Your own pages disagree about your product: run a facts reconciliation review

To reconcile your own pages, list every statement they make about a product, group the statements by fact, compare the values, and record which page is the source of truth for each fact. Then change the pages that disagree. It is a manual editorial review, not something a tool does for you. It matters for AI search because an engine that reads two of your pages may repeat either version, and you cannot tell in advance which. This post gives you the register, the process and a worked example.

In short

• A facts reconciliation review is a manual audit that compares what your own pages say about the same product fact and picks one authoritative value for each.
• Work fact by fact, not page by page. The unit of the register is the fact (price, limit, region, policy), and each row lists every page that states it.
• Rank conflicts by what a buyer would decide on the answer, then fix the page, not the AI answer.
• Once the values agree, put them in an approved-facts list and let monitoring check whether engines state them correctly. Monitoring does not find contradictions between your pages.

Why do your own pages contradict each other?

Pages contradict each other because they are written at different times, by different people, for different jobs, and nothing forces them back into agreement. A pricing page is updated at a plan change. A feature page is updated at a release. A help article is updated when a support agent notices. A launch announcement is never updated at all, because it is dated and feels finished.

The result is drift, not error in any one place. Each page may have been right when it was last edited, yet together they can describe three different products.

This is worth catching before you look at AI answers, because engines may draw on what they can read about you. If two of your own pages give different figures, you may have handed an engine both versions. We can observe what an answer says and which sources it cites (see citations), but that is an observation of one answer, not proof of why an engine chose one figure over another. Cleaning up the inputs is the part you control.

What counts as a product fact worth reconciling?

A product fact is a statement about your own brand that a buyer could check and that has one correct value at a given time. Prices, limits, supported regions, integrations, security claims, refund terms and company details all qualify. Adjectives and positioning ("fast", "trusted") do not, because there is nothing to reconcile.

The approved facts in DiscoveredBy's Fact check use the same idea. It groups facts into six categories: pricing, company, product, availability, policy and other. Those six are a practical checklist for this review too, and the register below uses them. In the product, a fact is one line of plain text between 5 and 300 characters, so a good register entry is a single checkable sentence, not a paragraph.

Skip statements that are legitimately different by context. A page for enterprise buyers and a page for small teams can quote different contract terms without conflicting. Write the scope next to each value so you can tell the difference between a conflict and a segment.

Which pages should you compare?

Compare every public page a buyer or a crawler could plausibly land on, starting with the ones that carry commercial facts. Work through this list and note which of them you own:

• Pricing and plans pages, including any comparison table.
• Feature, product and integration pages.
• Security, privacy and compliance pages.
• Support, help centre and FAQ articles.
• Terms, refund and cancellation pages.
• About, company and contact pages.
• Announcements, release notes and blog posts that state specifics.
• Partner listings and directory profiles that you control.

Do not forget older material. Announcements and blog posts are the usual place where an out-of-date figure survives, because nobody revisits them. Leave out pages behind a login; the review is about what is public.

Enlarge image

How do you run the review step by step?

Run it in six steps, and keep everything in one register so the next reviewer can see what you decided and why.

1. Freeze the scope and date. Pick the list of pages and note the review date. A reconciliation is a snapshot; without a date, nobody can tell later whether a mismatch is new.

2. Extract statements. Read each page and copy every checkable statement about the product into the register, with the page URL and the exact wording. Use the page's own words, not a paraphrase, because a paraphrase can hide a difference that matters ("up to 5 users" versus "5 users included").

3. Group by fact. Give each fact one row label, such as "seat limit on the base plan", and collect every page's statement under it. Two statements belong in the same row when a buyer would treat them as answers to the same question.

4. Compare values. For each row, mark the values as matching, conflicting, or different in scope. Check units and time as well as numbers: monthly versus annual billing, currency, per seat versus per workspace, and effective dates.

5. Choose the source of truth. For each conflicting row, name the one authoritative value and the person who confirmed it (usually product, finance, legal or security, depending on the category). A marketing writer should not guess. If nobody can confirm, mark the row "unverified" and do not publish a value.

6. Fix pages, then record. Edit the pages that disagree, note the date, and re-read them. Then add the confirmed value to your approved-facts list if you monitor AI answers.

If a page states something you cannot support, remove the claim or soften it to what you can support. Do not leave a number in place because it is convenient.

The reconciliation register

Copy this register into a spreadsheet. One row per fact, one column per page group. The last four columns are where the decisions live.

RECONCILIATION REGISTER
Review date:            Reviewer:            Scope (pages/site):

ID | Category | Fact (one checkable sentence) | Page A (URL + wording) |
Page B (URL + wording) | Page C (URL + wording) | Match? (yes / conflict / scope) |
Buyer impact (high / med / low) | Source of truth (value) | Confirmed by (name, date) |
Pages to change | Status (open / fixed / unverified) | Date rechecked

To rank the work, use buyer impact and reach. A wrong price or a missing compliance statement on the page most buyers read comes before a stale figure in a two-year-old blog post. But do not skip the old post entirely: if it is public and states a specific, decide whether to correct it, add a dated note, or retire it.

Illustrative example: Quillstone

Illustrative example: Quillstone and its competitors are fictional, and the numbers are made up to show the method.

