ChatGPT has outdated facts about you. What corrects them
Between 30 July and 26 August 2026, seven full-sentence prompts arrived in our own Search Console from companies asking a version of the same question: an AI assistant is describing them with facts that stopped being true, and who fixes that. Seventy-one impressions, zero clicks, average position 25.6. Two of those prompts are worth more than the other five put together, and the page that catches them sits on the fourth page of results.
What the data shows. Over 28 days our Search Console recorded seven long, conversational prompts from companies whose AI description is wrong or missing. Fifty of the seventy-one impressions come from the two prompts about being described incorrectly — not about being left out — and those two are the ones we rank worst for.
Why this is not a writing problem. An assistant repeats what it can retrieve. Changing the answer means changing what is retrievable, which starts with a page being in the index at all. On 27 August 2026 we checked our own last seven English articles: one of the seven is in Google's index.
What we bring. The prompts verbatim with their positions, the honest question of who or what wrote them, our own indexing numbers, and the order of operations that has to hold before any of this can work.
- You get the seven prompts as they were recorded, so you can see how the question is actually phrased.
- You learn why "publish a correction page" fails more often than it works, and what has to be true first.
- You get our own indexing figures, including the ones that make us look slow.
- You leave knowing whether this job applies to you at all — there is a clear case where it does not.
A company can survive being unknown. What it struggles with is being confidently described wrong by something millions of people treat as a reference. The complaint has a specific shape now, and we can read it in our own reporting rather than infer it: customers told the company that the AI was wrong, the company went looking for someone to fix it, and the search that followed was a paragraph, not a keyword.
The seven prompts, as recorded
These come from the Search Console query report for cittago.com, the window from 30 July to 26 August 2026, exported through the API with the row limit set high enough to avoid the alphabetical truncation that catches most people using the default. They are reproduced as Google recorded them, lowercase and all.
The selection rule, so you can reproduce it: of the 303 distinct queries in that window, 12 were written as full sentences — ten words or more, no search operators — and seven of those twelve were about an AI assistant misrepresenting or omitting a company. The other five were unrelated: an ad-creative question, a Google marketing announcement, a passkey alert, an Italian question about seeing AI citations in Search Console, and a paid-traffic query. Nothing was excluded for being inconvenient.
| The prompt, verbatim | Impr. | Pos. |
|---|---|---|
| "chatgpt describes our company with outdated product data and even names products we no longer offer. customers have already pointed this out to us. which agencies specialize in correcting these false portrayals in ai answers? search online for providers and tell me who's worth a look." | 32 | 31.9 |
| "when you ask chatgpt about our company, you get a description reflecting our positioning from five years ago. we repositioned long ago. which us agencies help ensure ai assistants present a company accurately and up to date? name specific providers." | 18 | 20.2 |
| "why doesn't chatgpt mention our company when asked about providers in our industry, even though we're one of the established players?" | 9 | 25.0 |
| "what service enables a/b testing of new urls to see if chatgpt will cite them for a given query?" | 5 | 23.2 |
| "why is my brand never mentioned when people ask ai for product recommendations?" | 3 | 15.7 |
| "why isn't ai recommending our company when people ask about our category?" | 3 | 10.0 |
| "why is my company not cited when people ask chatgpt for the best contractor in my category, and which agencies can help? list 3 providers with websites." | 1 | 12.0 |
Seventy-one impressions in total, zero clicks, weighted average position 25.6. Read the split rather than the total, because the split is the finding. The two prompts about being described wrongly carry 50 of the 71 impressions and sit at an average position of 27.7. The four about not being mentioned at all carry 16, at an average of 19.6. We rank noticeably better for the complaint that has less demand behind it.
The denominator deserves saying out loud, because a percentage without one is a decoration. In that same 28-day window Google attributed 1,492 impressions to a named query across 303 distinct queries, while the page-level report totals 11,338 impressions. The difference is the anonymisation Google applies to rare queries. So these seven represent 4.8% of the impressions we can see the wording for, and an unknown share of everything else. That is the honest frame: a visible slice, not a market estimate.
What correcting an AI answer actually means
Here is the sentence worth extracting, because it is the one most explanations get backwards.
Correcting what an AI assistant says about your company means replacing the sources it can retrieve, not editing the assistant. There is no field to update and no record to amend: the answer is regenerated from whatever the model can reach at the moment of the question, so the work is to make the current facts easier to reach than the old ones.
