Being named in ChatGPT, Gemini and Perplexity: what a service provider can realistically do

Imagine a tax advisor with a small team in a mid-sized town. Her website has been up for years, it appears on the first page for her town name and core service, and enquiries come in reliably. Then she notices that a new client says in the intro call: “I asked ChatGPT first, and then there were three names.” Hers was not among them.

This is a typical starting picture, not a documented individual case. In conversation with an assistant, people phrase questions longer and more specifically than in a search engine. Instead of “tax advisor town”, the question reads something like: “What should I look out for when I’m a freelancer looking for a tax advisor, and who is an option in my region?” The answer then appears ready-made in the chat window or as an overview above the results list.

This raises a new question for the business owner: what can she do so that her business appears in such answers? This study answers it without promises of salvation. It separates what is known, what is plausibly assumed and what nobody can know from the outside.

The topic now has several names, such as Generative Engine Optimization, or GEO for short. Here we mostly use the sober term “AI visibility”, because it describes what this is about: whether and how a business shows up in the answers of generative systems.

The first difficulty is fundamental. ChatGPT, Gemini, Perplexity and Google’s AI Overviews work differently. Some systems answer from what they learned in training, others additionally retrieve current web pages and cite them. Which route is taken for a specific question is decided by the system, not by the website owner.

The second difficulty is the answer itself. The same question can produce different sources at different times. A single observation (“today I’m included”) is therefore hardly meaningful, and there is no ranking in the classic sense. Anyone who promises you a fixed placement in an AI answer is claiming something that cannot be verified.

Third, the incentive to exaggerate is great. A market that is only just emerging quickly produces recipes that sound too good: special files that “convince the AI”, or masses of text meant to feign source authority. For most of these recipes there is no reliable evidence.

Fourth, measurement remains fuzzy. The providers do not publish reports like Google Search Console in which you can see your mentions. You measure through samples and indirect traces and have to report the uncertainty alongside.

The rough mechanism of retrieving systems is known. In response to a question, the system searches for suitable web pages, breaks them into sections, rates those sections by relevance to the question and formulates the answer from them. The page from which a used section comes is then usually the one cited. It follows that whole pages do not compete, but individual passages.

It is also documented that the systems draw on existing search indexes. For Google’s offerings that is the Google index; for other providers it is partly their own crawlers, partly search partners. A page that is not indexed or not retrievable has hardly any chance of being named. This is the part that can be checked cleanly.

The weighting is unknown. How much do domain authority, freshness, mentions on third-party sites and the wording of the passage count? The providers do not disclose it, and it changes with model versions. Everything that circulates publicly about it is observation or conjecture. That also applies to our own assessment, and we label it accordingly.

A detail that is often mixed up: for AI Overviews and AI Mode, Google states that no additional markup is needed. Anyone who is indexable and snippet-eligible for normal search is in principle also eligible for these. Google does not prescribe any separate “AI optimisation” in the technical sense.

From what is known we derive a strategy that is deliberately modest. We treat AI visibility as an extension of good content work, not as a separate playing field with tricks. The goal is this: every important question a prospect would ask has a clear, verifiable answer on the website, in a passage that is understandable even without the rest of the page.

The alternative is to build a special machine layer: additional files, landing pages just for bots, mass-produced question-and-answer pages. That costs effort, is hard to maintain and carries a risk. Content that exists solely for machines contradicts the search providers’ quality guidelines and loses out if it does not fit the visible offering.

A second alternative would be to rely on ads or directory listings. Both can make sense, but they do not replace your own source. Directories are sometimes used by systems as sources themselves, which argues for clean listings. It does not make your own website superfluous, though, because only there does the answer appear in your words.

Our decision: foundations first, then measure, then refine. That is slower than a promise, but it holds up when the systems change.

An example of the difference: the trick approach would generate twenty pages on “tax advisor for professional group X”, each with interchangeable text. The other approach writes one good page on the three professional groups the business actually serves and answers there, specifically, what is different for that group. Fewer pages, more substance, and each of them can be stood behind in front of a client.

The foundation is the entity, that is, the unambiguous identity of the business. A system has to understand: which company is this, what does it offer, where, for whom? That sounds banal but often fails on details. The home page has a short form of the name, the legal notice the full company name, the industry directories an old address. We standardise name, address, service description and catchment area at every place we can influence.

Then comes the passage. A paragraph that is meant to be cited has to stand on its own. It starts with the answer instead of a lead-in, names subject and place explicitly (“The firm advises freelancers in region X on …”) and avoids back-references such as “as described above”. A rule of thumb: if you copy the paragraph and paste it into an empty message, it should still make sense.

Fact blocks are short, easy-to-find areas with verifiable details: scope of services, catchment area, availability, how an intro call works, billing method, qualifications. Where prices cannot be stated as a flat figure, the block says what they depend on instead. That is more honest than a token number and just as useful to readers.

The FAQ pattern complements this. Real questions from consultations are given answers of two to four sentences. It matters that the answers really say something. An FAQ made of empty phrases delivers no citable section, only noise.

