A Website Is More Than a Data Source: It Is the Database AI Retrieves From
A data source is not enough for AI to return an accurate answer about a vendor.
Designing the website as a database increases the probability of a complete and accurate answer.
One topic per page, one answer per section, and each section readable on its own.
In AI-mediated shortlisting, buyers outsource research to AI, and they assume their research assistant returns accurate answers: What Is AI-Mediated Shortlisting. For AI to return an accurate answer about a vendor, it is not enough that the vendor's website is a data source for AI.
A data source can hold the answer in pieces across several pages, or hold multiple answers to one question. It can also hold the answer in one section that needs the rest of the page to be understood, and AI retrieves the section without the page. In each case, the answer AI returns is more likely to be AI's own than the vendor's.
Designing the website as a database increases the probability that AI returns one complete and accurate answer: one topic per page, one answer per section, and each section readable on its own.
Most marketing agencies treat a website as one of the data sources for AI answers, and stop there, because their goal is AI visibility. Accuracy is a different goal.
A Data Source Holds Facts.
A Database Returns the Right One.
A data source only has to contain facts. Like a flat file, it has no rules for how the facts are organized or for keeping them complete and consistent. A database has both: a structure that gives each fact one place, and integrity rules that keep each record complete and consistent. That is why it returns the right fact when queried. When AI looks for answers to a buyer's questions, it queries the website: it looks for specific answers and returns them to the buyer.
Answering a question is a different job than most websites were designed for. Websites were initially designed for people: a destination to visit, explore, learn, and form an impression over multiple visits. Then they were optimized to be found, first by search engines and now by AI answer engines. Neither the destination nor the visibility optimization is about answering a question.
In AI-mediated shortlisting, the goal is different. Once AI has found the candidate vendors, the buyer asks it to help build the shortlist: The 3 Buyer Jobs Framework. As a research assistant, AI is then not trying to find the vendor. It is trying to answer the buyer's questions about the vendor. That requires a website optimized for accuracy, not discovery.
How to Design a Website Like a Database
Designing a website like a database means treating a page like a database table, a section like a row, and a hyperlink like a foreign key.
A table holds one subject, so a page holds one topic.
Different topics go on different pages, the way customers, orders, and addresses go in different tables.
A row is one complete record, so the section under each heading is one complete answer to one question.
The row holds everything about one record. The section holds the complete answer to one question.
A foreign key defines the relationship between two tables, so a hyperlink shows AI the relationship between two topics.
When a page mentions a topic it does not own, it should link to the owning page. The link tells AI where the complete answer is. The clickable link should name the topic.
Databases are designed this way so that when queried, they return a complete and accurate answer. If the customer's address is stored in the orders table as well as the address table, then the day it changes in one place and not the other, the database has two answers to one question. Database designers use one fact, one home, and a foreign key to prevent that inconsistency.
A website that follows the same design should reduce the probability that AI returns an incomplete or inaccurate answer, for the same reason a database does: one fact, one home, and a link to it. The design has three rules.
One Page Owns Each Topic on a Website
Just like a normalized database gives each subject one table and each fact one home, a website designed like a database gives each topic one page. That page owns the topic.
Other pages may mention the topic, but only as a summary: less detail than the owning page, and a link back to it. Which pages a website needs, and what each one owns: The Foundational Pages.
Each Section Is One Complete Answer to One Question
Each section on a page is the definitive answer to one question. AI is asking questions, and it should find the complete answer in one place, not pieced together from several sections. Do not make AI think. Give it the answer it is looking for. The questions AI is asking: The Six Buyer Questions.
For the writer, the rule is simple. The H1 states the question the page answers, or the claim that answers it. Each H2 and H3 does the same for a supporting question, and the paragraphs under it give the answer. All the sections together answer the H1 question.
Each Section Stands on Its Own
Each section on a page has to be understood on its own. A person reads a page top to bottom, so when a section refers to something said earlier, the reader has the context. AI does not. It retrieves the section without the rest of the page, so the section cannot rely on the page to be understood. Each of these points to text that AI did not retrieve:
"As described above" or "as mentioned earlier"
"See below" or "the section above"
"The same approach" or "the same process"
"This" or "it," when the noun it stands for is in another section
A section has to name what it needs to be understood: the name of the service, the term being used, the question being answered, even if the section before it said the same thing and it sounds repetitive.
That repetition is not a competing claim, and it is not duplicate data. A competing claim is a second answer to the same question. Duplicate data is the same fact stored in two places. Repeating the name of the service, or the term, or the question is not repeating the answer. It is saying which answer the section gives, so the section can be read on its own.
AI Retrieves Blocks, Not Pages
The exact mechanism for reading a web page varies by system, but they all have one thing in common: none of them uses the page as a whole. Each uses pieces of it:
What gets used is a section, not a page.
On a page designed like a database, the block AI retrieves is the section the writer wrote: one heading and the paragraphs under it. "Block" is what AI retrieves, and "section" is what the writer writes. They should be the same thing.
AI Reads Content, Not Intent
AI reads content: the block of text on a page. It does not read intent: what the page was for. A person figures out what a page is for from where it sits in the menu, how it looks, what gets emphasis, and what comes first. AI reads the text.
So when AI retrieves a block, it does not know how that block relates to the rest of the page or what the page was designed to do.
A Block Is Analyzed Without the Full Context of the Page
AI analyzes a block without the full context of the page because the block is what AI uses, not the surrounding text. Some systems retrieve only the block. Others fetch the whole page and keep only the pieces that match the question.
What comes back is the block, the page title, and the address.
What does not come back is everything a person uses to read intent: the layout, the size of the type, what sits above the fold, which menu the page is in, and what the other pages say. The menu labels may survive as a list of links, but their prominence does not.
