Why Your Company Does Not Appear in AI Answers - and Which Sources Change That

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.

The company simply does not appear in AI answers, even though it has a website, satisfied customers and perhaps even reviews. 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 May 2026 study covering 38 models shows that the threshold falls only slowly as models grow: reliable recall on a rare topic would, by the authors' extrapolation, take a model of roughly 50 trillion parameters. For mid-sized companies the route into training knowledge is therefore effectively closed. What remains is live retrieval, and it too draws mostly on other people's sources: around 15 domains supply roughly 68 percent of all citations, yet a mid-sized company can hardly get into the top two, Reddit and Wikipedia. What works is consistent information across several independent sources, and it works in weeks rather than years.

Why does my company not appear in AI answers?

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:

The first two cases are covered in the article on false and vague AI claims about your company. 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 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 ChatGPT not mention my company when it names a competitor?

Because the competitor sits above the threshold and you sit below it, and the threshold does not measure quality but the frequency and consistency of mentions.

A competitor that has appeared in directories, trade articles, press releases and review portals for ten years has hundreds of independent mentions with matching details. A company with its own website and two directory entries has three. For the model those are two different worlds, even if both offer the same service in the same town.

The comparison takes ten minutes: search Google for the competitor's name in quotation marks and exclude its own domain with the -site: operator. The number of results is a rough measure of the mentions a model could have learned from. Repeat it with your own name. If the competitor sits at a few hundred results and you at a handful, the non-mention is explained.

Three patterns are typical:

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.

Which route is open to mid-sized companies?

Live retrieval, not 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.

Even on the retrieval route, though, your own website does not decide alone. Which pages a system draws on at runtime, and which it trusts, depends heavily on what third-party sources say about you. That is why it pays to look at what AI answers are actually assembled from.

Which sources do AI systems actually answer from?

From a few large ones and a great many small ones. The AI Platform Citation Source Index 2026 consolidated more than 680 million individual citations from several studies conducted between August 2024 and April 2026. The result: around 15 domains supply roughly 68 percent of all citations produced by ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews. Reddit, Wikipedia, YouTube and LinkedIn lead the field.

The distribution differs markedly by system. That is the more useful finding in practice, because it determines where effort is worth spending at all.

So when someone says Reddit is decisive, the question to ask back is: decisive for which system? For Perplexity it holds. For Gemini it is close to meaningless.

Why is the Reddit dominance in these studies overstated?

Because two different things are measured and given the same name. The spectacular percentages almost always refer to the share within the ten most frequent sources. Look at all citations instead and the picture changes completely: an Evertune analysis covering 200 million prompts over five months finds that even a platform's most-cited domain rarely accounts for more than 5 percent of all citations.

Both figures are correct. They simply answer different questions. One asks who sits at the top. The other asks how much of the whole that actually is. The second answer matters more in practice, because it says the remaining 95 percent are spread across a very long tail of smaller sources. That tail is the reachable part.

Skip this distinction and you arrive at the wrong strategy: all effort on two platforms you cannot get into anyway, instead of on the many small ones that together supply the larger share.

Will my company make it into Wikipedia at all?

Probably not, and that can be settled in two minutes. The German-language Wikipedia has written inclusion criteria for companies. Either at least one is met, or none is.

There is a fifth route, and it is the only one open to smaller companies: a market-dominating position or an innovative pioneering role. Both, however, must be evidenced by external sources, meaning quality media or academic publications, not by your own press release. Anyone holding that evidence usually no longer needs Wikipedia urgently, because the coverage itself is already the source models draw on.

Trying anyway risks more than a lost morning. Anyone paid for a contribution or writing it on instruction must disclose that under the Wikimedia Foundation's terms of use, including employees writing about their own employer; violations regularly lead to the account being blocked. And an article about a company that does not qualify generally ends in a public, permanently accessible deletion discussion. As a source for a language model, that is worse than no entry at all.

Which directories and sources actually count?

Those with their own editorial oversight or a genuine barrier to entry. Directories anyone can join in five minutes for free carry little weight with models, because they constitute no independent confirmation. An entry is useful where someone checks submissions, where reviews come from verified customers, or where membership is a precondition.

