Comparison Tables: The Format LLMs Cite Most Often

A restaurant SaaS team added a single HTML comparison table plus a JSON-LD ProductComparison block to 14 versus-pages. Within two weeks, ChatGPT started copying the table verbatim when users asked "thMenu vs competitor X". Across our test set, structured tables were cited 4.8x more often than the equivalent paragraph prose.

Why LLMs Love Tables

Large language models saw millions of <table> elements during pre-training. Row and column relationships are far cheaper to parse than free-form text, so when a user asks a comparison question the model often decides "do not summarise, just show".

Markdown tables work too, but JSON-LD-augmented HTML tables clearly dominate Perplexity's Compare feature and Bing Copilot. The key signals: a header row that names features, short cells with one fact each, and units or currency declared on the cell itself.

The JSON-LD ProductComparison Pattern

Schema.org has no official "ProductComparison" type, but you get the same result by wrapping multiple Product nodes inside an ItemList. Each Product carries its own "additionalProperty" array to list feature-value pairs.

Maintain 1:1 mapping between the visual table and the JSON-LD. If the column "Monthly price" maps to "offers.price", every row must follow. Any mismatch and the LLM falls back to citing prose paragraphs instead of the structured table.

Practical Checklist

• First column = product or competitor name; first row = feature labels.
• One fact per cell (price, integer, "Yes/No"); never write sentences inside cells.
• Place the JSON-LD immediately above the table; if a page contains multiple tables, emit a separate ItemList for each.

The restaurant SaaS team only saw organic traffic grow 18%, but their brand mention rate inside LLM answers jumped 7x over three months — a metric traditional SEO dashboards barely capture today.

FAQ

Will Markdown tables also rank? They will, but the HTML plus JSON-LD combination is what LLMs measurably prefer; reserve Markdown for blog platforms that cannot inject schema.

How many columns is ideal? Between three and seven. Past eight columns, mobile rendering breaks and LLMs revert to summarising rather than quoting.

Is naming a competitor risky for SEO? No — users with comparison intent already know the competitor's name, and LLM answers now dominate that query shape, so visibility matters more than caution.

Found this helpful? Share it.