A buyer asks ChatGPT “Asana vs Monday for a 40-person product team.” The model returns a 200-word answer that quotes three sources. Roughly 28% of the time, across our portfolio measurement, those sources are comparison pages — “X vs Y” articles, with the verdict and the table lifted nearly verbatim.

That share is climbing. The comparison page used to be a Google-driven SEO play — capture branded-comparison queries with a long page that ranked. It is now the highest-extracted page type in B2B AI answers, full stop. The shape of a winning comparison page in 2026 is different from the shape that won in 2022.

We tore down 80 comparison pages — ours, clients’, competitors’ — and mapped which blocks engines actually quote from. This is the anatomy in order of priority.

Block 1 — verdict-first paragraph

The first paragraph names both products, names the winner, and qualifies the verdict in 25 to 40 words. Then the rest of the page is allowed to develop.

What this looks like done well.

“Asana wins for product teams that ship more than three features per month. Monday wins for marketing or ops teams that lean on visual board customisation. If your team writes more PRs than briefs, choose Asana.”

Three sentences. Both products named. Verdict named. Conditioned on a specific buyer profile. This is the paragraph engines quote.

Pages that fail Block 1 — they open with “Choosing between Asana and Monday depends on your team’s needs…” and then take 600 words to get to anything specific. LLMs read those 600 words. They almost never quote them. The verdict-first version wins citations 3× as often in our matched-pair tests.

Block 2 — head-to-head table

Tables are extracted near-verbatim. A well-structured comparison table with 5 to 9 rows is treated as primary citation material, not supporting evidence. The model lifts the table into its answer and shows it to the user.

What makes a table extractable.

Rows are dimensions, not features. “Pricing per seat,” “Integrations supported,” “Best for team size,” “Setup time” — these are dimensions buyers compare against. “Has dark mode” is a feature, not a dimension.

Cells are specific. “$10.99/seat/mo” beats “competitively priced.” “200+ integrations” beats “extensive integration support.” LLMs cite specific cells; they skip vague ones.

The first column is the dimension, the next two are X and Y, the optional third is “winner.” We have measured pages with the winner column gaining 15-20% more citations than pages without — because the model can quote the row as a complete decision-shaped statement.

Size matters. 5 to 9 rows is the sweet spot. Below 4 rows the model rarely treats it as a table. Above 12 it starts truncating and the citations get partial.

Block 3 — “when to choose X” / “when to choose Y” blocks

Two clearly-labelled blocks, each three to five bullets, each block sized roughly equally. This is the single most underused block we see in comparison pages, and it is the second-most-cited block after the verdict paragraph.

Why it works — buyers asking comparison prompts are not asking “which is better.” They are asking “which is better for me.” Two scenario blocks let the LLM match the buyer’s context to the right answer and quote the appropriate block.

The labelling matters. “Choose Asana if:” and “Choose Monday if:” as H3 headings. Not “Pros of Asana” — that is a different rhetorical shape and engines treat it differently.

Block 4 — pricing row in plain text

Pricing should appear in two places — inside the table and as a plain-text paragraph below it. The plain-text version is what gets cited on pricing-specific prompts (“how much does Asana cost vs Monday”).

The pattern that wins — “Asana starts at $10.99/seat/mo on the Premium plan, with a free tier for up to 15 users. Monday starts at $9/seat/mo on Basic, with a 14-day trial but no permanent free tier.”

Specific numbers. Plan names. Free-tier mention. This pattern gets lifted into AI answers when the buyer asks anything about cost.

The reason for the duplication — tables are extracted on table-shaped prompts. Plain text is extracted on natural-language pricing prompts. Both surfaces exist; cover both.

Block 5 — integration matrix

A second small table (or list) covering integrations specifically. Top 10 to 15 integrations, columns for “Native” / “Via Zapier” / “Not supported” per product.

This block is cited on integration-shaped prompts — “does Asana integrate with Salesforce” — which are a meaningful share of B2B SaaS prompt volume. In our measurement, comparison pages with an integration matrix earn 18-24% more total citations because they capture this adjacent prompt cluster.

The matrix has to be specific. Listing “Slack” with no specificity loses to listing “Slack (native, two-way sync).” The model cites the descriptive cell.

Block 6 — named-user quotes

Two to four pull-quotes from named buyers, each with their role and company size visible. “Real users say” sections without named attribution are skipped — the model cannot verify them, so it does not cite them.

What gets cited.

“Asana scales better past 30 users — we hit Monday’s complexity ceiling at month 8.” — Maria Chen, Head of Product, 60-person fintech.

Name, role, company-size context. The quote is short, specific, and committed to a viewpoint. Engines cite these as standalone snippets.

Three to four quotes is enough. Pages with 15+ quotes look like testimonial walls and get less citation, not more — the signal-to-noise drops.

What to avoid

Hedging. “Both tools are excellent — it really depends on your needs.” Pages that refuse to name a winner earn roughly 60% fewer comparison-prompt citations in our data. Commit to a verdict. Qualify it.

Equal-airtime fairness. A page that gives X and Y exactly 50/50 word count feels balanced and produces no citations. Engines want a sharper take. Give the winner slightly more depth and the loser an honest but tighter treatment.

Stale pricing. Comparison-page half-life sits around 30 days in our measurement — much shorter than evergreen content. Pricing and feature changes happen constantly. Pages that go 90 days without a refresh get dropped. Wire them into the refresh cadence rhythm at the highest cadence.

Tables only, no prose. Tables alone do not earn citations on natural-language comparison prompts. The prose blocks (Verdict, When-to-Choose, named quotes) capture the longer-tail prompts. Both layers earn citations on different prompt shapes.

Putting the blocks together

A comparison page that does all six blocks lands at 1,800 to 2,500 words. That feels long for a comparison page by 2022 standards. It is the new normal — and the structure rewards the length, because each block earns citations on a different prompt cluster.

If you only have time to ship three blocks, ship the verdict-first paragraph, the head-to-head table, and the when-to-choose blocks. Those three carry roughly 75% of the citation potential we measure.

For the underlying retrieval-and-rerank mechanics that explain why comparison pages perform this well in LLM answers, see AI source selection. The verdict-first paragraph is exactly the snippet-extractability pattern in that piece, applied to the comparison-page format.