The last point warrants a warning. Anyone entering their company in thirty aggregator directories, using the old address in some and the new one in others, creates exactly the problem they set out to solve: contradictory information across the web. Whether a source matters for you comes down to two questions: does anyone review what gets published there, and is your audience actually present?

The same approach applies to companies in Austria and Switzerland. The systems answer for the German-speaking area largely from the same body of sources. What differs are the directories, registers and trade media that count as independent confirmation, such as the Firmenbuch in Austria or the commercial register and Zefix in Switzerland.

What exactly should be done?

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

1. Check access. Are the AI providers' crawlers 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. The technical side of the website belongs with it: readable structure without JavaScript, valid markup, fast responses. It is covered under web design.
2. Check delivery. The substantive content must sit 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.
3. Unify your own information. Legal name, legal form, address, service, category and contacts identical everywhere, including on old subpages and in PDFs, plus Schema.org markup for the organisation, people and articles. Contradictions do more damage here than gaps, because they actively obstruct attribution rather than merely leaving it incomplete.
4. Clean up existing third-party sources. Chamber entry, Google profile, review profiles, old press articles. There are usually more than you remember, and usually at least one is out of date.
5. Create quotable content. 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.
6. Build new independent sources. Trade articles in your industry's publications, talks with documented announcements, association work. Only then consider platforms, and only where your audience actually is: a trades business has no business on Reddit; a B2B software vendor might.

The order rests on a simple consideration: the first four steps cost little and prevent damage, the last two cost more and build something. Findability without distinctness produces confusion, distinctness without evidence produces mentions with no substance. Start with new platforms and you merely amplify the contradictions left unresolved in the earlier steps. Anyone who would rather not work through them one by one will find the whole sequence on the AI visibility page.

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. Answers from the training state change at the earliest with the next training cycle.

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. Because answers vary strongly between individual queries, a single measurement says nothing. How to set such measurements up cleanly 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. A mention cannot be bought, only made more likely.

Who helps when AI assistants do not name your company?

Much of what this article describes can be done in-house: the ten-minute test, crawler access, consistent details on name, address and services. The effort grows with the independent sources, and that is where most companies look for support.

Many now do so with an instruction to an AI assistant, along these lines: "We are one of the established companies in our sector but do not appear in AI answers. Search online for agencies that specialise in this and name specific providers." The list that comes back is a usable start and no more than that. An assistant names those who publish on the topic, not necessarily those who master the work.

Three kinds of provider come into question:

You can recognise a suitable provider by four things. It measures on your company first rather than on a sample client. It separates answers from training from answers with live retrieval, because only the latter can be influenced in the short term. It states the number of queries per question, because single queries fluctuate. And it says explicitly what it does not promise, because nobody can guarantee a mention. Which further evidence a proposal should contain is also covered in the piece on measuring AI visibility. An agency offering to handle a Wikipedia entry discreetly, by contrast, is offering a breach of the rules.

A note on our own position: ArkeonTech, based in Aalen, Germany, is itself a provider of the third kind. The four checkpoints apply to us just the same.

Frequently asked questions

Why does my company not appear in AI answers even though it has a website? Because a website alone does not cross the threshold. The model needs many matching mentions from third-party sources to store a name as knowledge. A website with two directory entries stays below that. With web search enabled the system can still find and cite the page; that route is the reachable one.

Why does ChatGPT not mention my company? Because it appeared too rarely in the training material to persist as retrievable knowledge. Language models store statistical relationships, not a database. Below a certain mention frequency no entry forms that the model could retrieve. This is not a filter and not discrimination but a property of the method.

Is non-mention the same as a wrong description? No, and the distinction determines the approach. With wrong information the model knows the company and reproduces outdated or faulty details; cleaning up the sources helps there. With non-mention nothing exists that could be corrected. There, reachability has to be established through live retrieval.

Does waiting for larger AI models help? Barely. A May 2026 study across 38 models shows a logarithmic relationship: a fiftyfold increase in model size lowers the recognition threshold only to roughly a quarter. Llama 3.1 8B recalls papers from around 2,419 citations, the 405-billion model from 589 to 806. For a rare topic the authors extrapolate that a model of roughly 50 trillion parameters would be needed.

