Comparison Pages: Put Yourself Next to Competitors

On the first Monday in June, a two-person team selling portable power stations to campervan builders types a question into Perplexity: “Which 1,000 Wh unit runs a 12 V fridge and a laptop for two nights off-grid?” The answer names two models, then “other brands are also available.” Their site has eleven van-life photos, a founders’ road-trip story, and “quiet, fast, and built to last.” No row can be set beside a competitor’s spec sheet; both named brands publish a table with conditions underneath.

A comparison page earns citations not by claiming the win but by stating who you are compared with, on which dimensions, and by which convention.

Models want rows they can align, not adjectives

Asked a comparison question, a model has two routes: lift a table already aligned, or stitch one sentence each from five pages. The second invites attaching Brand A’s wattage to Brand B.

A brag page fails not because it brags but because nothing on it converts. “Cools faster” carries no ambient condition, “quieter” no decibel reading, “30% longer runtime” never says longer than what. A person reads confidence; a model reads a claim it would have to vouch for, drops the page, and returns to roundups where someone else set the dimensions. That is the mechanism behind why AI keeps only three to five names.

Choose the dimensions buyers use to exclude

One test decides whether a row belongs: both brands must be expressible in the same quantity.

Dimension type	Writable row	Row to delete
Hard constraint	1,024 Wh, 26.5 lb, 12 V DC output	“High capacity,” “portable”
Operating condition	To 80% in 68 min at 77°F from a 120 V outlet	“Charges fast”
Cost basis	Landed price with shipping and state tax	“Great value”
Risk basis	5-year warranty, who pays return freight	“Worry-free support”
Time basis	2–4 days from a US warehouse	“Ships quickly”

Five to seven rows is enough; past ten, models grab the wrong line. Choose them from 30 days of support tickets and one-star reviews, sorted by the reason people did not buy — rarely the six the product team expects.

One row has to be a loss

A table you win on every line reads as advertising, and the model treats it that way. Asked “what are the downsides,” an assistant needs a source: if you wrote it, it uses your sentence; if not, it quotes a complaint from three years ago.

Three rules: neutral definitions, so your peak output is not set against a rival’s continuous rating; a source and date per competitor value, with “not published” where you cannot get one; and an explicit who-should-not-buy line, because weight that rules out carrying the unit up a ramp screens out returns before they happen. This is extraction efficiency, not virtue.

Example: one table, two versions

Example — brands, specs, and prices are invented; no product is described.

Before: a screenshot titled “Why choose us” with rows reading “More powerful,” “Faster,” “Built to last” — no units, no conditions, no selectable text.

After:

Dimension	Brand A 1,024 Wh	Brand B 1,200 Wh	Brand C 768 Wh
Capacity / weight	1,024 Wh / 26.5 lb	1,200 Wh / 34.2 lb	768 Wh / 19.8 lb
AC recharge to 80%	68 min	95 min	Not published
Continuous / peak output	1,000 W / 2,000 W	1,200 W / 2,400 W	800 W / 1,600 W
Warranty	5 years	3 years	2 years
US warehouse to delivery	2–4 days	3–5 days	5–7 days

Sources: manufacturer spec pages and manuals, 2026-05-20; capacity at 77°F, discharged to 80%. Brand A is the heaviest unit and the worst to haul into a cabin — the row we lose. Brand C’s empty cell is deliberate: a model that knows a value is missing leaves the comparison open, while an unreproducible number is what gets quoted back at you.

Four steps, then schema and links

Step	Action	Owner	Passes when
1 Pick rivals	Three names already in AI answers	Content, growth	Each seen in two samples
2 Fix the convention	Six exclusion dimensions, same units	Content, product	Hide the names and it still compares
3 Build the table	Native HTML; own Product/Offer per model	Front end, content	Visible with JS off; price and stock match
4 Wire the links	Comparison page to each PDP and back	Content	Three clicks from home, in the sitemap

Two failure modes recur: a table rendered as an image, where a parser sees only a filename, and a page behind parameters (/compare/?a=1&b=2) that is crawled inconsistently — give each pairing a static path. Schema does one job: each model’s own Product/Offer, matching the page, with the 2026 Product schema fields. Direction beats count: a PDP pointing here moves an interested buyer into the deciding step; the return link gives the model somewhere to read on.

Two weeks on, rerun the prompts the AI visibility baseline way and see whether answers repeat your conditions, not just your name. The content and semantics hub carries the rest of the June line.

Where this breaks

First, the dimensions cannot be objectified: mattress feel, fragrance, sofa comfort written as rows dress opinion up as specification.

Second, competitor data cannot be publicly verified. If your only evidence is a teardown you will not publish, leave the cell empty.

Third, the brand is not in the candidate set. Zero off-site mentions and three spellings leave even a fair table nothing to attach to; settle whether your brand entity exists first.

Fourth, the sale happens inside a marketplace, where a site-side page does little; fix the listing’s specification block first.

Fifth, comparative advertising has legal edges: EU rules require comparisons to be verifiable and not misleading, so have counsel read the page before launch.

This month, build one five-row, three-column table against a competitor already in AI answers: one row you lose, a source and date per cell, a static URL, links from one PDP and one category page. Ask six prompts a fortnight later and see whether the model has adopted your units.