How to correct a wrong Gemini claim about a business: evidence, reporting, and follow-up

You cannot edit a Gemini answer. What you can do, in this order, is preserve the exact response as evidence, find the public page that caused the error, correct that page so it is unambiguous and machine readable, report the response through Google's feedback route, and then retest on a schedule and log whether the wrong claim came back. The measurement that matters is a count: corrected claims that stayed corrected, out of the claims retested, with dates. As of 29 September 2026 we have not completed a controlled retest series on Gemini, so this page carries no such count and does not pretend to.

The reason the order matters is that most businesses start at step four. Reporting a response while the page that caused it is still live and still wrong achieves nothing durable.

Preserve the response and identify the claim

Do this before anything else, because the answer you are complaining about may not be reproducible in an hour.

• Save the full text of the answer, not a screenshot of part of it.
• Save every displayed source link, and open each one.
• Record the prompt exactly, including capitalisation and any typo, because a different prompt is a different test.
• Record the conditions: date, time, time zone, surface, the model name shown on screen, whether you were signed in, and whether memory or custom instructions were active.
• Record whether a search step appeared. An answer with no retrieval is a memory problem and cannot be fixed by publishing, which changes what you do next.
• Reproduce it twice more, in clean sessions. A claim that appears once may be variation. A claim that appears three times is the model's position.

Then write the claim down as a single testable sentence. Not "it says wrong things about our pricing" but "it states our starting price is X per month". One sentence per claim, one row per claim, because a bundle of four complaints cannot be scored later.

Find the source of the error

Errors come from a small number of places. Work through these seven in order, because the early ones are the common ones.

• Your own site says it. An old page, a blog post from three years ago, a footer that was never updated. This is the most frequent cause and the easiest to fix.
• Your own site does not say anything readable. A fact that exists only inside a script-rendered element or an image is a fact that is not published, as far as a machine is concerned.
• Two of your own public sources disagree. 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.
• A third-party listing is wrong, and it is being read in preference to you. On the same programme the prices the assistants quoted had come from an app store listing rather than from the company's own site. Claude, 6 August 2026, on clawlaw.in.
• An old press mention or directory entry carries a superseded fact that nobody has ever asked to be updated.
• The model is generalising from your category rather than from a page about you. The tell is that the wrong fact is the category norm, and that no source link supports it.
• A claim of your own could not be true, and the assistant checked it. A headline figure on one 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. This is the worst case, because the correction required is to a claim you chose to make.

Correct the authoritative page

Fix the source before you report the answer. Six things make a correction usable by a machine.

• State the fact in plain text in the page body, not only in an image, a PDF or a script-rendered block.
• Date it. "Effective from 1 October 2026" or "last checked on 29 September 2026".
• Say the old figure is superseded, in one sentence, on the same page. A page that quietly changes its number leaves the old number alive everywhere else with nothing to contradict it.
• Fix every public copy, including listings, app stores, directories and profiles. One stale copy is enough to keep the error going.
• Add the fact where a buyer would ask for it, not only on a corporate page nobody reads.
• Do not remove the old page if it has links pointing at it. Correct it and keep it, so the corrected version sits where the old one was found.

There is one more reason to make the fixed page specific and sourced. Two comparison pages on one site 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. A correct fact published without a source and a date is treated as an opinion.

Report the Gemini response

Google provides feedback controls on responses in its own products, and there are separate routes for legal and privacy matters. Use them, and use them with the same evidence you gathered above: the exact prompt, the exact response, the reason it is wrong, and the URL on your own site that states the correct fact.

Three honest points about reporting. First, we are not aware of any published commitment that a report changes future answers, so treat it as one input rather than a fix. Second, a report supported by a live, dated, machine-readable page is a stronger input than a report supported by an assertion. Third, keep your own record of what you reported and when, because you will need it in the follow-up log and nobody else is keeping it for you.

