An AI agent website checklist for product facts

By Ding Xiang · Published 2026-11-13 · Updated 2026-11-13

Before an agent can describe your product correctly, the facts have to exist in one consistent version, in public, where a machine can read them. This checklist is the preparation we work through before every audit — the same steps, in the same order, whether you do them yourself or hand them to us. Nothing here promises better recommendations. It removes the avoidable reasons an agent gives a wrong price, a wrong scope or a wrong contact address with complete confidence.

1. Inventory pricing, scope and contact sources

List every place your product facts are stated: the pricing page, the checkout page, the FAQ, the terms, the app-store or marketplace listing, the last announcement post. For each source, write down the price, currency, billing basis, scope limits and contact address it shows, with the date you checked it. Most sites discover at least one forgotten surface at this step — an old launch page, a directory listing, a PDF. You cannot fix a contradiction you have not listed, and an agent will find the stale copy even if you have forgotten it.

2. Identify conflicting or missing facts

Compare the inventory line by line. The usual conflicts: a launch price that was never updated after the introductory period, a “per month” on one page against “one-time” on another, a scope limit that exists in the terms but not on the pricing page, a support address that changed in one place only. Missing facts matter as much: no delivery window stated anywhere, no refund wording, no currency next to the price. Decide which version is true — the one checkout actually charges — and mark every other copy for correction in step 5. Do not average conflicting prices; pick the real one.

3. Check public accessibility and file existence

Facts behind a login, a script wall or a bot block do not exist for an agent. Open each key page in a private window, signed out, and confirm the price and scope are readable as served HTML. Then check the three root files: /llms.txt, /pricing.md and /robots.txt — do they return HTTP 200 as the right content type, or a redirect, or an error page? Our free file check runs exactly this presence test. While you are in robots.txt, confirm it is not disallowing the very pages you want read.

4. Record model, prompt, date and verbatim output

Now ask the questions a buyer would ask: “How much does [your product] cost?”, “What does it include?”, “What are the alternatives?” Record, for each answer, the model, the exact prompt, the date, and the answer pasted verbatim — never a summary from memory. Verbatim records are what make the next step possible: you can point at the sentence that is wrong and trace it to the page it probably came from. In our own audit reports this record is the core evidence, with the snapshot date printed on the cover, precisely because model answers drift and an undated claim about them is worth nothing.

Keep the record as one table — one row per question, one column per check — so the next run, whether in a month or after a price change, is directly comparable with the last. When we re-run a simulation for a Plus recheck, this table is exactly the comparison we send: same prompts, new date, new verbatim answers, with the differences visible at a glance instead of reconstructed from memory.

5. Prioritise corrections without claiming recommendation gains

Fix in this order: wrong price first, then wrong scope, then missing delivery and refund facts, then stale secondary pages. Correct the fact at its source page before adding any new file — a pricing.md that contradicts your pricing page helps no one. When the sources agree, publish the machine-readable copies (llms.txt and pricing.md) from the corrected facts, and re-run step 4 to confirm the record improved. What you can claim at the end is narrow and honest: your facts are consistent, public and readable. Whether that changes what a model recommends is the model's business, and any checklist that promises otherwise is selling you the one thing it cannot deliver.