AI Describes the Company You Were Three Years Ago

A client making LED driver power supplies ran into something odd: a buyer asked DeepSeek whether their power supplies had surge protection, and AI answered without hesitation — mainstream models, no surge-protection version found. Yet the surge-protection range had launched six months earlier, written on the website, printed in the brochure, shown at trade fairs.

What AI said was not a lie. It was stale.

One distinction matters here: AI not knowing your new material and AI misremembering your old facts are two completely different problems with different solutions. The first is about opening up retrieval; the second can only wait for training. Handle them as one and both money and time are wasted. This is the point most easily glossed over in GEO, and gloss it once and the whole plan points the wrong way.

The boundary first. This article is about the timeliness of information and does not cover brands being misreported or mistaken for someone else — that is a separate matter. Nor does it discuss whether to let AI crawl your site; that premise has been covered on its own, and this article assumes crawling is open. Finally, every action below only raises the probability that new information gets read; none can promise how soon it takes effect. Generative engine optimization has no official timetable, and anyone who hands you a definite number of weeks is guessing on your behalf.

First, sort "old" into three kinds

When a buyer complains that AI got you wrong, do not rush to change content. Establish which kind it is first. In AI search optimization this feedback recurs, and sorting the type saves a great deal of wasted work.

The first, expired facts. AI still states your address from three years ago, a phone number you have changed, a model you discontinued. The most conspicuous kind, and the easiest to misdiagnose as "the website was not updated".

The second, missing additions. A new product, a new certification, a new line — AI knows none of it. It is not wrong; it has simply never heard.

The third, a stale judgement. AI still files you under "domestic sales only" or "entry-level models only", and that is no longer true. The hardest to notice, because you never think to ask.

The three are handled in a different order: the second shows results fastest, the first is slowest, and the third depends on accumulating content over time. Get the order wrong and you get "we changed a pile of things and AI is unchanged". That is the most common cause of rework when GEO moves from plan to execution.

Two paths: one frozen, one running

This is the technical floor of the whole thing, and where AI search optimization differs most from traditional SEO in principle. It is worth two minutes.

The first path is training. Once a model finishes training on a batch of material, its parameters are fixed. Its picture of the world stops at that moment and does not update itself afterwards.

The second path is retrieval. Before answering, the system pulls a set of pages from the search index, puts them into context as evidence, then composes the answer. This path is live — a page you publish today could in principle be picked up today.

Google's documentation contains a definition that makes the two paths plain. The robots.txt token Google-Extended governs two things: whether content Google has crawled may be used to train future versions of the Gemini models, and whether it may be used for grounding — supplying search index content to the model inside Gemini apps. The same page states plainly that Google-Extended does not affect a site's inclusion in Google Search results and is not used as a ranking signal in Search.

The value of that definition: being searchable and being trained on are two separate things at Google. You can appear in search results and stay out of the training set. Conversely, content already in the training set does not disappear because you changed the rules later.

So the first question is not "how do I make AI update". It is "which path does my problem belong to".

The training path: slow, irreversible, no shortcut

If the problem is that AI's overall picture of you is dated, it is most likely in the training layer. No technical means accelerates that layer; three things can be done. This also explains why many GEO projects spend the most time here and see no immediate feedback.

One, keep producing. Training material is crawled, and crawl volume correlates with content volume. A site publishing two pieces a year and one publishing forty are not equally likely to be included. There is a further limit easily missed: arXiv:2410.22349 measured citation completeness of AI answers against source pages at only 20% to 25%. Even when a page is picked up, only a fraction is taken. The passage you want taken has to be written tighter and closer to the top.

Two, earn third-party citation. Training material does not come only from your own site. AirOps' 2026 study puts 85% of brand mentions on third-party pages, and the same logic holds in the training layer. The University of Toronto preprint is finer: AI's preference for third-party sources when composing answers runs between 53% and 95%, depending on query type. So "who else besides you is talking about you" counts on both paths.