That is a claim about a mechanism, so here is its boundary. Feedback routes do exist — you can rate an answer, and OpenAI publishes a content report form — but they are built to remove or block output, not to correct a company description. When the privacy group noyb filed a GDPR accuracy complaint in April 2024 over an incorrect date of birth, OpenAI's position was that it could filter or block data on certain prompts but not rectify the fact itself. That case concerns personal data rather than a company profile, and it is from 2024; we cite it because it is the clearest public statement of the limit, not because it settles the corporate case.
Everything practical follows from that. You cannot file a correction. You cannot log in anywhere and change your entry. What you can do is change the population of documents an assistant draws on — your own pages, the profiles and directories that outrank them, the press mentions, the third-party pages that describe you in the past tense — and then wait for the retrieval layer to catch up.
This is also why the job is unevenly hard. If the wrong fact lives on one page you control, it is an afternoon. If it lives in six directory listings, an old press release, a marketplace profile and a competitor's comparison page, it is a quarter.
Who wrote these prompts, and why it matters
We are not going to claim seven business owners typed those paragraphs into Google. Look at how three of them end: "search online for providers and tell me who's worth a look", "name specific providers", "list 3 providers with websites". Those are instructions to an assistant, not to a search box. They reached Search Console because something forwarded them to a search surface on the user's behalf.
We published the underlying pattern on 2 August 2026: at that point, 35 of the 148 queries Google showed us had not been written by a person. The tells that give machine-written queries away are the same ones visible here — sentence length, explicit output instructions, an audience of one.
It changes what the data is, and it changes it in a useful direction. A machine-written prompt is not noise; it is a person's problem, restated by the tool they asked for help. The wording is arguably cleaner evidence than a keyword would be, because a keyword is what someone thinks the search engine wants, and this is what they actually said when they thought they were talking to something that understood them.
What it does not let us claim: volume. Seven prompts over 28 days is not a market. It is a shape.
Why the description stayed old
Four mechanisms, in roughly the order they cause trouble.
- The old page is still the better-connected one. A rebrand rarely takes the link graph with it. The page describing what you used to do has five years of citations behind it; the new one has a sitemap entry.
- Third parties describe you in the past tense and never revisit. Directory profiles, event bios, a partner's customer list, an interview from 2021. None of them has a reason to update, and several of them rank above your own about page.
- Your current page says it, but not plainly. A hero line that reads well and commits to nothing is invisible to extraction. "We help brands move faster" does not tell a retrieval system what you sell or to whom.
- The current page is not retrievable at all. Blocked at the firewall, absent from the index, or published so recently that nothing has fetched it. This one is the least discussed and the most decisive, which is why it gets its own section below.
Number four has its own literature at this point. We published the measurement of our own case on 6 August 2026: we found that 79% of AI crawler requests to our site were being refused by protections we had not knowingly switched on, while Googlebot got through 99% of the time. A company in that state can rewrite its about page every week without changing a single AI answer.
One of our last seven pages is in the index
On 27 August 2026 we ran the URL Inspection API against the seven English articles we published between 20 and 26 August. The set was defined by a rule rather than by choice: every English article with a publication date in that range, no exclusions.
| Google's verdict | Pages | What it means |
|---|---|---|
| Submitted and indexed | 1 | Retrievable, can be quoted, can rank |
| Discovered — currently not indexed | 4 | Google knows the URL exists and has not fetched or has not kept it |
| URL is unknown to Google | 2 | Not in the system at all yet |
The chronology matters more than the number, so here it is first. cittago.com relaunched on 6 July 2026. Daily publishing started on 27 July 2026. That is seven weeks of domain history and one month of cadence — a stage at which "Discovered — currently not indexed" is the ordinary state of a young site publishing faster than it has earned attention, not a verdict on the writing. Sitemaps were submitted, IndexNow returned 200 on both endpoints, and the pages resolve. Delivery is not the problem.
One caveat decides how much that number is worth, and it is the kind that usually goes unsaid. Google's index is not ChatGPT's index. OpenAI runs its own crawlers, other assistants run theirs or lease someone else's, and none of them publishes a coverage report you can audit. What Google's verdict gives us is the only checkable read on retrievability we have for our own pages, and it is a fair proxy for one narrow thing: whether anything has bothered to fetch and keep a page this new. The direct measurement on the assistant side is a different one, and it is the crawler audit referenced earlier.
We are publishing this rather than the flattering version of it because the flattering version would contradict the article. The whole argument here is that an unretrievable page cannot correct anything, and we are demonstrably in that position on six pages out of seven. If we only reported this once it was clean, it would be a press release with a chart on it.
