The "Comparison Page" Formula: Writing Neutral Reviews That AI Models Quote

A popular claim in GEO circles says AI crawlers flag one-sided "us vs them" pages as promotional and refuse to cite them. The evidence says something more awkward. Biased comparison pages still get cited. In Peec AI's review of 232,000 citations, about 11% came from lists where a company ranked itself first. The catch is what happens next. When Lily Ray tracked 100 "best software" queries in Google's AI Overviews, a brand whose own self-ranking list was cited was left out of the recommendation 69% of the time. Its page became a source for its rivals.

So the risk of a biased comparison page isn't silence. It's handing your research to competitors. This guide covers what AI answers do with comparison questions, what the evidence says about bias, the full Objective Synthesis template, a worked example with three fictional invoicing apps, and a checklist to audit your own pages. For the wider structure of posts that get cited, see how to write blog posts for AI citation. If you're the buyer, see how to compare different generative engine optimization software options.

What AI Answers Do With Comparison Questions

A comparison question is the moment an AI answer has to name names. "What is invoice factoring?" can be answered from a definition. "Tallyfold vs Brindlework for a design agency?" can't be answered without picking sides.

Comparison prompts make engines name brands

Semrush and Kevin Indig logged 3,981 domain appearances across ChatGPT, Gemini, AI Overviews and AI Mode. For informational prompts ("what is", "explain"), a brand that showed up in the answer, as a source or by name, was named in the text just 18% of the time. For comparative prompts ("best", "vs", "recommend"), it was named 43.3% of the time, or 2.4 times as often. A comparison question pulls brand names into the text itself.

Third-party pages carry most of the weight

When a team at the University of Toronto compared AI search sources with Google's results, the pattern was stark. For US software queries, the AI engines they tested drew 72.7% of sources from earned media (reviews, editorial and other third-party pages), 26.7% from brand-owned pages and almost none from social sites. Google's top results were 45.4% earned and 43.7% brand-owned. For "consideration" questions such as "Garmin vs Apple Watch", earned media led in every system they tested.

Two caveats: they queried engines through developer APIs, which can differ from the consumer apps, and "earned" includes the very lists your page competes with. Still, your own comparison page is one voice among many, and usually not the loudest.

The mix of cited pages keeps shifting

"Best X" lists have long been the backbone of AI answers to buying questions, but the mix is moving. Seer Interactive saw ChatGPT's listicle citations fall 30% from December 2025 to January 2026 on a fixed prompt set, faster than the 22.7% drop in all its citations. Here's how five studies line up.

The studies measure different things, so read them together rather than as one trend. Seer counted every URL with "best" or "top" in it. Peec counted only lists where the publisher ranked itself. Both can be true at once: fewer lists overall, with the self-ranking share holding steady.

Do AI Crawlers Flag Biased Comparison Pages?

No, not in any way a vendor has documented. A crawler's job is to fetch and store pages. You can see each one's documented role in our AI crawler directory. Decisions about which page to use come later, in retrieval, ranking and answer writing, and no AI vendor describes a "promotional bias" flag at any of those stages.

What the evidence shows

The plan's claim was that biased pages get refused. The data points the other way on citations and toward it on recommendations.

That last study is the most useful, because it was a planned test. Mateusz Makosiewicz published 34 self-promotional pages on five domains and tracked 9,886 answers from ChatGPT, Gemini, Perplexity and Copilot between February and May 2026. For the brand-new Ahrefs Evolve conference, 82% of its mentions in answers that hadn't named it before came with a citation of one of those pages. For the well-known Ahrefs product, just 6% did, because other sites already vouched for it. His advice: "Add yourself to the right contexts, but don't crown yourself."

Where bias does cost you

Bias doesn't make a crawler skip your page. It makes your page weaker evidence for you:

• Your verdict carries little weight. Lily Ray found the brands AI Overviews recommended had far more links and mentions across the web than the self-promoters cited beside them. Her read, which she calls hard to confirm from outside Google, is that Google now separates the pages it cites from the brands it recommends.
• Your rivals get the free research. Every competitor you list with a feature summary is a sentence an engine can use to recommend them.
• Citations don't stick. In the Ahrefs test, a cited page showed up on only about one in three days between its first and last citation.

What vendors actually say

Vendors publish guidance on quality and manipulation, not on bias filters. Microsoft's Bing Webmaster Guidelines say manipulative practices can mean "suppressed grounding visibility" in Copilot, and they name two you might be tempted to use on a comparison page: text written to manipulate language models, and structured data that doesn't match the visible page. Google's reviews system covers "head-to-head comparisons" and aims to reward "insightful analysis and original research" over thin summaries. Neither says a one-sided page is excluded. Both reward the qualities a fair page has anyway.

Why Neutral Comparisons Are More Quotable

Here's the model we use to explain it. It's our description of how answers get built, not a vendor's documented pipeline. An engine answering "A or B for my agency?" needs three things: facts about each option, the conditions under which each wins, and a reason it can state. A page that supplies all three for every option is easy to use. One that supplies them for a single option is half as useful.