What Goes Wrong When a Website Is Not Designed Like a Database
The accuracy of the answer depends on how a website's content is organized, not only on how well it is written. A competing claim is when another page answers the same question in its own words instead of pointing to the owner. Now AI has two answers to one question, and the two answers take one of two forms: the pages say the same thing in different words, or the pages disagree.
When the pages disagree, AI will typically pick one and disregard the others. When the pages say the same thing in different words, AI still has more than one version and no way to tell which one the vendor meant. There is no rule that the page in the top menu counts for more than the one in the footer. The choice is AI's, not the vendor's.
The pages say the same thing. The common case is industry pages. A vendor solves the same problem the same way for four industries, so it writes four pages that say the same thing in slightly different words, because each page is written for a different reader. AI does not see four readers. It sees four versions of one answer, and which version it uses is not the vendor's choice.
The pages disagree. That is a contradiction, and it breaks the rule that each fact has one home. A homepage lists two counties as its service area because it is supposed to be brief. The service-area page lists three because it is supposed to be complete. AI does not know what either page is meant to be. It sees two pages describing the service area and picks one. It could pick the homepage, which lists only two counties, and the answer it returns is not completely accurate.
When No Page or Section Has the Whole Answer
Several pages, or several sections, that each hold part of the same answer give AI no single place with the whole answer. That is a split answer, and it breaks the database rule that each record is complete. A complete answer inside one block is more likely to be retrieved, and to produce an accurate answer, than the same answer in pieces across several.
Every extra block is another hop, and each hop adds two risks. AI may not retrieve the piece at all. If it does, it has to assemble the pieces into an answer of its own instead of finding the answer stated in one place.
When the answer is in one block, AI quotes it or restates it, and the answer is the vendor's. When the answer is in pieces, AI writes it, and the answer is AI's, like the last player in the telephone game. Accuracy falls with each hop.
For example, a vendor describes how it works in three places: a line on the services page, a paragraph in a case study, a section on the about page. Each is true. None is the whole method. AI has to find all three, then combine them into its own version of the methodology, filling the gaps from its training data and the biases built into it. What it returns is AI's version, not the vendor's.
When a Section Depends on the Text Around It
When a section needs the rest of the page to be understood correctly, it is unlikely that AI will read the rest of the page to get that context. AI uses just the block. Research on retrieval finds that a block which stands on its own is retrieved more accurately than one that depends on text outside it.
Topic Clusters Increase Visibility but Decrease the Probability of Accurate Vendor Comparisons
Traditional SEO and AEO add pages to a topic to cover more queries, because a site that matches more queries is found more often. In database terms, that stores one subject in several tables, which breaks the design of a website as a database.
In a topic cluster, the pillar page holds the main topic, and each cluster page is written for a different but related search query or question. The rule is that one page owns each topic. Cluster pages break the database rule in two ways. First, they give AI more than one version of the same answer. Second, they split the answer.
What a Topic Cluster Is Built to Do
A topic cluster is built for search rankings and AI visibility. Marketers call the goal topical authority: the site that covers a subject most broadly is treated as the expert on it. The pillar page covers the topic broadly. Each cluster page covers one narrower but related query, a long-tail keyword, and links back to the pillar "so search engines can see the relationship." A pillar on leadership development gets cluster pages for first-time managers, for manufacturing companies, for remote teams, and so on. More pages, more queries matched, more chances to be found.
It works for what it is built for, and for a vendor whose problem is being found, it is a sound strategy. The industry's own analysis of 54 studies puts ranking for multiple related queries among the strongest citation factors. That is discovery, or visibility. The trade-off is accuracy. When a buyer is comparing vendors, accuracy is what matters, and the cluster method works against it.
Cluster Pages Are Not Duplicate Content
Cluster pages are written to be different: each page describes the same topic in different words, optimized for a different query or prompt. To a search engine, that is three different pages. To AI, it is three versions of one answer, with no way to know they were meant to be the same. Restating does not help: paraphrased copies of the same content scored no better than the original, and a single paraphrase scored below it, for every model tested.
Duplicate content is one text at two addresses, and a canonical tag settles it. This is one topic in three texts, and there is nothing identical for a canonical tag to point at. Where a cluster page instead covers one part of the topic the pillar leaves out, the topic is split across pages.
The Pillar Page Is Not the Problem
The pillar page is not the problem. A comprehensive standalone page on one topic is a database page. The marketing definition of a pillar, "a comprehensive, standalone page that covers a topic in-depth, all on one page," is one page per topic. The same industry analysis that scores query coverage so highly also scores self-contained passages, "important statements can stand alone without additional context," as a citation factor in its own right.
Traditional SEO and modern AEO tactics and the database concept agree on how to write the pillar page. They disagree on the value of the cluster pages. Cluster pages add visibility and rankings, at the cost of being compared accurately. Which a vendor needs depends on which problem it has: being found, or being compared.
The Rules Are for AI. The Website Is Still for People.
The database rules say how a website's content has to be organized, so that AI has the highest probability of returning an accurate answer when it queries the website. That matters when the buyer is building a shortlist. Once a vendor makes the shortlist, the buyer visits the vendor's website to validate it: The 3 Buyer Jobs Framework.
The same page that answered AI's question has to work for a person: easy to navigate, the right brand, the right look and feel, and the emotional appeal of a vendor worth choosing. A page that works for AI and loses the person reading it has not done its job. Balancing the two is why a vendor needs a good website designer.
About the author
David Lee
David Lee is the founder of Do What Works. At Toyota Motor Sales, he was part of a team that selected a vendor, from gathering requirements to the final decision. He then ran a B2B marketing agency for ten years. Today he runs buyer conversations through ChatGPT and Claude to see how AI researches and shortlists vendors.