Skip to content
Etalo

Method

When AI gets you wrong

A brand that is highly visible in assistant answers, with a wrong price, does not have a visibility problem. It is the one case where a good score is worse than a bad one — and the one visibility grids cannot see, because they count citations without reading what they say.

Founder of Etalo, SEO and generative visibility consultant

Read on 17 September 2026. Every study cited carries its date in the text.

The essentials in thirty seconds

There is today no form, no channel and no documented turnaround for having a factual error about your company corrected inside an assistant's answer. The only method that works is repairing the sources the model reads — your site first, then the third-party pages it cites — and then checking that the answer changed. How long it takes depends on where the error comes from: an error from a web page is fixed in days, an error from the model's memory waits for its next training run.

  • Inaccuracy is frequent and confident: across 1,600 queries put to eight AI search tools, more than 60% of answers attributed a source incorrectly, with huge gaps between tools — 37% for one, 94% for another (Tow Center, Columbia, February 2025).
  • The worst part is not the error, it is the confidence: in that study one tool got 134 of 200 wrong while signalling doubt just fifteen times, and never declining to answer.
  • Two kinds of error, two timelines: what the model reads on the web is fixed at the speed of indexing; what it "knows" from memory only changes at the next training run, over which you have no control.
  • A Google tool exists, but is widely misunderstood: the outdated content tool only applies to pages you do not own, and only once the third-party page has actually been changed or removed.
  • What can be steered: one up-to-date reference page of your own, consistent facts everywhere else, and a measurement that checks answers against reality rather than counting mentions.

The case where a good score is worse than a bad one

Every generative visibility grid answers the same question: do you appear? None asks the next one: is what is said accurate? Yet the two lead to opposite actions.

A brand absent from the answers has an awareness or readability problem: it produces content, opens its crawlers, waits. A brand highly present with an outdated price has a different and more urgent problem: every extra citation makes the damage worse. The score rises while the harm spreads.

The errors that cost deals are always the same: a price that no longer exists, a discontinued offer still attributed to you, a wrong location, an executive who left two years ago, a certification you do not hold, or — most common — an isolated forum opinion repeated as if it described your company.

How often tools get the citation wrong

Share of answers attributing a source incorrectly, across 1,600 queries: twenty publishers, ten articles each, eight tools, in February 2025.

Perplexity 37%
Average of eight tools over 60%
ChatGPT Search 67%
Grok 3 94%
Source: Tow Center for Digital Journalism, Columbia, February 2025. What this study measures: the ability to trace the source of a news excerpt. What it does not measure: the accuracy of what is said about a company. Transposing it as it stands would be dishonest; what it establishes is that error is frequent and stated with confidence.

Two errors with nothing in common

Before any correction you need to know where the error comes from, because the two origins are neither repaired the same way nor in the same timeframe.

A retrieval error: the model read a page, and that page says something false, outdated or badly interpreted. This is the most common case and the only genuinely repairable one: fix the page, the engine reads it again, the answer changes. You recognise it because the answer cites a source — and opening it, you find the error.

A memory error: the model asserts without citing, from what it learned during training. No change to your site will fix that in the short term: what it learned is frozen until its next training run, whose date and contents you do not know.

The test that separates them takes one query: ask the same question with and without web search. If the answer changes, the error comes from a page. If it persists with no source at all, it comes from memory — and your job is no longer correction, it is making the truth so present on the web that it eventually weighs.

The procedure, in order

Six steps. The first two take an hour, and the last is the one everybody skips.

01

Record it properly, several times

An assistant's answer is not reproducible: a single query proves nothing. Ask the question five times, on each surface, and keep the raw answers with their date. Without those readings you will not be able to show the correction happened.

02

Identify the origin

With or without web search, with or without a cited source. That is what determines whether you face a repairable error or a training legacy.

03

Repair the source of truth, on your own site

One page stating the truth, dated, unambiguous, and not leaving the old version lying elsewhere on the site. If the information lives on three pages in three versions, you have manufactured the contradiction the model repeats.

04

Get third-party pages corrected

Directories, company listings, press articles, partner pages, collaborative encyclopedias. This is the long part, because it depends on people you do not command — and on encyclopedias, conflict-of-interest rules forbid editing your own entry: you go through the talk page, with sources.

05

Request a refresh, where possible

Google offers an outdated content tool, but it has two strict conditions: it only applies to a page you do not own, and only if that page has been removed or substantially changed. It does not correct generated text: it forces an update of the search result feeding it.

06

Check again, and keep the trail

Ask the same questions, the same number of times, and compare with the starting readings. It is the only way to know whether the work paid off — and the only evidence you can put in front of a meeting.

Five concrete cases, and what actually works

All met in audits. The timeframe given is the lever's, not a promise of outcome.

An outdated price repeated by three assistants
Almost always a retrieval error: an old pricing page, a forgotten PDF brochure, or a comparison site that never updated. Fix the page, pull the PDF, write to the comparison site. Days to weeks.
A discontinued offer still attributed to you
The trap is deleting the page: a model that finds nothing keeps what it knew. Better to keep the page, stating clearly that the offer ended, with the date. You are handing the model the sentence that corrects it.
A forum opinion presented as fact
The hardest, because the source is legitimate in the engine's eyes and you do not control it. What works: publishing a factual, dated, verifiable answer that becomes the best available source on the subject.
An executive who is no longer there
Often a memory error, kept alive by directories and old articles. Consistency everywhere — site, listings, professional networks, press releases — then patience: this is the kind of error that fades at the pace of training runs.
Confusion with a same-named company
Frequent and underestimated. The lever is the entity: an unambiguous about page, consistent public identifiers, and outbound links tying your name to your industry rather than the other one's.