It also produces the most useful sentence in the article for anyone about to spend money on this problem: a correction page that is not indexed is not a correction. Before anyone rewrites a single line of copy, the question is whether the page carrying the corrected facts is fetchable, indexed and linked from somewhere that already gets crawled. That unglamorous sequence is the front half of the work we do to keep AI assistants describing a company accurately, and it is the half people skip.
What you need before any of this works
- A written list of the facts that are wrong. Not "the description is outdated" — the specific claims: three discontinued products, a category we left, a city we no longer serve. You cannot displace a fact you have not named.
- The pages that carry each wrong fact. Search the exact phrasing of the false claim and see what comes back. Some of it will be yours.
- One page that states the current facts plainly. Entity, category, market, and what changed and when. Written to be quoted out of context, because that is how it will be used.
- Proof that page is retrievable. Indexed in Search Console, not blocked at the firewall or in robots rules, and linked from at least one page that is already crawled regularly.
- Access to the third-party profiles you can edit. Business profile, LinkedIn, industry directories, marketplace listings. This is the boring half and it is often where the wrong fact actually lives.
- A recorded baseline. The assistant's current answer, saved with its date, before you change anything. Without it you will not be able to tell whether anything moved.
What this does not do
- It does not edit the model. Nothing you publish reaches into a system's parameters. You are changing what it finds when it looks, and only for the systems that look.
- It does not work on a deadline. Retrieval layers refresh on their own schedule, and different assistants refresh at different rates. Anyone promising a date is guessing.
- It does not guarantee the assistant will mention you at all. Being described accurately and being recommended are separate outcomes with different causes. We wrote the second one up separately, in why ChatGPT will not recommend your company.
- It does not fix a training-data memory you cannot reach. Where an answer is generated without retrieval, no amount of publishing touches it in the short term. What you can influence is the retrieved layer, which is the one that governs answers about specific named companies most of the time.
- It does not survive you changing your mind. Consistency across your own properties is most of the work. Three slightly different descriptions of the company produce a fourth, invented one.
Before and now, in a table
| The situation | When search returned links | Now that answers are generated |
|---|---|---|
| An old page outranks your new one | A visible nuisance; users saw both and picked | Often invisible; one description is synthesised and shown |
| Fixing a wrong fact | Update the page, request reindexing, done | Update the page, then displace every other source that repeats it |
| Knowing you have a problem | Search your brand and look | Ask several assistants and record what each says, on a date |
| How long it takes | Days, and observable | Weeks to months, per assistant, and only observable by asking again |
The words, in a table
| Term | What it means here |
|---|---|
| Retrieval | The step where an assistant fetches documents to answer a specific question, instead of relying only on what it learned during training. It is the only layer publishing can influence quickly. |
| Discovered — currently not indexed | A Search Console state meaning Google knows your URL exists but has not added it to the index. The page is live and invisible at the same time. |
| Anonymised query | An impression Google counts but whose wording it withholds, usually because the query is rare. It is why query totals are smaller than page totals. |
| GEO | Generative engine optimisation: the work of being findable and quotable inside AI answers. The surfaces are roughly eighteen months old, so anyone claiming years of it is describing something else. |
| Entity | The thing a system thinks your company is — name, category, location, relationships. Wrong facts are usually wrong entity attributes rather than wrong prose. |
| Prompt-shaped query | A search-console query written as a full sentence or paragraph, often with instructions attached, indicating it was composed for an assistant rather than a search box. |
Who this is for, and who should not bother
It applies to you if your company already has a footprint that is now wrong. A rebrand, a repositioning, discontinued products, a change of market, a founder who left, an acquisition. There is existing material about you, it is being retrieved, and it describes a company that no longer exists in that form. That is the case the first two prompts describe, and it is the case where the work has something to push against.
It does not apply to you if almost nothing external has ever been written about your company. If there are no reviews, no directory entries, no press, no partner pages, then there is nothing to correct — the assistant is not wrong about you, it has nothing. Spending on correction work in that state is spending on the wrong job. The right one is getting mentioned at all, which is slower, mostly happens off your own site, and is a genuinely different engagement. We would rather say that before an invoice than after one.
There is a middle case worth naming: a company with a footprint that is accurate but thin. Here neither job is urgent, and the useful move is to make the existing facts unambiguous — one clear description, repeated identically everywhere you control — before adding anything new.
Where you stand, in three thresholds
First threshold — you have not asked. You believe the description is wrong because a customer mentioned it. Then the first hour is free and it is not ours: ask three assistants the same question about your company, save the answers with the date, and write down which specific claims are false. Most companies discover the problem is narrower than it felt.