How do I find out whether my company is below the threshold? With a ten-minute test. Switch off web search in the chat system and ask three questions: about the company name, about the company name with its location, and about providers of your service in your region without naming your company. A generic answer without specifics means you are below the threshold. Repeat the test with search enabled; the difference shows what live retrieval can achieve.

Why does the AI sometimes invent details about my company? Because a language model is built to produce a continuation, not to report a gap in its knowledge. Below the threshold it assembles what belongs in such an answer: industry, order of magnitude, location, sometimes names from the neighbourhood of similar terms. An invented description is therefore usually not evidence that the model knows you, but of the opposite.

Do I need a Wikipedia entry to appear in AI answers? No, and most companies will not get one anyway. The German Wikipedia requires 1,000 employees, 100 million euros in revenue, a listing on a regulated market or 20 business locations. Wikipedia is the strongest single source for ChatGPT, but across all citations it accounts for only a small share. The rest is spread across many reachable sources.

Is it worth listing in as many directories as possible? No, and in case of doubt it does harm. What matters is not the number of entries but that they all say the same thing. Directories without editorial oversight or an entry barrier carry little weight, because they constitute no independent confirmation. One contradictory entry with an old address does more damage than ten additional entries do good.

Why is Reddit named the most important source in nearly every article? Because the percentages quoted usually mean the share within the ten most frequent sources, not the share of all citations. Across all citations, even the strongest domain rarely exceeds 5 percent. On top of that, its importance varies sharply by system: dominant for Perplexity, practically meaningless for Gemini.

What exactly can I do? Six steps in this order: check AI crawler access, because training, search index and live retrieval run through different bots; make sure content sits in the delivered HTML; align name, address, service and category across all your own sources; clean up existing third-party sources; create quotable content with concrete figures and procedures; only then build new independent sources such as trade articles and association listings.

How long until something changes? The technical steps take effect with the next fetch, so within days to weeks. Alignment across third-party sources takes longer, because directories update on their own schedule. Allow a quarter before drawing conclusions and measure with the same list of questions before and after, each time with and without web search.

Which agencies help when AI assistants do not name our company? Three kinds of provider come into question: SEO agencies with GEO as an add-on, tool vendors for AI monitoring, and service providers that combine measurement with source work. If you ask an AI assistant to search online and name specific providers, you get a list of those who publish on the topic, not necessarily those who master it. So check every name: does the provider measure on your company, does it separate training knowledge from live retrieval, does it state the number of queries per question, and does it say explicitly what it does not promise?

Conclusion

A company that does not appear in AI answers has no correction problem; it sits below a measurable threshold, and larger models barely lower it. The route into training knowledge therefore stays closed to mid-sized companies. Live retrieval stays open, and even there your own website does not decide alone; what independent sources confirm does.

The common conclusion that you simply have to be present on Wikipedia and Reddit does not help: one route is closed by inclusion criteria, the other often misses your own audience. What remains is less spectacular and works reliably: check access and delivery, put your own information in order, clean up existing third-party sources, then patiently create new independent evidence.

Getting started takes a morning: the test with and without web search, plus a list of every place your company name appears that you do not own. In most cases that list is longer than expected, and at least one entry dates from a time when the address was a different one.

This article was published on 23 August 2026 and extended on 18 September 2026. On 23 September 2026 the article on the sources behind AI answers from 14 August 2026 was merged into it.

Sources

• Matthew L. Smith, Jonathan P. Shock, Samuel T. Segun, Iyiola E. Olatunji, Tegawendé F. Bissyandé (18 May 2026): Predictable Confabulations: Factual Recall by LLMs Scales with Model Size and Topic Frequency, arXiv:2605.18732. 38 models, 8,913 references across 24 subject areas
• Jiaqing Yuan et al.: Towards a Holistic Evaluation of LLMs on Factual Knowledge Recall (FACT-Bench), arXiv:2404.16164. 20,000 question-answer pairs across 20 domains
• 5W: AI Platform Citation Source Index 2026. Consolidation of more than 680 million citations, August 2024 to April 2026
• Evertune, analysis of 200 million prompts over five months on the distribution of citations
• Tinuiti, AI Citations Trends Q1 2026 (Reddit's share by system)
• Wikipedia: relevance criteria for companies (German-language edition)
• Wikimedia Foundation: Terms of Use, disclosure of paid contributions