Why AI Does Not Mention Your Company at All

There is a difference between a wrong answer and no answer at all. Anyone who asks ChatGPT about the providers in their sector and does not find their own company in the list has no correction problem. They have a threshold problem.

That is a different starting point, and it calls for a different route. What helps against wrong information does not help against not being mentioned.

In brief: Language models reproduce knowledge from training reliably only when a topic appeared often enough in the training material. Below that threshold the result is not a wrong answer but no answer. A study from May 2026 covering 38 models shows that model size and topic frequency together explain around 60 percent of recall quality, and that the threshold falls only slowly as models grow: reaching reliable recall on a rare topic would, by the authors' extrapolation, take a model of roughly 50 trillion parameters, about thirty times the largest one tested. For mid-sized companies the conclusion is that the route into training knowledge is effectively closed. The route that remains runs through live retrieval, and it works in weeks rather than years.

Why does AI not mention my company?

Because for the model it does not exist. That sounds blunt, but it describes the mechanism more precisely than any assumption about bias or filtering.

During training a language model does not learn a database but statistical relationships. Whether it can retrieve a name later depends on how often that name appeared in the training material and in how many different contexts. A company with a handful of mentions on its own website and in two directories sits below what consolidates into retrievable knowledge.

The distinction from the two other cases, which are often confused, matters:

We have covered the first two elsewhere, once on false AI claims and once on why AI describes companies vaguely. This piece covers the third.

What does the research say about this threshold?

That it is measurable and follows a regular curve.

A study by Matthew L. Smith, Jonathan P. Shock, Samuel T. Segun, Iyiola E. Olatunji and Tegawendé F. Bissyandé dated 18 May 2026 (arXiv:2605.18732) examined 38 models from one billion to 405 billion parameters for how reliably they reproduce academic literature. The basis was 8,913 scholarly references across 24 subject areas.

The result is a sigmoid curve driven by two factors: the size of the model and the frequency of the topic in the training material. Together they explain around 60 percent of the variation in recall quality, and within individual model families 74 to 94 percent. Model size alone contributes 42.1 percent, with topic frequency adding a further 17.8 percent.

The second figure is the interesting one. Topic frequency is not a side condition but an independent factor of the same order as a substantial part of model size.

The authors describe three regions a topic can occupy:

A typical mid-sized company sits in the floor region. That explains two observations which belong together: the absence of a mention and the occasional invented detail. Both are the same phenomenon, once as omission and once as improvisation.

Does waiting for larger models help?

Barely. This is the practically most important finding of the study.

The authors measured how rare an academic paper may be for a model still to know it. Llama 3.1 with 8 billion parameters recalls papers with a median of 2,419 citations. The 405-billion model of the same family reaches 589 to 806 citations.

Working that through: a fiftyfold increase in model size lowers the recognition threshold to roughly a quarter. The relationship is logarithmic, not linear.

The authors extrapolated the curve for a rare topic. Reaching a recall quality of 0.90 there would require a model of roughly 50 trillion parameters, about thirty times the largest model in the test. Their own conclusion: for topics below the threshold, retrieval that bypasses parametric recall entirely is the appropriate response.

Translated into the position of a mid-sized company: the next model generation will not know your company. Nor will the one after that. This is not a pessimistic reading but the result of an extrapolation along the measured curve.

A second study supports the order of magnitude. The FACT-Bench work by Jiaqing Yuan and colleagues used 20,000 question-answer pairs across 20 domains to test how strongly the prominence of an entity affects recall. GPT-4 reached 65.9 percent exact matches on well-known entities and 52.3 percent on rare ones. The gap of 13.6 percentage points arises within a group of entities that all still appear in the model's knowledge. A company that does not appear at all sits below that scale entirely.

How do I tell whether my company is below the threshold?

Through a test that takes ten minutes and needs no software.

Ask three questions in a chat system that demonstrably does not search the web but answers from training. Most providers allow web search to be switched off.

First, the direct question about the company name. Second, the same question with the location added. Third, a question about providers of your service in your region, without naming your company.

The answers sort themselves:

A correct description means you are above the threshold. Your topic is then not this article but the accuracy of the description.

A description that clearly means a different company puts you in the floor region with name confusion. The name is too weakly tied to you.

A generic answer without specifics, or a note that nothing is known, means you are below the threshold. This is the most common case.

Repeat the test with web search enabled. The difference between the two runs is the genuinely interesting number, because it shows what live retrieval can do for you.

Why does the model sometimes say something anyway?

Because a language model is built to produce a continuation, not to report a gap in its knowledge.

In the floor region, according to the study, the model produces templated answers. It assembles what belongs in an answer of that kind: an industry, an order of magnitude, a city, sometimes names drawn from the neighbourhood of similar terms. The result reads fluently and is not true.

An uncomfortable insight follows for practice. An invented description of your company is not evidence that the model knows you and merely mixed something up. It is usually evidence of the opposite.

This is why the obvious route misleads. Anyone who responds to a wrong answer by correcting details is correcting something that was never stored. The work has to start somewhere else.

The route runs through live retrieval, not through training

And that is the good news in this matter.

Modern assistant systems answer along two routes. Either from training knowledge or through a search at runtime, in which pages are fetched and evaluated. For unknown entities the second route is the only one open, and it has three properties that suit mid-sized companies.

It is independent of mention frequency. What counts is whether a page is found, fetched and understood, not how often it appeared in a training run.

It works quickly. Changes take effect once the affected pages have been fetched again, so within weeks rather than with the next training state.

And it is open to influence. Access, structure and clarity of your own pages are things a company decides for itself.

What exactly should be done?

Four steps, in this order, because each builds on the previous one.

First, access. Check whether the AI providers' crawlers are permitted to fetch your pages at all. This is not a formality: training, search index and live retrieval run through different bots at every provider, and a robots.txt that clears only one of them shuts out the others. Blocking here leaves you as invisible on the retrieval route as in training knowledge.

Then, delivery. Check whether the substantive content sits in the delivered HTML rather than appearing only through JavaScript. Most AI crawlers execute no JavaScript. A page that gains its text only in the browser is empty to them.

Next, clarity. Company name, legal form, address, service and category must match across every source that exists about you. Contradictions do more damage here than gaps, because they actively obstruct attribution rather than merely leaving it incomplete.

Finally, quotability. A system searching at runtime needs something to quote. Concrete figures, procedures, responsibilities and answers to real questions are quotable. General self-descriptions are not. Which sources actually count is set out in our separate piece on the sources behind AI answers.

How long does it take, and how do I measure it?

Allow a quarter before drawing conclusions.

The technical steps work quickly: a changed crawler permission and server-side delivered content take effect with the next fetch, so within days to weeks. Clarity across third-party sources takes longer, because directories and portals update on their own schedule.

Measure with the same list of questions before and after, with and without web search, and record how often you are mentioned at all. That rate is the only metric that matters for the case described here, and it is a different one from the accuracy of the description. How to set such measurements up cleanly, so that chance does not determine the results, is covered in our piece on measuring AI visibility.

One expectation is worth dropping: that your company will one day be named out of training knowledge. On the extrapolation cited above that is not a realistic target for a mid-sized firm. Something else is achievable, and for practical purposes it is entirely sufficient: that a system searching at runtime finds you, understands you, and can quote you.