Quillstone sells document-review software to mid-sized legal and compliance teams. Its team reviews seven public pages: pricing, features, security, help centre FAQ, refund policy, about, and a launch announcement from last year. They extract 21 statements and group them into 9 facts.

Five facts turn out to conflict, three match, and one is a segment difference (the enterprise page quotes a longer contract term than the base plan, and says so). Here are the five conflicts.

Arithmetic check: 5 conflicts, 3 matches and 1 segment difference make the 9 facts. Of the 5 conflicts, 3 are high impact, 1 is medium and 1 is low.

The team fixes the three high-impact rows first. The help FAQ and launch post are corrected for F1, the pricing page FAQ for F2, and the about page for F3. F4 needs a features-page edit, and F5 gets a one-line correction in the launch post. Each row moves to "fixed" with a date. Nothing in this example says whether AI answers improve afterwards; that is what monitoring is for.

Enlarge image

Where does DiscoveredBy help, and where does it not?

DiscoveredBy helps once you have agreed values: it checks whether AI engines state them correctly. It does not compare your pages with each other, so the reconciliation itself stays a manual job. The docs describe no feature that detects contradictions across your website, and this post does not imply one.

What the Fact check screen does is hold a list of approved facts about your own brand and check AI answers against them. A weekly fact study asks the chat engines available on your plan, plus ChatGPT (app), one question per category about your brand (the Google AI surfaces and Gemini (app) are not asked), and up to 10 tracked prompts you choose are checked as their answers arrive. Findings show the engine's quote, an AI-written explanation, and the sources the engine cited for that sentence, where there are any. A finding is the model's verdict; your team reviews the ones marked as contradictions.

Two related capabilities are worth knowing about for this review:

• Suggest from my site reads your homepage plus up to 7 pages linked from it whose path contains words such as pricing, plans, about, features, product, faq, terms, refund, security or contact. It drafts proposed facts, each with the page it came from and the sentence it quoted. It is a shortcut for the extraction step. It does not compare pages, and it will not read an announcement post unless the homepage links to it and the link path matches one of those words. Read each draft against its sentence before you approve it, because the statement is the model's wording.
• Editing a fact gives it a new version, and existing findings keep the wording they were checked against. So once your reconciliation changes a value, edit the approved fact, and expect its tracked accuracy to start again. Answers that were already checked are not re-checked against the new wording.

Approved facts are limited (50 active at a time) and are about your own brand; claims about competitors are not checked. Claims that match none of your facts are not reported either, so a fact must be on the list before it can be checked. Fact check is not available on every plan; see plans and limits.

What this review cannot tell you

A reconciliation review shows that your public pages agree with each other, and nothing more. Be clear about the limits before you report it as done.

• Agreement is not accuracy. Three pages can agree and all be wrong. The "confirmed by" column is what protects you.
• It does not show what engines say. Engines may draw on third-party pages, older copies or their own training. Consistent pages are a precondition, not a guarantee.
• It does not change an answer. Fixing your pages cannot force any engine or publisher to update. For a wrong statement on someone else's site, see what to do when AI gets your pricing or product facts wrong and requesting a correction from a publisher.
• It is a snapshot. New releases reopen old rows. Rerun it on a schedule and after any pricing or policy change.
• It skips what you cannot see. Cached copies, syndicated content and partner pages you do not control are outside the register unless you add them.

Common mistakes

• Reconciling page by page. You see each page as internally fine and miss the cross-page conflict. Group by fact.
• Letting the newest page win by default. The latest edit might be the wrong one.
• Fixing the visible pages and forgetting the announcement archive.
• Skipping units and dates, so "$49" and "$49 billed annually" look identical.
• Adding a fact to your monitoring list before the value is confirmed. You then measure engines against a guess.
• Treating a segment difference as a bug and flattening it, which removes information a buyer needs.

Frequently asked questions

How often should we run a facts reconciliation review?

Run a full review at least once a quarter, and a targeted one whenever pricing, plans, policies or regions change. The full pass catches slow drift; the targeted pass keeps a launch from creating new conflicts on the day it ships.

Who should own the register?

Give it one editorial owner, usually in content or product marketing, but let the confirming owner differ by category. Finance confirms pricing, legal confirms policy, and security confirms compliance and data claims. The register owner chases confirmations; they do not decide the values.

Can a tool find these contradictions automatically?

Not in DiscoveredBy. Its Fact check compares AI answers with facts you approved, and "Suggest from my site" drafts facts from a limited set of your pages, but neither compares your pages against each other. Use it after the manual pass to keep watching the agreed values.

Should we fix the AI answer or the page first?

Fix the page first. Your pages are the part you can change, and a wrong answer that comes from your own contradiction will not clear until the contradiction does. Then monitor whether the answers change, without expecting a particular timeline.

What if two teams disagree on the correct value?

Escalate to whoever owns the underlying decision, such as the pricing owner, and record their answer with a date. Until then, mark the row "unverified" and consider removing the claim from public pages rather than publishing a guess.

Next step

Once your register has confirmed values, add them as approved facts and let DiscoveredBy check how engines describe your product. Start in Fact check; a related read is Where do AI answers contradict approved product facts?. If you are planning a larger change, see monitoring before and after a rebrand. Sign in or create an account to add your first approved facts.