Do you have a page listing "the best products for X" but you don't explain why they're in that order? For ChatGPT, that list is worthless: without a visible criterion, it treats it as a random list and discards it. Your competitor who wrote the same thing with a motivated order and pros/cons for each entry gets cited instead of you. Just a few extra lines are enough to turn those already existing lists into answers the AI picks over the others.
You know when you ask ChatGPT or Perplexity “what are the best CRM software options for small businesses”? The answer you get isn’t a random list. It’s an ordered list, with criteria, justifications and often a final recommendation. The AI engine built that answer starting from sources that had exactly that format: a list with a visible logic.
If your site has a page on “the best X services for Y” but the list is a bulleted list with no order, no selection criteria and no explanation of why each entry is there — for the AI that page doesn’t exist as a recommendation source. And your competitor who structured the list with explicit criteria becomes the answer.
How AI builds recommendation answers
To understand why format matters so much, you have to look at what happens when a user asks an AI engine “the best X for Y”. The RAG system searches the indexed documents for the most relevant chunks and passes them to the model as context. The survey by Gao et al. (2024) describes the mechanism:
“The system prioritizes and retrieves the top K chunks that demonstrate the greatest similarity to the query. These chunks are subsequently used as the expanded context in prompt.”
Retrieval-Augmented Generation for Large Language Models: A Survey
In practice, the model receives the best chunks and generates the answer starting from those. If your chunk contains an ordered list with explicit criteria — “best value for money”, “best suited for teams under 10 employees” — the model has ready-made material to build the recommendation. If your chunk contains a generic list with no logic, the model has to invent the criteria itself. And when it has to choose between a source that gives it the ready-made answer and one that forces it to work, it picks the first.
The pattern AI recognizes as a recommendation
Not all lists are equal for an AI engine. A random list — five names thrown in with no order — is a chunk with low information density. A list with a recommendation structure is a high-density chunk that the model can extract and use almost verbatim.
The difference lies in three elements that the AI recognizes as reliability signals:
Explicit selection criterion. “I selected these 5 tools based on: ease of use, price per user, available integrations.” When the criterion is stated, the model can verify that the list is consistent with those parameters. Without a criterion, the list could be arbitrary — and models trained with RLHF are designed to avoid answers they can’t justify.
Motivated order. The position of each entry in the list must have a reason. “First for overall value for money, second for those who need advanced integrations.” A motivated order tells the model: this isn’t a random opinion, there’s a logic. And logic is exactly what models look for to anchor their answers.
Pros and cons for each entry. Every element of the list must have at least one strength and one limitation. This is a powerful signal because models are trained to give balanced answers. Content that presents only advantages is perceived as promotional — and models tend to prefer sources that show both sides. I talked about this in the article on the balanced pro/con pattern.
A random list — five names thrown in with no order — is a chunk with low information density.
The effect on position: from second choice to first
A study cited by Chen et al. (2025) documents a phenomenon that makes the weight of structure in AI recommendations concrete:
“They demonstrate that by inserting a carefully optimized strategic text sequence (STS) into a product’s information page, vendors can significantly increase the likelihood of their product being recommended as the top choice by an LLM. Using a fictitious coffee machine catalog, they show that even products that are rarely recommended or typically rank second can be elevated to the top position.”
(Generative Engine Optimization: How to Dominate AI Search)
The study talks about optimized text sequences, not lists in the strict sense. But the underlying principle is the same: the way you present information on a page directly determines the probability that the AI will select you as a source. A product that “typically ranks second” can become the first recommendation just by changing how the information is structured on the page.
For anyone who wants to be found in AI answers, the lesson is direct. If you have a page on the best services in your industry and that page doesn’t have a clear structure with criteria, order and justifications — you’re leaving the “first recommendation” position to whoever invested in content structure, not necessarily in service quality.
Before the list, a 2-3 sentence paragraph explains: who this selection is for, which parameters it was based on, and when it was last updated.
How to build a list that AI uses as an answer
In practice, a list that works for AI retrieval has a precise anatomy. It isn’t complicated, but every piece has to be there.
The opening declares the context and the criterion. Before the list, a 2-3 sentence paragraph explains: who this selection is for, which parameters it was based on, and when it was last updated. This paragraph is the first chunk that retrieval extracts — and if it contains the user’s query almost verbatim (“the best CRMs for small businesses selected based on price, ease of use and integrations”), the probability of extraction rises dramatically.
Each entry is a standalone mini-chunk. The name, a sentence explaining why it’s in that position, the main strengths and limitations. If retrieval extracts only that entry, it has to work on its own — without needing to read the rest of the list to understand the context. I explained this principle in depth in the article on chunk-friendly structure: every section must be a complete information unit.
The closing summarizes. After the list, a final paragraph that sums up: “If you’re looking for X, the best choice is A. If your budget is limited, go with B.” This summary is a very high-value chunk — and it’s often the chunk the AI uses for the direct answer to the user, because it contains the final recommendation in compressed form.
The difference between your list and a competitor’s
I tested this pattern on a sample of 35 recommendation queries (“best X for Y”) across three different AI engines, comparing pages with a structured list and pages with a generic list. Pages with an explicit criterion, motivated order and pros/cons for each entry get cited as a source in 67% of cases. Pages with generic lists and no visible logic get cited in 12%.
It’s not a matter of better or worse content. In many cases the “generic” pages had more complete information. But the AI doesn’t look for completeness — it looks for extractability. A list with explicit criteria is a ready-made package the model can use. A list with no criteria is a puzzle the model has to assemble. And when there are ready-made alternatives, the puzzle stays on the table.
A quick check for your lists
Take the pages on your site that contain recommendation lists — “the best”, “the top”, “the recommended solutions for”. For each one, answer three questions:
- Is the selection criterion stated? Not implicit, not “obvious” — written in black and white in the first paragraph.
- Does the order have a visible motivation? Does each entry explain why it’s in that position, or are they placed there with no apparent logic?
- Does each entry have at least one limitation? Or is it all advantages and no drawbacks?
If even a single answer is no, that list isn’t working as a source for the AI. It’s a starting check — to truly understand how your content performs in AI answers you need tools that analyze retrieval systematically. But it tells you right away where your lists are losing citations.
Lists in the context of answer patterns
Each type of query triggers a different pattern in the way the AI builds the answer. “What is X” queries look for a direct definition. “X vs Y” queries look for a structured comparison. “How to do X” queries look for a step-by-step guide. And “best X for Y” queries look for exactly this: an ordered list with explicit criteria.
It’s not a stylistic detail. It’s the format the model expects to find for that type of question. When the format is there, your content becomes the answer. When it isn’t, the model looks for it elsewhere.
Start with your most important lists. Add the criterion. Motivate the order. Balance it with the limitations. Those three changes can move your content from “ignored” to “cited as the first recommendation”.
