How to build a comparison site with AI agents

Last updated October 14, 2026 · By Orenston

Short answer: To build a comparison site with AI agents, design the comparison data first, have the agent collect every fact from the brand's own pages with a source and date, generate all pages from that one data file, put the answer and table at the top, route every affiliate link through one tracked redirect, disclose next to the recommendation, and keep thin pages out of the index. A person reviews before anything is published.

This guide comes from comparison sites we have built ourselves, including a telehealth price-comparison site and orenston.com's own product database, and from the FTC's and Google's published guidance. It does not promise rankings or income.

The ten steps

1. Pick one comparison intent. Choose a question buyers actually search, such as "best home sauna under $5,000" or "cheapest semaglutide program". One page answers one question. Keep unlike things apart: on a telehealth price site we run, oral, injectable, brand-name and compounded treatments are never ranked against each other.
2. Design the data before the pages. Write a schema for the facts you will compare: product, brand, price components, billing terms, availability, source URL and verified date for each fact. Our telehealth build stores intro, recurring, upfront and effective monthly price as separate fields, because merging them ranks a 12-month prepay above a cheaper month-to-month plan.
3. Collect facts only from primary sources. Read prices, specs and policies from the brand's own pages and record the URL and date for each one. If a fact is not published, record that ("price not disclosed") instead of guessing. Never fill gaps with invented numbers, reviews or ratings.
4. Generate every page from one data file. Keep facts in one source file and build pages, the public JSON feed and the sitemap from it with a script. On orenston.com, products.json, product pages and sitemap.xml all come from one build script, so a hand edit to a generated file gets wiped on the next build.
5. Write the comparison answer first. Put the short answer and the comparison table near the top, in server-rendered HTML. Show price type and billing terms beside every price, a plain-language methodology, caveats and the date facts were checked. Explain why each pick wins for a specific use, as Google's reviews guidance asks.
6. Route every affiliate link through one place. Send each outbound commercial link through a single redirect or component that logs the click and adds rel="sponsored". Orenston uses /go/<slug> links that record the click and 302 to the brand, with noindex on the redirect. Never let commission size decide the order of a ranking.
7. Disclose next to the recommendation. Add a clear disclosure beside the comparison table and links, in the same language as the page. A footer-only or linked "Disclosure" button is not enough under FTC guidance.
8. Use conservative structured data. Use WebSite, BreadcrumbList, Article or CollectionPage and FAQPage only where the visible page supports them. Do not add Product, Review or rating markup that claims hands-on testing you did not do.
9. Gate what gets indexed. Mark pages noindex until they are genuinely useful, and generate the sitemap from the same rule, so a page enters the sitemap when it qualifies and leaves when it stops.
10. Ship, submit, then keep facts fresh. Submit the full sitemap URL in Search Console and Bing Webmaster Tools, request indexing for the strongest pages, and recheck prices on a schedule. Show the "last checked" date on the page; a stale price hurts trust faster than a missing one.

Copyable prompts for agents

Data schema

You are building an affiliate comparison site for [NICHE]. Propose a JSON schema for the facts buyers compare. Every fact needs a source_url and verified_at field. Keep intro price, recurring price and upfront price as separate fields. Do not include commission data in the public schema.

Fact collection

For each of these brand URLs: [URLS], extract only facts stated on the page into the schema above. Quote the source sentence for each value. If a value is not stated, set it to null and add a note. Do not estimate, infer or invent any value.

Comparison page draft

Using only the attached JSON, write a comparison page for the query "[QUERY]". Start with a two-sentence answer, then a table, then why each option is best for a specific buyer, then methodology and caveats. Put price type and billing terms beside every price. Add this disclosure above the table: "We earn a commission if you buy through links on this page."

Pre-publish audit

Audit this page against its data file. List any claim without a source, any price shown without its billing terms, any superlative without a reason, and any missing disclosure. Output a pass/fail list; do not rewrite the page.

Pre-launch checklist

• ☐ One search intent per page, with a descriptive H1
• ☐ Data schema with source_url and verified_at for every material fact
• ☐ No invented prices, specs, reviews or ratings; unknowns labelled as unknown
• ☐ Like-for-like comparisons only
• ☐ Pages, JSON feed and sitemap generated from one data file
• ☐ Answer and comparison table near the top, in server-rendered HTML
• ☐ Price type and billing terms beside every price
• ☐ Methodology, caveats and a visible "last checked" date
• ☐ Every affiliate link through one tracked redirect with rel="sponsored"
• ☐ Ranking order never driven by commission
• ☐ Disclosure next to the comparison, not only in the footer
• ☐ Conservative schema that matches visible content
• ☐ noindex on thin pages; sitemap generated from the same rule
• ☐ Human review of AI drafts before publishing
• ☐ Sitemap submitted to Google and Bing; strongest pages requested for indexing
• ☐ Scheduled fact rechecks

What we learned building our own

• FAQ rich results are rare now. Our telehealth build found valid FAQPage markup produced no rich result, because Google limits FAQ rich results to authoritative government and health sites. Keep FAQs for readers, not for snippets.
• Request Indexing has a per-account daily quota. The quota is shared across every property in the Google account, so a second site can use it up. Spend it on the homepage, hubs and best comparisons first.
• "Discovered, currently not indexed" is normal at first. New comparison sites often see the homepage indexed within about a day while deeper pages wait in the queue. Check that the live URL test passes, then wait.
• Submit the full sitemap URL. On orenston.com Search Console rejected a bare "sitemap.xml" and accepted https://orenston.com/sitemap.xml.
• Return a real 404. Without a 404 page, our static host served the homepage with a 200 status for any URL, which creates soft 404s.
• Let one agent deploy. Two agents deploying the same site overwrote each other's builds on orenston.com. Now one agent publishes and others hand it patches.
• Keep private economics out of the page. Commission rates and payout terms stay in a private config; the public page never needs them to render a comparison.

What not to do

• Don't mass-generate a page for every product pair or city. Google's spam policies call out scaled content and thin affiliate pages.
• Don't claim hands-on testing you didn't do. The FTC says you can't talk about your experience with a product you haven't tried.
• Don't call an old screenshot, network listing or referral link a current offer.
• Don't publish AI drafts without checking every fact, including titles, descriptions and structured data.

Want products to compare? Browse the Orenston product database, where each product has a tracked link, instructions and prompts. See also how to disclose affiliate links in AI content.

Browse the product database All guides