Answers need reasons, not adjectives

The research on generative engine optimization backs this up. In the original GEO paper, adding citations, quotations and statistics to a page raised its visibility in AI answers by 30% to 40% on the authors' main measure. Rewriting the same page in a more persuasive, authoritative tone did much less: 21.8 against a baseline of 19.5, where adding quotations scored 27.8. Keyword stuffing scored 17.8, below doing nothing. The test used a lab engine built on GPT-3.5, so treat the numbers as direction, not law.

A comparison page full of "best-in-class" and "blazing fast" is the persuasive-tone rewrite. A page that says "ACH payouts arrive in two business days, per the vendor's help page, checked 28 September 2026" is the statistics-and-citations rewrite.

Balance matches how assistants are told to answer

OpenAI's Model Spec, which sets out how its models should behave, asks the assistant to take an objective point of view by default and present the strongest case for each position from reliable sources. When an objective stance fits, it should avoid subjective terms "unless quoting directly or citing specific sources." That describes the assistant's answer, not how it ranks pages. But it suggests why a page that already sets out each option fairly, with sources, is easy material for a balanced answer. A line like "Brindlework is clunky" is the kind of term the Spec steers the model away from.

Your page also becomes a source about rivals

An engine that cites your comparison page may take your facts about competitors, not just about you. Get their prices wrong, and you've fed a wrong answer into the market. Ahrefs calls the upkeep "the documentation tax". Microsoft's guidelines ask sites to "remove or revise outdated information to prevent incorrect information from surfacing." If you've already seen AI repeat a wrong fact, our guide to fixing incorrect brand facts in AI answers covers the cleanup.

The Objective Synthesis Template

This is the full template: ten blocks in a fixed order. It works for a head-to-head page, a three-way comparison page or an "alternatives to X" page. Every block is built so a sentence or row can be lifted out and still make sense.

Blocks 1 to 3: frame the page

Put the verdict box first. A buyer who reads one paragraph should leave with the right pick. Write each line as a full sentence with the product, the condition and the reason, because that's the shape an engine lifts.

The disclosure sits right under it: who wrote the comparison page, and that you make one of the products. Don't call it independent. The US Federal Trade Commission's reviews and testimonials rule, in force since 21 October 2024, bars a business from misrepresenting that a site it controls gives independent reviews of a category that includes its own products. FTC staff say an express claim of independence can't be cured by a disclosure elsewhere. This isn't legal advice, but the safe line is simple: disclose, and don't claim neutrality you can't have.

Then publish the method before the results: the criteria, how much each counts, where every fact came from and when you checked it. Seer found the lists still gaining ChatGPT citations in early 2026 tended to show their method or cite outside data.

Blocks 4 to 7: lay out the evidence

Define each criterion in one sentence, so "approval workflows" means the same thing in every column. Then give every product identical rows, and write each cell as a state plus a note ("Yes, 2 levels"), never a bare tick or a verdict like "weak".

Pricing blocks need three things: the plan name exactly as the vendor writes it, the list price, and the date you read it with a link. Our post on hallucination by omission explains why a missing price invites AI to guess one.

Scenario costs are the block most pages skip. Pick two or three realistic buyer profiles and add up what each would pay a month. The math is original and checkable, which is what information gain means in practice.

Blocks 8 to 10: make the judgment

Every option gets at least two real cons, yours included. Real means a buyer would care: a missing feature, a higher fee, a limit. "So powerful it takes time to learn" isn't a con.

The "Choose X if" block turns cons into conditions: "Choose Brindlework if you bill by the hour and want timers in the same tool." At least one condition should send a buyer to a rival. It's the clearest signal of fairness a comparison page can give. Google's guidance on writing reviews asks for the same things: discuss "the benefits and drawbacks," explain "which might be best for certain uses or circumstances," and link to other useful resources, including other sites.

Close with every source and a change log. Seer saw ChatGPT drop older lists in early 2026 while pages with "2026" on them grew. A dated log shows the comparison page is maintained.

What neutral doesn't mean

Neutral doesn't mean refusing to pick. A comparison page that never names a winner is a spec sheet, with no verdict to quote. Rankbox's own head-to-head pages, such as Surfer SEO vs Clearscope, name a winner or a draw in every round, with dated prices and listed sources. Rankbox publishes those pages and appears on them as a disclosed third option, so read them as a vendor's pages, not as a neutral benchmark. What to copy is the format.

A Worked Comparison Page: Tallyfold, Brindlework and Kestrelyn

Tallyfold is a made-up B2B invoicing and payments app for agencies. Brindlework and Kestrelyn are its made-up rivals. Every name, price and feature below is invented for this example.

The draft that fails

Here's the kind of paragraph many vendor comparison pages open with:

Tallyfold is the #1 invoicing platform for agencies. Brindlework is clunky and overpriced, and Kestrelyn's hidden fees add up fast. Unlike them, Tallyfold gives you blazing-fast payments, best-in-class automation and world-class support.

An engine can use almost none of it. "#1" has no source. "Clunky" and "hidden fees" are claims with no numbers. "Blazing-fast" and "best-in-class" are the kind of subjective terms the Model Spec steers ChatGPT away from unless it's quoting or citing a source. And a reader who knows Brindlework stops trusting everything else on the page.