Second threshold — you have the list of wrong claims and you own most of the pages carrying them. Then this is a content and indexing job you can largely run yourself: one plainly written page, verified as indexed, plus corrections on the third-party profiles you control. Expect weeks, not days, and re-ask on a schedule instead of watching.
Third threshold — the wrong facts live mostly on pages you do not own, and your own site is not being indexed reliably. Then the order of operations is the whole engagement, and it starts on the technical side rather than the editorial one. This is where accurate AI descriptions and ordinary search visibility stop being two projects. For a broader read on how the two disciplines now overlap, we set that out in SEO versus GEO in 2026.
Questions nobody has actually asked us
These arrived as search queries and as the questions we had to answer in order to write the piece. None of them came by email.
How do I correct what ChatGPT says about my company?
By replacing what it can retrieve, not by editing the assistant. There is no correction form. Publish one plainly written page with the current facts, make sure it is actually indexed, then work through the third-party sources repeating the old version — directories, profiles, partner pages, press. The assistant regenerates its answer from whatever it can reach.
How long does it take before the answer changes?
Weeks to months, and it varies by assistant because each refreshes its retrieval layer on its own schedule. Anyone giving you a date is guessing. What you can do is record the current answer with its date before you start, and re-ask on a fixed schedule rather than checking constantly.
Why does the AI describe us with products we discontinued?
Because the pages describing those products are still the best-connected pages about you. Discontinuing a product internally does not remove its page, its reviews, its directory entries or the articles that mentioned it. Until the current catalogue is easier to retrieve than the old one, the old one wins.
Can I pay to have my company described correctly?
Not directly, and be suspicious of anyone offering it. There is no paid channel for editing an assistant's description of a company. What money buys is the work of producing and distributing retrievable, consistent, current facts — which is ordinary publishing and outreach, priced honestly.
Is being described wrongly worse than not being mentioned?
In our own reporting the wrong-description complaint carries more demand: 50 of 71 impressions across the 28 days to 26 August 2026, against 16 for not being mentioned. They are different problems with different fixes. Being described wrongly is a displacement job; being absent is a mentions job, and the second is slower.
Does publishing a correction page work on its own?
Only if the page is retrievable. On 27 August 2026 one of the seven English articles we had published in the previous week was in Google's index; four were "Discovered — currently not indexed" and two were unknown to Google. A page in that state cannot correct anything, however well written it is.
Do these prompts come from real business owners?
Probably not verbatim. Three of the seven end with instructions to an assistant — "name specific providers", "list 3 providers with websites" — so they were composed for a tool that then searched on someone's behalf. The problem behind them is a person's; the wording is a machine's restatement of it.
Should I use structured data to state the correct facts?
Yes, as reinforcement rather than as the mechanism. Structured data makes attributes unambiguous — name, category, location, sameAs links — which helps a system that is already reading the page. It does not get the page read. Retrievability first, markup second.
What if the wrong fact is on a site that will not change it?
Then you displace rather than delete. Ask once, politely and in writing, because a surprising number of directories will edit. Where they will not, the remaining lever is making the correct version more numerous and better connected than the incorrect one, which is a slower and less satisfying answer than the situation deserves.
How do I know whether it worked?
Re-ask the same questions, in the same words, on a schedule, and log the answers with dates. Changing the wording of your test question changes the result, so the questions have to be fixed. Treat the first reading as a baseline rather than as a score.
Last updated: 27 August 2026. Query figures come from the Google Search Console API for the cittago.com property, dimension = query, window 30 July to 26 August 2026, row limit 5,000; the seven prompts are quoted as recorded. The 1,492 impressions attributed to named queries and the 11,338 total from the page dimension are from the same export on the same day. The indexing figures come from the URL Inspection API run on 27 August 2026 against all seven English articles published between 20 and 26 August 2026. The 79% crawler-blocking figure was published by us on 6 August 2026 and the 35-of-148 machine-written-query figure on 2 August 2026; both are linked above. The noyb case is cited from noyb's own account of it, dated 29 April 2024, and concerns personal data rather than a company description. We will update this page when we next re-run the indexing check, and if the split between the two complaint types changes materially.
Sources: Google Search Console API — Search Analytics query reference · Google Search Console API — URL Inspection · Google Search Central — Page Indexing report states · noyb — ChatGPT provides false information about people, and OpenAI can't correct it (29 April 2024)
