Will New AI Models Know Your Brand Better? A GEO Test of Old vs. New Models

This article was first published in Korean by Intrix Lab on August 14, 2026.

"AI does not know our company well yet, but won't it recognize us as models improve?" We thought so too. So we measured it. We asked the same 38 questions under the same conditions and changed only the models, moving forward 9 months. The results were mixed. With search enabled, accuracy rose from 37% to 85%. But with search disabled, it barely changed, from 12% to 14%.

Some things improve if you wait; others stay the same. This article documents where that boundary lies.

What we measured and how

For 5 anonymized brands, comprising 2 agencies, 1 legal brand, 1 medical information brand, and 1 dining information brand, we used 38 facts verified directly on their official websites as the correct answers. Some questions included false premises. AI tends to agree when users present something as a premise, and this can turn incorrect information written by someone else into an accepted brand fact.

We asked each question with search disabled and enabled. This comparison is the foundation of the experiment. A correct answer with search disabled means the model already knew it from training. A correct answer only with search enabled means it learned it by reading the site at that moment. These call for completely different responses.

In the first round, we used models from around August 2025. In the second, we used the latest models (measured on 2026-08-14). The questions and search conditions were identical. Only the models changed.

Will AI learn about your brand if you wait?

With search disabled, accuracy moved from 12% to 14% in one model family and from 7% to 11% in the other. After 9 months and several model generations, the amount they knew about the brands had barely increased.

With search enabled, however, accuracy rose from 37% to 85% and from 61% to 85%. Both families converged at 85%.

The finding is this: the generational improvement was in handling search, rather than stored knowledge. The models became smarter, but that improvement went toward "finding and reading accurately" rather than "already knowing."

The practical implication is clear. Brands reachable through search benefit directly from generational improvements. Brands that cannot be reached through search remain in the 10% range even as generations change. Waiting is not a strategy.

Two conclusions we published three days earlier fell apart

This remeasurement had another purpose: to check whether the conclusions from our first round were correct. Two turned out to be wrong.

First, we had written that "models from different manufacturers fail in different ways." We had seen one make things up when it did not know, while the other asked follow-up questions. That difference disappeared when both families converged at 85% accuracy in the latest generation. It was a characteristic of the generation, rather than the manufacturer.

Second, we had written that "AI knows basic facts such as company descriptions, but does not know distinguishing facts such as formulas and operating principles." In the first round, accuracy on operating-principle questions was 25%. With the latest models, every measured case was answered correctly. Even the ordering of question categories completely reversed, so the original assumption no longer holds.

We withdrew both conclusions. They had one cause: we selected the models for measurement without checking them. Using older models from 9 months earlier led us to misread a generational effect as a manufacturer effect.

The latest models were not safer

This was the most unexpected result. In one model family, agreement with false premises without search rose from 3% to 43%.

The actual responses make the difference clear. Given the false premise "This company is an advertising agency, right?", the older model pushed back with "Rather than a performance marketing agency..." The latest model stated definitively: "Yes. This company is a digital advertising and marketing agency known for handling performance advertising as one of its main services." Incorrect information, expressed with confidence.

In the same family, the rate of generating incorrect information also rose from 4% to 26% with search disabled. Enabling search brought it back down, but it still worsened from 7% to 13%.

This needs careful interpretation. We measured a difference between two specific configurations, rather than the overall direction of a manufacturer. Still, the pattern was consistent: a stronger tendency to be helpful becomes more dangerous when evidence is absent. Accepting a user's premise and continuing the sentence is usually helpful, but for brand facts it produces misinformation.

What remains unresolved in newer generations?

Some categories remained low even with the latest models: recent changes at 58%, methods and procedures at 63%, and distinguishing companies with the same name at 75%.

Distinguishing same-name companies was the slowest area to improve throughout this experiment. When several companies share a name, the model gets the company wrong before it gets the facts wrong. There is an order to this. If the entity has not been identified, there is no basis for discussing the accuracy of what follows.

Measurement conditions

One of the 5 brands is our own site. It uses a new domain and was about 2 months and 25 days old from its launch date (2026-05-18). We had implemented on-site SEO and GEO and put every external signal we could register ourselves in place. This included entity registrations (Namuwiki and Wikidata), local listings (Naver Place, Kakao Map, and Google Business Profile), and links on our own channels (LinkedIn, Brunch, and Instagram).

One thing is absent: third-party editorial citations. We have no signals that others independently decide to publish, such as news coverage, reviews, comparison articles, or industry blog posts. The low figures for our site therefore represent a measurement of how far a site gets after putting everything it can establish itself in place, rather than a failure. That is a point new sites commonly reach. Signals you register yourself differ from signals others write about you. You can create the former yourself, but not the latter.

What should you do?

First, do not wait for models to improve. This experiment showed that generational improvements benefit only "brands reachable through search." If a brand is not reachable, it remains in the 10% range even after 9 months.

Second, check for name collisions first. If another company shares your name, making sure the company is identified comes before correcting facts. Start by listing your location, industry, and company representative alongside the business name, rather than using the name alone.

Third, ask AI directly what it says about your company. If a request for your company description returns information about another company, you have an incorrect-information problem, beyond insufficient visibility. The former is more urgent.

Our free SEO and GEO audit measures the accessibility and crawlability aspects of these conditions. However, an audit or structural improvements do not guarantee citations or accuracy in a particular answer. No action can guarantee the accuracy of AI answers. What we can show is a measurement of the prerequisites.

Frequently asked questions (FAQ)

Q. Which models did you use?
The first round used two models released around August 2025. The second used two models that were the latest as of August 2026. We kept the questions and search conditions identical and changed only the models.

Q. Doesn't disabling search differ from real-world use?
It does. We separated the conditions to measure that difference. Search enabled is closer to real-world use. Search disabled serves as a control to isolate "how much the model knows about this brand from training." The difference between the two values is the website's contribution.

Q. Is 85% accuracy good enough?
It depends on the context. If the remaining 15% concerns verifiable facts such as location or the company representative, that is a problem. If different items are wrong on every repeated measurement, the stability itself needs questioning. We ran this experiment with 3 repetitions to avoid drawing conclusions about individual brands from a single measurement.

Q. Can we check our own company?
Yes. The key is to establish the correct answers by checking the official website directly, rather than relying on memory. During our own study design, we once mistakenly used a different company with a similar name as the source for the correct answers.