Three, accept the lag. How often a model is updated is not yours to decide and is not publicly committed to. Consider the scale for a moment: Google's AI Overviews now cover more than 100 languages, and a corpus pool at that size cannot turn over quickly. Some companies ask how soon they will see a change after an update. There is no answer to that question at present, and anyone who supplies one is guessing.

One thing has to be said plainly: content already in the training data does not disappear because you changed the rules. Google states this. So "block the crawler now, switch it back on once I have fixed the content" has no effect on what is already in, and only affects the future.

The retrieval path: fast, with three conditions

If the problem is that AI knows nothing about something new, it is in the retrieval layer. That layer is actionable, and it is the fastest return on effort in AI search optimization.

Condition one: the page has to be crawlable. Elementary to state, and the failure rate is not low. One wrong line in robots, a server returning different results to different user agents, a site returning an empty page to requests without a particular header — any of these makes crawling fail outright. You cannot see these problems in a browser, because your browser sends a pile of headers automatically.

Condition two: the date has to be readable. The most overlooked condition. "What Gets Cited", accepted at SIGIR 2026, ran 252,000 controlled experiments with one clear finding: a newer timestamp has a stable positive effect on being cited, as does explicit price information. AirOps' figures the same year put 83% of commercial citations on pages updated within the previous year.

The two are not the same kind of evidence — one controlled experiment, one vendor statistic — but they point the same way: time is a signal that gets read, not page decoration.

Condition three: the structure has to be cuttable. From an eight-hundred-word block of company blurb, a system struggles to extract the fact that a particular model has surge protection. Written as subheads and lists, the cost of extraction drops sharply. The Yonsei preprint measured a 22% lift in retrieval hits from structured treatment — a relative change under laboratory conditions, not an online result, but pointing the same way as the two conditions above.

How to date a page without faking it

The first reaction many people have on hearing that timestamps matter is to change the date at the bottom of the page to today. That is wrong, and it backfires. Every time-related action in GEO rests on one premise: it must be true.

The right approach is to let the date reflect a real event.

• Use datePublished for first publication and dateModified for substantive revision, both with true values.
• Write the sentence into the body: "surge-protection range added March 2026". That sentence is itself an extractable fact, worth more than any tag.
• Let lastmod in the sitemap follow real edits; do not batch-stamp everything to one day. Hundreds of URLs sharing a single modification time tells the system the field cannot be trusted.
• On product pages, state when a model launched and when it was discontinued. Fewer people write discontinuations than launches, but when a buyer asks whether a model is still available, the answer exists only if someone wrote it.

Another common misconception is piling on structured markup to chase rich results. Google has already said AI features need no special schema or markup, and FAQ rich results stopped showing on 7 May 2026. Do the markup that is due, but do not expect exposure in return. The claim that "add a tag and you get recommended" circulates widely in AI search optimization and holds up badly to checking.

Which layer is your problem in?

There is a simple way to tell: ask two kinds of question about the same subject.

First, a specific fact: "does company X have a surge-protection power supply range in 2026". If AI answers correctly, the retrieval layer is open and your page can be picked up.

Second, an open evaluation: "what is company X like, what do they mainly do". If that still returns the picture from three years ago while the first question is answered correctly, the problem is in the training layer.

If neither is answered correctly, check crawling before writing anything. With crawling broken, writing is wasted at any volume. This is the step most often skipped in generative engine optimization and the most expensive to skip — it produces nothing to show anyone, but if it is broken every later piece of work returns zero.

One more case worth testing separately for B2B companies: ask about your model numbers directly. A model number is the hardest fact there is; if AI cannot answer even that, the product page has no position in the retrieval layer, and brand awareness has nothing to do with it. Export factories in the Pearl River Delta carry many models and iterate fast, which makes this especially visible: a product is revised every other year while the model number is often carried over, and AI cannot tell the new one from the old.

