What to do when AI gets your pricing or product facts wrong

When an AI engine states the wrong price or wrong product fact about your company, do four things in order: write down the approved fact, capture the answer and the sources it cited, record a human verdict on whether the answer is really wrong, and then decide whether the error traces to a page you own or to someone else's. Fix your own page first. Ask a publisher for a correction only when the wrong claim sits on their page. Neither monitoring nor a correction request can force an engine or a publisher to change anything, so treat every step as evidence gathering and every fix as a hypothesis you re-check later.

In short

• Start from an approved fact written down once, in your own words, before you look at any answer.
• Keep three things apart: what the engine said, what source it cited, and whether that source really says it.
• A person, not the model, decides whether a flagged answer is wrong.
• If the wrong claim traces to your own page, fix it there. If it traces to a publisher, send a dated evidence packet and record the request.
• You cannot compel an engine or a publisher to change; you can only remove wrong evidence you control and watch for a comparable result later.

Why do AI answers get pricing and product facts wrong?

We cannot see inside a model, so we cannot say why a particular answer is wrong. You can observe the answer, the quote that states the claim, and the sources the engine listed. Those point at places to look (an old pricing page, a review site with a stale plan table), but a cited source is an observation, not proof of cause.

A wrong price repeated across engines with the same source attached tells you where to look first. An answer can also be wrong with no source at all, and that is a finding too.

Step 1: Create an approved fact

An approved fact is one plain-text statement about your own brand, written in your words, that you have checked against your own records. Write it before you read the answers, so the answer cannot shape the fact.

In DiscoveredBy's Fact check screen, a fact is one line of text in one of six categories: pricing, company, product, availability, policy or other. The statement is 5 to 300 characters. Owners and editors can add facts, and a project can have 50 active facts at once. There is also a Suggest from my site option that drafts facts from your public pages, and the docs are explicit that a draft's wording is the model's, so you read each one against the sentence shown with it before approving.

Whatever tool you use, write facts a stranger could check: one claim per fact, with the unit and period (per user, per month, billed annually), the effective date, and the page that is your source of truth. Skip claims about competitors; facts are about your own brand.

Step 2: Inspect the answer and its cited evidence

Read the exact sentence the engine wrote and note which source, if any, it attached to that sentence. This is where most of the diagnosis happens.

In Fact check, answers reach the check two ways, as the docs describe it. A weekly fact study asks each eligible engine one question about your brand for each category that has at least one active fact (pricing, company, product, availability or policy; the "other" category is never asked). And checked prompts are tracked prompts you choose (up to 10), whose answers are checked as they arrive. Studies run automatically once a week for projects whose plan includes the feature, and owners and editors can press Run now, with a 24-hour gap after a study that counts. Which engines are asked depends on your plan and on which engines are available; the docs say Google AI Overviews, Google AI Mode and Gemini (app) are left out of the study, so do not assume every engine you track is covered.

For each finding, the screen shows the fact as it was checked, the engine's quote, an AI-written explanation, and the sources the engine cited for that sentence, or "No source was tied to this sentence." A few limits matter here:

• The verdict (supports or contradicts) is the model's judgement. What the product guarantees is that the quote really appears in the answer and the fact is really one of yours.
• Source links are matched to a quote by where the engine placed its citation. A linked source shows where the engine put its citation; it is not a check that the page says what the quote says. Open the page yourself.
• Some engines record no citation positions, so findings from those answers show no sources. The docs name Claude here, and describe a separate limitation for Gemini (API) with non-Latin scripts. Read Citations for how citations are recorded, and see the mentions, citations and visibility explainer for why a citation and a mention answer different questions.
• Only the first 16,000 characters of an answer are checked, and a claim you have no fact for is not reported.

For each cited source, open it and record three things: does it state the wrong fact, does it carry a date, and does it belong to you or to someone else.

Enlarge image

Step 3: Review the finding

A review is your team's verdict on the model's verdict. Owners and editors choose one of four outcomes for each contradicts finding, and the choice changes what the numbers count.

You can add a note (up to 500 characters) and reopen a reviewed finding.

Do not skip the "our fact is out of date" branch. The most uncomfortable outcome of this workflow is finding that the engine is right and your approved fact is stale, for example because a price changed and only some pages were updated. In that case, fix your fact, then fix whatever page still disagrees with it.

If a fact contradiction alert brought you here, the daily watchdog raises one per fact with an open contradiction from the previous 26 hours, and not again for the same fact within 7 days. Opening it takes you to the review queue filtered to that fact. See Alerts.

Step 4: Decide between an owned-page update and an external correction request

Enlarge image

Choose by where the wrong claim can be traced. If a page you own states or implies the wrong fact, or fails to state the right one clearly, update your page. If the cited source belongs to someone else and states the wrong fact, ask that publisher for a correction. If no cited source carries the claim, you have nothing to correct externally, and the useful move is to make the right fact easy to find on your own site.

What goes in an external correction request

A specific, easy-to-verify request is easier to act on than a general complaint, though nothing obliges a publisher to act. Include:

• The exact statement on their page, with its URL and the date you saw it.
• The current official fact, with the URL of your page that states it and its effective date.
• The change you are requesting, worded as a replacement sentence they can paste.
• A plain, courteous tone. You are asking; you have no claim on their editorial decisions.