Test again and document whether it changed

The retest schedule

• Retest at day 7, day 14, day 30 and day 60 after the correction is live.
• Use the exact original prompt, in a clean session, three times per retest date.
• Score one of four outcomes per claim: unchanged, corrected, changed to a different wrong claim, or the claim no longer appears at all.
• Record whether a source link now supports the corrected statement.
• Keep the answers, not only the scores.

The figure this produces, and what it is worth

The figure is claims that were corrected and stayed corrected, over claims retested, with dates. Even with a clean log, it is weak evidence of cause. Between your correction and the retest, the model may have been updated, the retrieval may have changed, and another site may have published something. A sequence is not a cause, and the honest way to write the result is "we corrected the page on this date and the claim read correctly on these later dates", with no "because" in the sentence.

Case log, with what we actually hold

Gemini correction cases retested by us: none, as of 29 September 2026. We publish no correction durability count for Gemini. What we hold is the diagnostic half of the work, on clawlaw.in on 6 August 2026: a pricing page whose prices a crawler never received, two public price lists that disagreed with each other, prices quoted from an app store listing instead of the company's own site, and a headline claim that could not be true and cost the company an answer once an assistant checked it. Those four are real, dated and caused by the site itself, which is where most of these errors start.

What this approach cannot do

• It cannot edit an answer. There is no control panel for what an assistant says about you.
• It cannot promise the correction will stick. We will not claim a durability rate we have not measured.
• It cannot fix a memory-level error quickly. If the wrong fact comes from training rather than retrieval, publishing a page may not reach it for a long time, and no route we know of shortens that reliably.
• It cannot handle defamation or legal harm. That needs the legal or privacy reporting route and, usually, a lawyer, not a content fix.
• It cannot cover errors on other people's pages that they refuse to correct. There you are left with publishing a clearer, dated, better-sourced version of the fact yourself.

Common questions

How fast can a corrected fact show up?

We cannot give you a reliable figure, and the one dated publish-to-citation event we hold is a different thing: ChatGPT cited a named clawlaw.in URL on 6 August 2026, fourteen days after that page was published. That is one page, one assistant, one date, and it is about a new page being used rather than about an old claim being replaced.

Should I report the answer or fix the page first?

Fix the page first, always. A report about a page that still says the wrong thing is asking somebody else to override your own published facts.

What if the wrong claim comes from a review site?

Ask them to correct it, in writing, with evidence. Meanwhile publish your own dated version of the fact, because our capture suggests the sources these answers lean on are mostly official pages and vendors' own pages rather than review sites. Perplexity, September 2026, on aiknowsus.com.

Does deleting the old page help?

Usually not. Deleting removes the corrected statement along with the wrong one, and the old figure survives in other people's copies with nothing authoritative to contradict it. Correct in place and say what changed.

The assistant is right and we do not like the answer. What then?

That is not a correction, it is a disclosure problem, and the honest fix is to publish the limits yourself. A page that says who should not buy from you is quoted more often than most businesses expect, because a careful recommendation needs exactly that material.

Can you get an answer about us changed?

No, and neither can anybody else who is being straight with you. We can measure what is being said, find the pages causing it, and log what happens after you fix them. We cannot guarantee a position or a wording in any assistant's answer.

Case log, limitations, and official links

Use Google's in-product feedback controls on a response, Google's help centre routes for the Gemini apps, and the separate legal and privacy removal routes for content that causes legal harm. Read those pages on the day you use them and note the date, because the routes change.

Our own observations are from the clawlaw.in programme from July 2026 and the aiknowsus.com audit of September 2026. The 6 August 2026 findings about the pricing page, the two public price lists and the app store prices concern pages that were public at the time and can be checked from outside; our counts taken over capture files cannot be, until the captures are published.

Change log. First published 29 September 2026, with no Gemini correction durability figure and the gap stated. If we complete a retest series, the numerator, the denominator, the prompts and the dates will be added here.

What to do first

Take the single wrong claim that costs you the most money, reproduce it three times in clean sessions today, and save all three answers with their source lists. Then find the page that says it, on your site or anybody's, and fix that one page with a date on it. One claim, closed properly, with a log, teaches you more than a long list of complaints filed against a moving target.