Do not test on a single engine. Ask Doubao, ERNIE Bot and ChatGPT once each; the answers may be entirely different — the overlap between the pages different systems cite is low to begin with, and one correct answer does not mean all are correct.

Three common questions

We have updated the site. How soon will AI read it?

The retrieval layer usually runs in days to weeks, depending on crawl frequency and site authority; the training layer runs in months or longer, and is not disclosed. Neither figure can be promised, because neither is yours to control. This is the question most easily blurred in GEO conversations, and writing it into a contract is worse still.

Should we run a News section just to look active?

With real content, yes. If it means republishing old pieces under a new date, no — such pages are not cited, and they drag down how the site's overall content quality is judged. GEO wants an increment of citable facts, not an increment of dates.

A buyer says AI got us wrong. Can we ask it to correct that?

There is no direct channel. What you can do is write the correct information in a form more easily retrieved, and get third parties to say the same thing. The role of generative engine optimization at this layer is to raise the odds that the correct version is picked up, diluting the share held by the old one.

Where this leaves you

Back to the client at the start. The surge-protection range had been live six months, written on the site, printed in the brochure, shown at a trade fair, and AI did not know. The reason is not complicated: in those six months, no third-party page other than their own mentioned the range.

At this layer GEO is not a contest of technique. It is who else besides you is talking about you. It sounds like common sense, and it explains most of the puzzlement behind "I wrote it, why will AI not say it".

Yihang Network is based in Shenzhen and started out in Alibaba.com-related services before turning to GEO in 2024. Testing whether a site can be crawled before touching content is a habit earned the hard way: one engagement saw three months of content changes produce no response at all, and the cause turned out to be a server returning an empty page to any request without a particular header. Three months, wasted.

Data and effect boundaries: the statements cited from Google's documentation (Google-Extended governing training and grounding only, with no effect on search inclusion or ranking; AI features requiring no special schema or markup; FAQ rich results withdrawn on 7 May 2026) are platform rules whose effect is limited to the Google ecosystem, and the Google-Extended description covers only Gemini and Vertex AI rather than other vendors' crawler policies. The 252,000 controlled experiments in "What Gets Cited" (accepted at SIGIR 2026) are findings under specific experimental conditions, not a guarantee of online mechanism; AirOps' 83% and 85% are vendor-published observational statistics, not independently verified; the University of Toronto range of 53–95%, the 20–25% in arXiv:2410.22349 and the Yonsei 22% are preprint findings, not peer-reviewed, and represent relative changes under laboratory conditions; the number of languages supported by Google AI Overviews is an official help-centre figure that changes between versions. None of the above represents any specific company's indexing or citation outcome. The methods described raise the probability that new information is retrieved and cited; they guarantee no appearance in any AI answer at any point, and promise no timeline and no business outcome.

Sources

1. Google for Developers: "Google's common crawlers" (Google-Extended has no separate HTTP user agent and controls training and grounding only; explicitly no effect on search inclusion or ranking), accessed 2026
2. Google Search Central: "AI features and your website" (AI features require no special schema or markup)
3. Google Search Central: rich results documentation update (FAQ rich results withdrawn from 7 May 2026)
4. Google Help Centre: languages supported by AI Overviews (more than 100, subject to change between versions)
5. "What Gets Cited", accepted at SIGIR 2026: 252,000 controlled experiments (newer timestamps and explicit price information have a stable positive effect on citation; peer-reviewed acceptance)
6. AirOps: "State of AI Search" 2026 (83% of commercial citations from pages updated within the previous year; 85% of brand mentions from third-party pages; vendor-published)
7. arXiv:2410.22349 (citation completeness of AI answers against source pages, 20–25%; preprint, not peer-reviewed)
8. University of Toronto: arXiv:2509.08919 (preference for third-party sources 53–95%; preprint, not peer-reviewed)
9. Yonsei University: arXiv:2602.12187 (structured treatment lifts retrieval hits by 22%, laboratory conditions; preprint, not peer-reviewed)