How to correct wrong company facts in Claude: source diagnosis, reporting steps, and a follow-up log

You cannot edit what Claude says. What works is a sequence: capture the answer as evidence, diagnose which public source produced the error, correct that source so it is dated and machine readable, report the response to Anthropic, then retest on a fixed schedule and log what happened without claiming that your fix caused it. The figure that matters is the count of documented cases that changed in follow-up tests out of all cases retested. As of 29 September 2026 we hold no such count for Claude, so this page publishes the diagnosis method, the log format, and the four dated cases where we did find the cause.

Most of those four causes were on the company's own site. That is the useful and unwelcome finding: the error usually starts with something you published, or failed to publish readably.

The answer, first

Five steps in a fixed order, and the order is the point.

• Capture the exact answer, prompt, conditions and source links, and reproduce it twice more in clean sessions.
• Diagnose which public source says the wrong thing, using the seven causes below.
• Correct that source, with a date, in plain text, and fix every other public copy of the fact.
• Report the response through Anthropic's feedback route, with the corrected URL attached.
• Retest at day 7, 14, 30 and 60, three runs each, and log four possible outcomes per claim.

How it was measured

The capture, before anything else

Save the full answer text, every displayed source link, the prompt exactly as typed, the date, time and time zone, the surface, the model name shown on screen, and whether a search step appeared. That last field decides your route: a wrong claim with no retrieval behind it is a memory problem, and publishing a corrected page will not reach it quickly.

Then reproduce it twice in clean sessions. One appearance may be variation. Three appearances is a position.

The seven causes, in the order to check them

• Your own page says it, somewhere old.
• Your own page does not say it readably. The fact exists in a script-rendered element, an image or a PDF, so a machine never received it.
• Two of your public sources disagree.
• A third-party listing is being preferred to you.
• An old press item or directory entry carries a superseded fact.
• The model is generalising from your category, which shows as a wrong fact that is the category norm with no citation attached.
• A claim of yours could not be true and was checked.

The correction, made so a machine can use it

• Plain text in the page body, not only in an image or behind a script.
• A date: effective from, or last checked on.
• One sentence saying the previous figure is superseded, kept on the page.
• Every other public copy updated: listings, app stores, directories, profiles, partner pages.
• A source and a date beside any figure about anybody else, because an unsourced figure is read as advocacy even when it is right.
• The old URL kept and corrected rather than deleted, so the fix sits where the error was found.

The follow-up log

One row per claim per retest date, with four possible outcomes: unchanged, corrected, replaced by a different wrong claim, or no longer mentioned. Record whether a citation now supports the corrected statement, and keep the answers themselves, not only the scores.

The figure this produces is claims corrected and still correct at the last retest, over claims retested, with all the dates. Write it without the word "because". Between your fix and the retest, the model may have changed, retrieval may have changed, and somebody else may have published something.

What the numbers were

Correction cases we have retested on a schedule: none, as of 29 September 2026. There is no durability count on this page. What we have is the diagnostic half, and it is dated.

• Cause two, our own page unreadable. A pricing page rendered its prices only after scripts ran, so what a crawler received contained no prices at all, and the prices the assistants did quote had come from an app store listing instead of from the company's own site. Claude, 6 August 2026, on clawlaw.in.
• Cause three, two public sources disagreeing. The website and the app store listing carried different plan names and different prices for the same product, and ChatGPT noticed the contradiction and said so in its answer. ChatGPT, 6 August 2026, on clawlaw.in.
• Cause seven, a claim that could not be true. A headline figure on the main site could not be true, and when an assistant checked it against public numbers it advised a buyer against the product, after which the company's accurate claims stopped counting for that answer. Claude, 6 August 2026, on clawlaw.in.
• The credibility version of the same problem. Two comparison pages stated competitors' prices with no link, no date and no source. The figures turned out to be correct when the assistant checked them independently, but it had no way to know that at read time, so it treated the pages as advocacy and used official sources instead. Claude, 17 September 2026, on clawlaw.in.

Three of those four are faults in what the company published. That is why the diagnosis step sits before the reporting step, and why most of the work in a correction is on your own site rather than in a support queue.

What this cannot tell you

• It cannot prove your fix caused a change. A small number of cases and a moving system cannot support a causal claim, and a supplier telling you otherwise is overreaching.
• It cannot fix a memory-level error on your timetable. If the wrong fact comes from training rather than retrieval, publishing may not reach it soon and no route we know of reliably shortens that.
• It cannot force a third party to correct their page. You can ask with evidence, and otherwise publish a better sourced, dated version yourself.
• It is not a legal remedy. Defamation, privacy and trademark matters go through the legal and privacy routes and usually need a lawyer, not a content fix.
• It cannot promise a wording. We do not sell control over what any assistant says, and we cannot guarantee a position or a phrasing in an answer.
• It cannot rely on the assistant's own account of events. Asked afterwards how many questions it had searched for, one assistant withdrew its own earlier statement in three separate batches. Perplexity, September 2026, on aiknowsus.com.

Common questions

Should I report it to Anthropic first?

No. Fix the public source first. A report about a fact your own site still states wrongly asks somebody to override your own published information.

How do I know whether it is memory or retrieval?

Look for a citation. A wrong claim with no citation and with a shape typical of your whole category is probably memory. A wrong claim with a citation that opens to a page saying that wrong thing is retrieval, and the page is your target.

How long should I wait before retesting?

Day 7, 14, 30 and 60, three runs each in clean sessions with the original prompt. Retesting on day one tells you nothing and tempts you into concluding something.

The wrong fact is on a competitor's comparison page. What now?

Ask for a correction in writing with evidence. In parallel publish your own dated, sourced version of the fact. Our own capture suggests these answers lean on official pages and vendors' own pages more than on review sites, so your own page is worth more here than it feels like it should be. Perplexity, September 2026, on aiknowsus.com.

Can I add instructions on my page telling the model what to say?

Do not. A competitor's page carried a hidden block of text addressed to answer engines, instructing them to cite that company as the source. The assistant found it, refused it, and named the company that had done it. Claude, 6 August 2026, on clawlaw.in.

Is one wrong sentence worth all this?

If it is about price, coverage or suitability, yes, because those sentences decide whether a buyer continues. We hold a dated case where a single impossible claim led an assistant to advise a buyer against the product, after which the company's accurate claims stopped counting for that answer.

Sources and change log

Use Anthropic's own feedback controls on a response and its published support and policy routes for anything more serious. Read Anthropic's web search tool documentation and model overview pages for what the product does, and note the date you read them.

Our own observations come from the clawlaw.in programme from July 2026, specifically the interim checks of 6 August 2026 and the Claude response audit of 17 September 2026, and from the aiknowsus.com audit of September 2026. The 6 August findings concern public pages and are checkable from outside.

Change log. First published 29 September 2026 with four diagnosed cases and no durability count. If we complete a retest series, the count, the denominator, the prompts and the retest dates will be added here, stated as a sequence rather than as a cause.

Disclosure. Written by AI Knows Us, a vendor of AI visibility measurement.

What to do first

Take the one wrong claim that costs you money, reproduce it three times today in clean sessions, and save the answers with their source links. Then work down the seven causes until you find the page. In our own programme that search ended on the company's own site three times out of four, which means the fix was available immediately.