And if the statements are damaging

A factual error and a statement that causes harm are different things. The first is handled as above. The second falls under a legal framework that depends on your country, the nature of the statements and their reach, and we are neither lawyers nor in a position to advise you on it.

What we can speak to is method, not law: in that situation the value of measurement changes in nature. Timestamped, repeated readings that preserve the raw answer and the exact question amount to a dated record — and that is generally the first thing an adviser will ask for. An isolated screenshot of a non-reproducible answer is worth very little.

Three markers for deciding

5

minimum readings per question before concluding an error is systematic. Below that you are measuring chance: an assistant's answer is not reproducible.

Scoring rule applied by Etalo

2

possible origins, and only one is quickly repairable — the one coming from a page the engine went and read.

Retrieval or memory

0

public procedures for correcting a factual error about a company, across assistant vendors. It is the most uncomfortable finding in this article.

Read on 17 September 2026

What Etalo does here, and what it does not

Etalo checks the claims in the answers against the reality of the offer, and produces the list of inaccuracies ranked by severity, with the question, the surface and the date of each finding. An inaccuracy verdict not reviewed by a human blocks the report export: accusing an answer of being wrong is too heavy a claim to publish on the word of one model judging another.

What it does not do: it corrects nothing for you, contacts no vendor, and has no privileged channel — nobody has one. Nor does it replace legal advice. It establishes a dated, reproducible record, and that is already what is most often missing when the subject reaches the table.

Frequently asked questions

Can I ask a vendor to correct an answer?
No public procedure allows it today for information about a company. Feedback links exist in the interfaces, with no documentation on how they are handled or how long it takes.
How long before a correction shows?
If the error comes from a web page, the delay is re-indexing — days to weeks. If it comes from what the model learned, it waits for a new training run, whose date is not public.
Should I delete the page carrying the old information?
Rarely. A deleted page does not correct what has already been learned, and deprives the engine of the sentence that states the truth. Better to keep it, clearly indicating what changed and since when.
Is Google's outdated content tool for this?
Only partly. It covers pages you do not own, and only after the page has been removed or substantially changed. It speeds up the update of a search result; it does not rewrite generated text.
Is an error on a single surface serious?
It depends which one and on your market. An error in AI Overviews reaches people who have never opened an assistant; the same error on a niche surface does not weigh the same. One more reason to measure surface by surface rather than on average.
How do I prove the error was there?
By keeping, for each finding, the exact question, the raw answer, the surface, the mode and the timestamp — and by repeating. A reproducible record is worth infinitely more than a screenshot.
  • Why my brand doesn't appear in ChatGPT

    You asked the question, and a competitor came out instead of you. Before rewriting anything, work out which of the four causes applies: they share neither the…

  • How much does a GEO audit cost

    Two of the three most-cited tools no longer show any price at all. Here are the ones still public, what the API calls behind them really cost, and why the gap…

  • AI prompt volume does not exist

    Every AI visibility tool ends up printing a number next to each question: "2,400 searches a month". Nobody measures that. Here is where the figure actually…

  • Getting into Google AI Overviews

    This is the generative surface most people see without ever looking for it: it sits above the results, in front of users who have never opened an assistant…

  • Building your question corpus

    It is the part of a GEO audit nobody wants to do, and it decides everything else. A badly written corpus produces impeccable figures about questions nobody…

  • GEO audit: what it measures, how, and what makes an audit wrong

    Public figures on generative search contradict each other: AI Overviews appear on "more than 20%" of searches by one measurement, on 75.24% by another. Both…

  • How to get cited by ChatGPT

    There is no button. There is a technical door that many sites close without knowing it, three editorial levers that have actually been measured, and a source…

  • GEO or SEO: what really changes

    Three serious studies measured the same thing — the share of AI citations coming from pages that rank well on Google. They found 16.7%, 52% and a correlation…

  • AI visibility tracking tools in 2026

    Seven tools, read on the same day from their public pages. Two publish no price at all. And once you divide published prices by the number of answers actually…

  • robots.txt and llms.txt: letting AI read your site

    It is the one generative-visibility job that can be settled in an afternoon, and the one that makes every other job pointless if you get it wrong. Three files…

  • Etalo.ai review

    You are looking for reviews of Etalo and you cannot find any. That is normal: the tool is young and has no clients to parade. Rather than manufacture…

  • Best Peec.ai alternative

    Peec.ai does one thing well: it tracks your visibility in assistant answers, daily, in a dashboard that is easy to read. If that is what you need, it is a good…

  • Best Promptwatch alternative

    Promptwatch has one real strength few tools have: crawler logs, that is, what AI robots actually come and read on your site. If that is your subject, keep it…

Sources

Citation error rates: Tow Center for Digital Journalism, Columbia Journalism Review, February 2025 — 1,600 queries, eight tools.

Conditions for the outdated content tool: Google Search Console help, read on 17 September 2026.

How Google's generative answers work, and the absence of specific markup: official guide, May 2026.

Know what they say, and whether it is true

A demonstration on your brand: what the ten engines answer, and the list of claims that do not match your offer.