B2B buying has a fixed question pattern: "X or Y?", "which CRM is better?", "best invoicing software for small business?". Buyers now ask these questions in Claude more than in Google. If your brand misses the answer, you never make the shortlist — and you cannot see this loss in web analytics, because the user never visits your site.
This guide covers Claude's comparison mechanics, the conditions for entering the candidate set, and a concrete 30-day action plan.
Anatomy of a comparison question
Claude answers a comparison question in three steps: it sets the candidates, ranks the criteria, and evaluates each candidate against the criteria. The critical gate is the first step — for a brand outside the candidate set, evaluation never starts.
The candidate set feeds on the model's training knowledge and current context. Being "known" in the category — holding consistent entity information — is the precondition. The second condition is comparability: the model needs concrete data about you on criteria like price, features, and fit. The third is distinguishability: of two brands described in identical sentences, the model usually picks the one appearing in more sources.
Practical test: ask Claude your category's 10 most common comparison questions and note the candidate set in each answer. How many include your brand is the first row of your citation rate panel.
Questions demanding reasons: "why?"
Every "which is better?" is followed by "why?". When Claude builds reasoning, it tries to ground each claim in a source. If the source for claims about you is you — crisp feature pages, documentation, pricing — you hold the narrative. If sources are third parties (review sites, forums, competitor comparisons), the story is written without you.
Three things decide reasoning quality:
- Criterion clarity. Before "which is better?", the model sets "better at what?". You win when criteria land on your strong dimensions (say, setup speed or local-language support). State your strong dimensions in identical language on every platform so the model knows them.
- Evidence density. "Fast setup" alone is weak; "average 3-day setup, 40+ SMB references" is strong. Claims carrying numbers, dates, and proper nouns enter the reasoning.
- Source hierarchy. The model looks for independent sources confirming your claim. Your own site's claim is the start; repetition in press, customer reviews, and industry directories turns it into reasoning.
Practical upshot: do not avoid comparison pages — feed them. Write your own pages with the clarity the model can quote, instead of thin pages that look neutral but carry no information.
Searched and unsearched answers
Claude answers some questions with web search and others from training knowledge. Each route wants different content:
| Searched answer | Unsearched answer | |
|---|---|---|
| Signal | Source links present | Fluent text, no links |
| Winning content | Fresh, quotable passages | Consistent entity information |
| Change speed | Weeks | Months |
| First job | Passage clarity + dating | Entity consistency audit |
Most B2B comparison questions swing between the two routes. Record each question's route separately in your panel; Citation Delta is computed with a route split. Route detection is not a one-off but a standing part of the panel — details in the measurement methodology.
B2B question patterns: build your panel on these
Real buyer questions fall into five patterns; your panel should hold questions from each:
- Direct comparison: "X or Y?", "alternatives to X?" The candidate set is narrow where a competitor is named; appearing here is the highest-intent traffic there is.
- Category recommendation: "Best invoicing tool for small business?" The candidate set is wide; criterion strength decides.
- Criterion query: "Which CRM has local-language support?", "fastest ERP to deploy?" These are your strong dimensions' questions; you should write the answer.
- Risk query: "Is X reliable?", "complaints about X?" This is where negative narratives get written; silence helps the competitor.
- Switching query: "How to migrate from X to Y?", "switching cost from a competitor?" The closest questions to the purchase moment.
For a 20-question starter set, pick 4 questions per pattern; run them across 4 models for 80 observations. The set is frozen and repeated identically on days 30, 60, and 90.
How to structure your comparison page
The comparison content on your own site is the primary source the model consults while writing reasoning. The structure:
- Make the question the headline. The answer to "X or Y?" is sought on a page carrying that question in its headline.
- Write criteria in the buyer's language. "Total cost", not "TCO"; "setup time", not "onboarding".
- Put concrete data on every criterion. Numbers, dates, scope — generic praise never enters reasoning.
- Do not hide your weak criterion. The model completes missing information from competitor sources; telling it in your own frame is always better.
- Date the page. Comparison content ages fast; an undated page looks suspect to the model.
This structure is also passage optimization: each criterion section is a self-contained unit the model can copy.
30-day action plan
Days 1–7: Detect. Collect 20 questions across the five patterns, drawing on what sales hears, competitor FAQs, and community discussions. Ask each in Claude and record the candidate set. Measure your current state with a citation baseline.