The technical side starts with access. The robots.txt states which crawlers may retrieve which areas. The big providers use separate identifiers, for example for training models and for search or for retrievals on behalf of a user. Anyone who blanket-blocks all unknown bots may also block being named. We check the file line by line and record the decision: allow training or not, allow search retrieval or not.

This is a real trade-off. Blocking training does not necessarily prevent being named in answers that work via search. Allowing everything means handing over content for training purposes. For many service providers with a public information offering, the interest in visibility argues for allowing it; for exclusive specialist content the weighting is different. The decision belongs in the hands of company management, not of the technical side.

Structured data according to schema.org complement the entity in machine-readable form: Organization or LocalBusiness, Service, FAQPage, Person. This helps understanding but does not guarantee a mention. What matters is that the markup matches the visible content. We check it with the Rich Results Test and the Schema Validator.

As for llms.txt: it is a proposed format, a Markdown file with a curated overview of the most important pages. Google does not evaluate it, and reliable use by other providers is not documented. We treat it as an optional extra with little effort, not as a lever. It replaces none of the other measures.

The most common case in practice is a contradiction between your own sources. A business has moved, but a directory still shows the old address, and an assistant outputs it. No measure on the website alone helps here. We keep a list of the places where the business is listed on the web and reconcile them. That is tedious and unspectacular, but it removes a real source of error.

A second case is content behind scripts. If the main text is only generated in the browser via JavaScript, some crawlers see just an empty shell. We check with the retrieved HTML, not with the rendered view, whether the core text is present. For classically built or server-rendered pages this is usually unproblematic.

A third case: the system mentions the business, but wrongly. It confuses it with a company of the same name or names a service that is not offered. This cannot be “repaired” at the source. What helps are unambiguous self-descriptions, links to established profiles and, if in doubt, the correction routes that individual providers offer.

A fourth case concerns content with liability relevance, for example in law, tax or health. There, a briefly cited sentence without its context can mislead. We therefore word key statements so that qualifications sit in the same paragraph (“as a rule”, “depending on the individual case”) instead of putting them in a footnote at the bottom of the page.

Finally there remains the case that for a question there are simply better sources: a specialist portal, a chamber, an authority. Then the business is not named, and that is not wrong. It makes sense to recognise such questions and start where the business has something of its own to contribute, such as local experience or concrete procedures.

We deliberately name no figures here, because there is no verified project data, and a forecast would be invented. What can realistically be expected: the website becomes clearer, the content becomes more usable for readers, technical obstacles disappear. Whether mentions result only shows in the measurement, and the result can also be “no change”.

The measurement has three parts. First, a fixed prompt set of twenty to thirty realistic questions a prospect would ask. They are put to the relevant systems every few weeks, and the answers are saved and evaluated: is the business named, linked, described correctly? Because the answers fluctuate, what counts is the trend over several runs, not the single hit.

Second, analytics. Visits via assistants sometimes appear with their own referrer, such as chatgpt.com or perplexity.ai, and sometimes without a recognisable origin. We set up a segment for these sources in the reporting and treat the number as a lower bound. Third, Search Console for the classic part: impressions, clicks and index coverage of the revised pages.

We allow at least two to three months before drawing conclusions. What cannot be guaranteed is each individual mention, an increase in enquiries and the time until a provider picks up a change.

The most important insight is a connection: the largest part of the work is classic SEO with a different weighting. Indexability, clear page structure, content with real substance, clean references and an unambiguous local presence are needed in both worlds. Anyone who neglects these foundations will hardly achieve anything even with additional AI measures.

What is added is a changed perspective. In classic search you optimise a page for a topic and hope for a click. With AI answers you optimise a passage for a question, and the click sometimes does not happen because the answer is already in the chat. The benefit shifts from traffic to awareness: the name comes up, and the prospect later searches for it specifically.

This includes mentions outside your own site. Specialist articles, association pages, press and maintained profiles increase the likelihood that a system links the business with its service. How strong this effect is, we do not know; the effect is plausible and costs nothing if the contributions are useful to people anyway.

A warning applies: not everything that looks good to machines helps readers. A text consisting only of short answers is citable but not enjoyable to read. We write for people first and then check whether the passages also stand on their own.

For the tax advisor from our example this would concretely mean: first an audit clarifies whether all important pages are indexed, retrievable and readable server-side, and what the robots.txt actually regulates. In parallel the fixed prompt set is created, so that the starting state is documented before we change anything. Without this point of comparison nothing could be put in context later.

After that the service pages are reworked, with an opening definition, fact block and a real FAQ. Structured data is added, the directory entries are reconciled, and the llms.txt can be created as an extra if time permits. The effort is manageable; the time goes mainly into clarifying the content: what exactly does the business offer, and how do you say it so that it stands on its own?

After eight to twelve weeks the first evaluation follows, honest even if it turns out thin. If nothing moves, we first check access and indexing, then the passages, and only after that do we look at mentions outside the website. The conclusion of the study is unspectacular: AI visibility cannot be bought, but the ground for it can be prepared.

If you want to know where your business stands today, we begin with the audit and the prompt set. Both are done in a short time and provide a basis on which to decide whether further work makes sense.