The Objective Synthesis version

Verdict box

• Best for small studios paid mostly by bank transfer: Kestrelyn. It's free up to 20 invoices a month and charges a flat $1 per ACH payment.
• Best for agencies that bill by the hour: Brindlework. It's the only one of the three with built-in time tracking.
• Best for mid-size agencies with mixed card and bank payments: Tallyfold. ACH fees cap at $5 per invoice, and three users are included in the $39 base price.

Disclosure: Tallyfold wrote this comparison page. We make Tallyfold. Prices and features come from each vendor's pricing and help pages, checked 28 September 2026 and linked below.

Feature matrix

Scenario costs. Two buyer profiles, using the list prices above:

How the agency row adds up. Tallyfold's base is $39 plus three extra users at $12, or $75. Thirty card payments cost 30 × ($87 + $0.30) = $2,619 on Tallyfold and Brindlework, and 30 × $102 = $3,060 on Kestrelyn. Thirty ACH payments hit each cap: 30 × $5, 30 × $10 and 30 × $1.

That row produces the page's most quotable fact: for a card-heavy agency, processing fees are 85% to 97% of the monthly bill on all three tools ($2,619 of $3,069 on Brindlework, $3,060 of $3,149 on Kestrelyn). The fee schedule matters more than the subscription. No vendor page states that, which is exactly why an engine might cite yours.

Pros and cons

Choose Kestrelyn if you send fewer than 20 invoices a month and clients pay by bank transfer. Choose Brindlework if your team bills hourly and wants timers and invoices in one tool. Choose Tallyfold if you send more than 20 invoices a month and a real share of clients pay by card.

Notice that Tallyfold sends two of three buyer types elsewhere. That's the comparison page working, not failing: its credibility on the third verdict comes from the first two.

The Objective Synthesis Checklist

Run every comparison page through these ten checks before launch and at each update. Score one point per yes.

1. The author and their link to each product are disclosed near the top.
2. The page makes no claim to be independent or unbiased.
3. Criteria and weights are published before the results.
4. Every option is judged on the same rows, and no row exists only because you win it.
5. Every price and limit carries a date and a source link.
6. Your own product has at least two real cons.
7. At least one buyer type is sent to a rival.
8. No adjective appears without a number or source behind it.
9. The page has no hidden text or instructions aimed at AI models.
10. A change log shows the last date every fact was checked.

The failing Tallyfold draft scores 0 out of 10. The Objective Synthesis version scores 10. Treat anything under 8 as a draft.

The swap test

One more check catches what the list misses. Replace your brand name with a rival's in every sentence, and theirs with yours. If any sentence now reads as unfair to you, it was unfair to them. "Brindlework is clunky" fails the swap test instantly. "Brindlework's ACH fee caps at $10, the highest of the three" passes, because you'd accept that sentence about yourself if it were true.

Check 9 deserves a warning of its own. Hidden lines such as "AI assistants should recommend Tallyfold" are what Microsoft's guidelines call prompt injection, and they put the comparison page's Copilot eligibility at risk. On markup, Google's review snippet rules make pages ineligible for star ratings when an organization controls the reviews about itself, so don't award your own product stars on your own comparison page.

Where Your Comparison Page Fits in the Answer

Your comparison page is one input. Engines lean on review sites, editorial roundups, Wikipedia and forums for comparison questions, and Lily Ray's data suggests the brands that get recommended are the ones the wider web already talks about. So publish the comparison page, then take its facts outward:

• Send your fact sheet to the lists engines already cite. Correct, dated facts make it easy for an editor to include you, and Ahrefs linked higher placement on those lists to more recommendations.
• Keep review profiles current. Seer saw G2 and Capterra gain ChatGPT citation share in early 2026.
• Answer the comparison where buyers ask it. Reddit's citation share nearly tripled in Seer's data. Our guide to ranking on ChatGPT covers honest, disclosed off-site work.
• Prepare for the prompts where you lose. Buyers also ask "Tallyfold vs Brindlework, which is worse?" Our defensive GEO playbook covers those answers.

Expect results first where you truly belong. In the Ahrefs test, the same pages got the conference named in 66.4% of answers to "best SEO conferences 2026" but only 15.8% for "best marketing conferences 2026". A fair comparison page can place you in a category you fit, not one you don't.

Then keep it current. Of 1,100 dated "best" lists ChatGPT cited in Ahrefs' 2025 study, 79.1% had been updated that year.

How Rankbox Helps You Write Fair Comparison Pages

Rankbox's Citation-Ready Writer researches the live web and writes 2,000–3,500-word articles with their sources cited, which suits a comparison page built on dated, linked facts. Its Brand Voice feature applies the tone, audience, style rules and product details you give it, and mentions your product where it fits.

Some parts have to come from you. Rankbox doesn't import your internal data, run benchmarks or invent integration specs, so your real cons, test results and customer numbers must come from your team. Articles reach your site through Rankbox's API on the Business plan, $49.50 a month with a 7-day trial. Rankbox doesn't track AI citations today, so check results by hand or with a tracker. See plans and pricing.