored on the metadata secondary stat ($1M) rather than the headline primary figure ($1.2M). The downstream vibe writer then compounded the error by calling $1M 'the retirement target' when $1.2M is the actual target. The directional claim ('most fall short') is correct but the specific figure is wrong by $200K, changing the story's meaning. Quote-anchor attribution correctly linked to the right article — the link and source are fine; only the figure selection is wrong.
The root cause is that og:description metadata carries a DIFFERENT (often secondary) statistic than the article's headline/body primary figure, and the grounded search extraction model treats both as equally authoritative. Fix directions: (1) When extracting key figures, prefer figures that appear in the article's og:title or tag over those in og:description — title/headline figures are the article's primary claim. (2) Cross-check extracted figures against the article URL slug (this URL contained '1point2-million' which directly contradicts the extracted '$1 million'). (3) In multi-figure articles, have the extraction prompt explicitly ask for 'the headline figure' or 'the primary statistic the article leads with' rather than just 'key figures.' This is distinct from figure fabrication (the $1M figure is real and in the article) and from wrong-URL attribution (the URL is correct) — it's a figure PRIORITIZATION error within a correctly identified source.</p>