DiscoveredBy does not contact publishers or engines for you; the docs say plainly that the product shows the contradiction and the cited sources but does not change any answer. Sending and follow-up happen outside the tool. If you already track outreach to cited third-party pages, the Earned sources screen lets you record outreach you handle externally; the docs note it does not send messages or find contacts, and it is built for placements rather than corrections, so treat it as an optional place to keep notes.

The deliverable: a correction log

Keep one row per wrong claim per engine per check. A log makes a slow process visible, and it stops you counting a fix as done before anything was re-checked.

Correction log

ID:                    CL-001
Date first observed:
Fact (approved wording, version, effective date):
Fact category:         pricing / company / product / availability / policy / other
Engine + model:
Origin:                fact study / checked prompt (name the prompt)
Question or prompt asked:
Exact quote from the answer:
Cited source(s) for that sentence (URL, or "none tied"):
Source owner:          us / third party
Does the source page actually state the claim? (yes / no / could not verify)
Source page date:
Review verdict:        confirmed wrong / answer agrees / our fact out of date / not relevant
Reviewer + date + note:
Route chosen:          owned-page update / external request / monitor only
Action taken + date:
Person responsible:
Publisher contact + date sent (external only):
Publisher response + date:
Re-check date planned:
Re-check result:       fixed / still wrong / newly wrong / no comparison / not addressed
Notes on what changed elsewhere in the same period:

Two habits keep the log honest. Record what else changed in the same window (a new pricing page, a press release), because a later "fixed" result may not be caused by your edit. And write "no comparison" or "could not verify" when that is the truth.

The Change column in Fact check gives you a comparison you can copy into the log: it labels each fact Newly wrong, Fixed, Still wrong, Still right or No comparison against the previous comparable study. Per the docs, the comparison is with the most recent earlier study that asked the same question wording in the same language, over the engines with a result in both, and a fact edited between studies gets "No comparison". The column is per fact, not per engine, so it can read "Still wrong" while one engine has improved; keep per-engine results in your log. The docs also note that adding or editing a fact does not re-check answers already checked, so any effect of a page change can only show in a later study or in later checked answers, and the docs do not say how long an engine takes to reflect a change.

Worked example: Quillstone and the wrong seat price

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

Quillstone sells document-review software to legal and compliance teams. Its approved pricing fact, written on 3 March: "Pricing: the Team plan is billed at 40 dollars per user per month when paid annually. Source: quillstone.example/pricing, effective 1 March."

The weekly fact study asks three engines the pricing question. Results:

The Latest study column reads "Wrong on 2 of 3 that addressed it", and the Change column reads Newly wrong, because no engine contradicted the fact in the previous study.

The Quillstone team reviews both findings and opens the sources. The blog post does say 30 dollars, published before the March price change, and it is still live. The roundup page also says 30 dollars, and is undated. Both reviews are marked Confirmed wrong.

The team logs two rows and chooses two routes:

Both rows get a re-check date after the next scheduled study. Suppose that study shows one engine still contradicting the fact and the other now supporting it. The team records "still wrong" and "fixed" on the two log rows (the fact-level Change column would read Still wrong, because one engine still contradicts it), notes the blog update and the pending publisher reply, and does not claim the edit caused the improvement. If the publisher never answers, CL-002 stays open, and that is a legitimate state to report.

Common mistakes and what this cannot tell you

• Treating the model's verdict as final. It is a judgement; read the quote and the source.
• Trusting a linked source. A link shows where the engine placed a citation, not that the page says what the quote says.
• Skipping the "our fact is out of date" check. Sometimes the engine is right and your fact is stale.
• Approving drafted facts unread. A drafted statement is the model's wording of a sentence on your page.
• Declaring victory after an edit. Monitoring cannot make an engine or publisher change, and it cannot show your edit caused a later result. You record a later comparable observation, nothing more.
• Reading silence as correct. Facts are checked only where an answer addresses them, for the engines and prompts in scope. "Not addressed" is not "correct".

Frequently asked questions

Can I make ChatGPT or another engine correct a wrong answer?

No. Neither DiscoveredBy nor any correction request forces an engine to change an answer. What you control is the evidence on your own pages and the accuracy of what you publish. Monitoring then shows whether later answers to comparable questions differ.

How long until a fix shows up?

It cannot be promised. Fact check only re-checks when a new answer arrives, in the next weekly study or in later answers to your checked prompts, and it never re-checks old answers. Whether and when an engine's answer changes after you edit a page is outside your control and outside what the docs promise.

What if the AI answer is wrong but cites no source?

Log it with "none tied", record the exact quote, and make sure your own site states the fact plainly with a date. There is nobody to send a correction request to, so the log entry serves as your baseline for a later re-check.

Do I need a paid plan for this?

Fact check is available depending on your plan; see Plans and limits. Any project member can read the screen when the feature is included, but only owners and editors can change facts, run studies or review findings. You can also do the same workflow by hand with a spreadsheet and a few saved answers.

Can I export the findings for my own log?

The docs describe a Fact checks file export, a customer API route and an MCP tool. All three need the Fact check feature on your plan; the file export also needs data exports, and the API route and MCP tool need customer API access and a key with the right permission. One row per finding carries the quote, the model's and the effective verdict, the source URLs and the review status and note. See Exports and activity.

Next step

Write down your five most important pricing and product facts, then run them through a first study and review what comes back. You can do that in the Fact check workflow, which is part of Brand perception; sign in or create a project to try it, and bring your correction log with you.