Can AI actually read your product pages? A 3,000-page extraction experiment
What AI needs from a product page is not pretty layout but extractable structured facts: specs, ingredients, price, certification.
Key numbers
- Scope: 3,000 food product pages from three online retailers, with the goal of extracting key attributes such as ingredient lists and nutrition tables.
- Two LLM extraction strategies were compared: direct extraction and indirect extraction via generated functions. The indirect route reached 96.48% accuracy, 1.61 percentage points below direct extraction.
- But it cut the number of required model calls by 95.82%, a substantial efficiency and cost gain at scale.
- The authors conclude indirect extraction is the more scalable option for template-heavy product pages.
Method and sample
arXiv paper “Evaluation of LLM-based Strategies for the Extraction of Food Product Information from Online Shops” (2025-06-17): direct versus generated-function extraction evaluated for accuracy, efficiency and cost on 3,000 food product pages from three online shops.
What it means for brands
AI does not read pages the way people do: it needs structured facts it can lift reliably. A page that looks good but scatters its information is hard to extract from — which is exactly why structured data matters for being cited and compared.
How SourceSight handles it
The site audit checks extraction signals (structured data, heading hierarchy, H1, language declaration, link patterns), and the content factory generates structured data alongside page copy; then the same prompts are re-run to compare.
Limits
The experiment covers food product pages only and does not generalise to every category or site; extraction accuracy is model performance on a specific dataset, not the quality of the answer a user finally sees.
References
Run it on your own project
What the same prompts return across 10 engines is only knowable by running them.
Related research
Same engine, different surface, different answer: test app, web and API separately
On the same platform and the same questions, app and web cite systematically different source sets — and the API shows something else again.
Methods · 2026-10-07How to design a retest you can trust: five moves so noise does not pass as progress
AI visibility is a distribution, not a point: fix the prompts, repeat the runs, add controls, paraphrase, validate by hand — or you cannot attribute the change.