Why Are AI Answers About Your Company Out of Date?

Why are AI answers about your company out of date?

Because the assistant is describing whatever sources it can reach, and your old sources outnumber your new ones. A model does not know that your pricing changed in March. It knows what the pages it has read say, and most of those pages were written before March.

This is the complaint I hear most often now, and it usually arrives with some frustration. Somebody asks an assistant about their own company and gets back a product they discontinued, a positioning they abandoned, or a price they have not charged in two years.

It is fixable, but not by the route most people try first. Writing a new page and waiting is the slowest possible approach, and it is what almost everyone does.

Where does a model get its idea of your company?

From two different places with two different clocks. Some of it comes from what was absorbed during training, which is fixed until the next model. The rest comes from retrieval at the moment of the question, which reflects whatever is reachable now.

The distinction matters because only one of them responds to anything you do this week. You cannot edit what a model learned during training. You can absolutely change what it finds when it goes looking, and that is where all the available leverage sits.

You can often tell which one you are dealing with by whether the answer carries citations. An answer with links is showing you its retrieval, and you can go and read the sources it used. An answer with no links and confident detail is more likely reciting, and that is a harder problem with a slower fix.

Which of your own pages are feeding the stale answer?

Usually a page you forgot about. An old pricing page, a launch announcement, a comparison page from a previous positioning, a careers page describing a product line you closed. These rank, they read as authoritative because they are on your domain, and nobody has looked at them in years.

Find them by searching your own site for the outdated claim rather than for the topic. If an assistant says you charge per seat and you no longer do, search your domain for the per seat language. In my experience the source is on the company's own site about half the time, which is the good half because you can fix it today.

Be thorough about where content hides. Blog posts, old landing pages, PDF downloads, help articles, and the archived versions of pages that were never properly retired. A PDF from 2023 sitting on your own domain is a perfectly readable source and it has no publication date on the page telling anyone it is stale.

What about pages you do not control?

Directory listings, review sites, old press coverage, and partner pages are the other half. You cannot edit them, so the work is different: contact the ones that matter, and make sure your own corrected page is more findable and more specific than theirs.

Prioritise by what actually appears in the assistant's citations rather than by what annoys you. A wrong listing on a site nobody retrieves is a small problem. A wrong listing on a site the assistant keeps citing is the whole problem, and it deserves a real email to a real person asking for a correction.

Keep a record of who you asked and when. Third party corrections take weeks and are easy to lose track of, and the same directories go stale again on a predictable cycle. This is one of the things worth carrying forward from a proper measurement pass, which I described in how to build a baseline for AI visibility before you change anything.

Why does deleting the old page not fix it immediately?

Because removal from a search index and removal from a model's picture of you are different events. Google's own documentation says its Removals tool will remove a page hosted on your site from Google's search results within a day, which is fast, and also that requests made in the Removals tool last for about six months.

That six month expiry is the part people miss. A removal request is a temporary measure, not a fix. Google's documentation is explicit that to make a removal permanent you should remove or update the content on the page, password protect it, or add a noindex tag. If you only file the request, the page comes back.

My preference is almost always to update rather than delete. A deleted page loses whatever authority it had accumulated. A rewritten page at the same address keeps that authority and redirects it toward the correct claim, which is a better outcome for both search and retrieval.

How do you publish a correction that actually gets picked up?

Put the current fact on the page that most obviously answers the question, state it plainly in a full sentence, and make sure it is not contradicted elsewhere on your own site. Contradiction is what keeps a stale claim alive, because the model has no way to tell which of your pages is current.

Write the correction as a standalone statement rather than as an implication. A pricing page that shows the new model implies the old one is gone. A sentence saying that as of this year pricing works this way and no longer works the other way is directly quotable, and quotable is what gets retrieved.

Dating the claim helps more than most people expect. A sentence that carries its own time reference gives a retrieval system something to prefer over an undated competitor, and it gives a human reader a reason to trust it. This is the practical end of the freshness question I looked at in how fresh a page has to be for AI search.

How long should you expect this to take?

Longer than a site edit and shorter than you fear, and genuinely variable. Retrieval based answers can change within days of a page being recrawled. Answers that come from training do not change until a model does, and no amount of publishing accelerates that.

I would not promise a client a timeline here, and I would be suspicious of anyone who does. The honest framing is that you are improving the odds on every future answer rather than editing a record. The work is worth doing on that basis, and it is worth doing whether or not any particular answer flips this month.

What you can control is the measurement. Re-run the same questions monthly and record what came back, so that when the answer does change you know it, and so that you can tell the difference between a real improvement and a lucky session.

What should you do before the next thing changes?

Keep a list of every page that states a fact likely to change, and update all of them on the same day the fact changes. Pricing, positioning, product names, team size, integrations, and anything describing what you do not do. The list is usually shorter than people expect.

The discipline is treating a change of fact as a publishing event rather than an internal one. When a company changes its pricing, the pricing page gets updated the same week and the nine other pages mentioning the old model get updated eight months later, if ever. Those nine pages are what the assistant will find.

Write the list once and attach it to whatever process governs the change. It costs an hour to build and it prevents the situation where, two years from now, somebody asks an assistant about you and gets an answer you would not recognise.

What should you do next?

Ask an assistant three direct questions about your own company today, from a clean session, and write down anything that is wrong. Then search your own domain for the wrong claim rather than for the topic, because that is how you find the page that is causing it.

Fix what is on your site first, by updating rather than deleting, and only then start on the third party sources. If you file a removal request as a stopgap, remember that Google's documentation says it lasts about six months, so pair it with a permanent change rather than relying on it.

If an assistant is describing your company in a way you do not recognise and you cannot work out where it is getting it from, send me the question and the answer. I am happy to help you trace the source.