Every day thousands of people ask the AI "is X or Y better?" in your industry — and if you don't have a page that answers that question with clear criteria and a verdict, your competitor builds the answer instead. It's not a complicated format to create: it follows a precise, repeatable schema. Creating it just once is enough to start intercepting those citations rather than handing them over.
Try this exercise. Open any AI engine and type a query with “vs” or “or”: “shared hosting vs VPS”, “in-house agency or external consultant”, “WordPress vs headless CMS”. Look at the structure of the answer. You’ll never find a generic “it depends”. You’ll find comparison criteria, strengths and weaknesses for each option, and a contextual recommendation.
This type of answer needs raw material. And that raw material is content that places two options side by side, with explicit criteria. If no one in your industry has written a structured comparison page, the model builds one by assembling fragments from different sources. If someone has written that page, the model takes it and cites it.
That someone could be you.
Why comparative queries are different from all the others
“X vs Y” queries are a special case in the way the RAG system retrieves content. When the user asks for a definition, the model looks for a single chunk that contains the answer. When they ask for a comparison, it has to bring together information about two distinct entities, evaluate them against the same criteria, and produce a balanced judgment. It’s a more complex task — and that’s precisely why the system looks for shortcuts.
In the survey by Gao et al. (2024) on RAG systems, a mechanism is described that applies perfectly to this scenario:
“Routing in the RAG system navigates through diverse data sources, selecting the optimal pathway for a query, whether it involves summarization, specific database searches, or merging different information streams.”
Retrieval-Augmented Generation for Large Language Models: A Survey
Routing selects the optimal pathway for each type of query. For a comparative query, that optimal pathway is the shortest possible one: a page that has already done the work of comparing. If instead the system has to retrieve a chunk about X from one page, a chunk about Y from another, and then merge the information — it’s doing more work with more risk of inconsistency. The comparison page eliminates this complexity. And the system rewards it.
The structure the AI recognizes as a comparison
Naming two options on the same page isn’t enough to have comparative content. The model recognizes the comparative structure when it finds specific elements: the evaluation criteria are explicit, the two options are analyzed against the same parameters, and there’s a conditional recommendation — “if your case is X, then A; if your case is Y, then B”.
In the same survey, Gao et al. introduce a concept that explains why structure matters more than generic content:
“Propositions are defined as atomic expressions in the text, each encapsulating a unique factual segment and presented in a concise, self-contained natural language format.”
Each proposition is a self-contained unit of information. In a well-structured comparison page, every cell of the comparison — “cost of X: high; cost of Y: low” — is an atomic proposition. The retrieval system can extract it, evaluate it, and use it in the answer without having to interpret a discursive paragraph. A comparison written in narrative prose forces the model to do the logical parsing on its own. A criteria-based comparison serves it up ready-made.
The difference is between giving the model an already-chopped ingredient and handing it a whole vegetable to peel. It makes the same dish, but with the first it takes less time and risks fewer errors.
The comparison “Our product vs the competitor” where your product wins on every criterion isn’t a comparison — it’s a landing page with a deceptive structure.
What I saw testing comparative queries
I analyzed 40 comparative queries rephrased in various ways, spread across four different AI engines: ChatGPT, Perplexity, Gemini, and Claude. The queries covered different industries — from technology to professional services — and all followed the “X vs Y” or “is X or Y better” schema.
The clearest result: in 78% of cases where a page with an explicit comparison structure existed — criteria, evaluation for each option, recommendation — that page was cited as the primary source in the answer. When no dedicated comparison page existed, the model assembled the answer from 3-4 different sources, often with inconsistencies in the criteria used for the two options.
The second finding: comparison pages that used tables or structured lists by criterion were cited almost verbatim. Pages that made the comparison in narrative prose were heavily paraphrased — the content was there, but the author’s words were lost in the model’s reformulation.
This connects to what I explained in the article on HTML tables as a structured chunk: the format in which you present the comparison determines whether you get cited in your own words or whether you get digested and regurgitated in a generic version.
Define the criteria before talking about the options.
How to build a comparison page the AI extracts in full
The structure that works best isn’t complicated, but it requires discipline. It’s not an opinion piece along the lines of “why I prefer X to Y”. It’s a criteria-based analysis where the reader — and the model — can find the answer to their specific case.
Define the criteria before talking about the options. If you’re comparing two solutions in your industry, open with the 4-5 criteria that matter to your customer: cost, complexity, scalability, implementation time, support. These criteria become your headings — and each heading becomes a self-contained chunk that retrieval can extract.
Analyze both options for each criterion. Don’t dedicate half the page to X and half to Y. Dedicate one section per criterion, and within each section analyze both. This way every chunk contains the complete comparison on that parameter — and the model can cite it without having to cross-reference distant sections.
Close with a conditional recommendation, never an absolute one. “If you need X, choose A. If your priority is Y, choose B.” As I told you in the article on the balanced pros/cons pattern, models are trained to reward answers that acknowledge the complexity of the context. An absolute recommendation — “A is always better” — triggers the same signal as promotional content.
Include numbers where possible. Numbers, percentages, price ranges, timelines — every concrete data point is one more atomic proposition that retrieval can extract. As I wrote in the article on the numeric pattern, the AI anchors answers on verifiable data because it lowers the risk of hallucination.
The mistake I see most often
Many people make comparison pages that are actually promotional pages in disguise. The comparison “Our product vs the competitor” where your product wins on every criterion isn’t a comparison — it’s a landing page with a deceptive structure. The model perceives it exactly that way.
The comparison pages that work are the ones where both options have real strengths and real weaknesses. Where the reader understands there’s no absolute winner — there’s a better choice depending on the context. This honesty isn’t generosity: it’s the signal the model has learned to recognize as trustworthy during its training.
If your industry has natural pairs — in-house vs external service, solution A vs solution B, traditional vs innovative approach — and no one has compared them with explicit criteria on a dedicated page, there’s a gap the AI is trying to fill by assembling pieces. Whoever creates that page first settles into that gap.
How to check whether you have comparative coverage
The check is simple. Take the 5-10 most searched pairs in your industry — the ones your customers ask you about on the phone when they have to decide — and search for them on the AI engines. For each pair, look: who does the model cite? Is it you? If not, who is it? Do they have a dedicated comparison page, or is the model assembling?
If the answer is “it’s assembling”, you have a concrete opportunity. Write a structured comparison page for that pair — with criteria, analysis by criterion, conditional recommendation — and you become the source the model was looking for.
This is a starting point for understanding where you’re exposed. For a complete picture of how your content performs on the comparative queries in your industry, you need analysis tools that map AI citations systematically. But just knowing which pairs you aren’t covering already gives you a direction on where to act first.
The AI doesn’t invent comparisons out of thin air. It builds them from whoever has already written them. And between a generic page that names the two options in separate passages and a dedicated page that places them side by side with clear criteria, the model chooses the second. Every time.