Days 8–14: Candidate-set repair. Isolate questions where you never appear; the cause is usually missing entity consistency. Equalize your brand information across platforms and add structured data.
Days 15–21: Reasoning content. Write criterion-based pages or sections for your 5 most valuable comparison questions. Support every claim with numbers and dates.
Days 22–30: Re-measure. Run the same 20 questions under the same conditions. The change in candidate-set entry rate is your first victory metric; citation rate yields a meaningful curve from the second month on.
Three common mistakes
Never naming competitors. A "we don't mention rivals on our site" policy blocks candidate-set entry on comparison questions. The model feeds on sources where two brands appear side by side.
Measuring one question type. A panel tracking only "best X" questions hides losses on risk and switching questions. Distribute the five patterns evenly.
Evidence-free claims. Unsupported adjectives like "leading", "best", or "pioneering" get skipped by the model. Every adjective needs a number or a source beside it.
Silent drops from the candidate set
The sneakiest loss is silently dropping out of a candidate set you once held. Causes: a competitor published fresh content, your pages aged, a model update shifted weights. No alarm rings — your site traffic looks unchanged, but your share of AI answers erodes.
Counter it with a drop alarm in your panel: any question where you appeared last round but not this one gets investigated within 48 hours. Most drops trace to content freshness; updating the page and refreshing its date usually brings you back next round. In the 90-day Filmfolk study the alarm rang twice and both were recovered with corrective publishing — the smooth curve reflects discipline, not luck.
How comparison content relates to FAQs
Comparison content feeds your FAQ page and feeds on it. Three to five FAQs specific to the comparison, placed under the "X or Y?" page, serve user and model alike: the model uses the main passage as the reasoning body and the FAQ items as detail evidence.
But your page FAQs must not repeat your panel questions verbatim — that teaches the model an answer template instead of measuring real visibility. Panel questions use real user language, page FAQs editorial language; both cover the same topic in different sentences. Our site-wide rule is explained on the FAQ page.
Score your own page
Before publishing a comparison page, ask these five questions; each "yes" scores a point:
- Does the headline carry a real comparison question?
- Is each criterion a separate, self-contained section?
- Does every claim carry a number, date, or scope?
- Is your weak criterion told in your own sentences?
- Are publish and update dates visible on the page?
Pages scoring 4–5 enter reasoning; pages below 2 get skipped by the model. Repeat this checklist quarterly — the bar rises as competitor content changes.
Frequently asked questions
How does Claude pick candidates in comparison questions?
The model first builds a candidate set from brands it knows in the category, then ranks criteria and evaluates each candidate. For a brand outside the candidate set, evaluation never starts; the precondition is consistent entity information.
Should content differ for searched and unsearched Claude answers?
Yes. Searched answers reward fresh, quotable passages; unsearched answers rest on consistent entity information in training data. Appearing on both routes takes a page-plus-passage pair.
How should a B2B comparison page be structured?
Write separate, self-contained sections per criterion: pricing, fit, integrations, support. Name criteria in the buyer's language, back claims with concrete data, and date the page.
Conclusion
Comparison questions in Claude are a three-layer game: enter the candidate set, look strong on criteria, become reasoning evidence. In the Filmfolk study the Claude citation rate rose from 0% to 78% in 90 days; the full case is on the results page, and the complete method in the Adaptive RAG methodology. Start by measuring your current state with a citation baseline. Remember: candidate-set discipline is a quarterly routine, not a campaign; repeating the same questions under the same conditions each round turns visibility from luck into a manageable metric. Read your first round's results on day 30 against this article's framework.
Related Articles
GEO for SaaS & B2B
How B2B brands win AI citations across the buying committee's questions.
Which Metric Matters for B2B?
AI citation ranking vs traditional citations — what B2B should track.
What Is Citation Rate?
The core metric of AI visibility: definition, measurement, and benchmarks.
How to Fix a Zero Citation Rate
A diagnostic sequence for brands AI answers never mention.
About the author
Bora Kurum is the founder of RAG Signal and a Ph.D. researcher at Istanbul Bilgi University, where his work focuses on LLM retrieval behavior and brand representation in generative AI systems. Read more →

