Getting found in ChatGPT, Gemini and Perplexity

A year ago a customer started his search at ten blue links. He searched, clicked, compared and made up his own mind. Now he types his question into ChatGPT and gets one answer back, with a handful of names in it. If you are in that answer, you get considered. If you are not in it, then for that person you briefly do not exist.

Take an entrepreneur looking for an accountant in his region. He used to type "accountant Alkmaar" into Google and look at the first five results. Now he asks ChatGPT which accountant suits a small business with a lot of foreign invoices. The model gives three names. The accountant who is among them gets the conversation. The other twelve in town never come up at all.

You can imagine what that does. The businesses AI names get the enquiry, the conversation and often the work, without the customer ever having clicked a search result. Most entrepreneurs have no idea whether AI names them. That is exactly the problem we solve.

An AI model does not invent names. It draws on what it has learned about the world and on what it finds on the web at that moment. Broadly speaking it looks at four things.

Whether you are mentioned, and where

Models lean heavily on what other sources say about you: trade publications, media, review platforms and communities. The more often independent sources mention you around the same topic, the sooner a model dares to recommend you. Research across large numbers of brands shows that brand mentions on the web correlate more strongly with visibility in AI answers than the number of backlinks does.

Whether your entity holds up

A model wants to be sure who you are. If you are described slightly differently everywhere, with another company name, another address or a vague description, your profile stays hazy. When it is consistent everywhere you become a recognisable entity that the model can name with confidence.

Whether you are the answer

AI cites content that answers the question directly, in plain language, without beating about the bush. A page that says straight away what something is and who it is for is easier to cite than a page that only gets to the point after four paragraphs.

Whether it can read you

Before a model can say anything about you, it has to be able to get in. If your server or your CDN blocks the AI crawlers, if your text only appears once JavaScript has loaded, or if your server response takes too long, the crawler drops out and the model misses your whole story. Access is the base that everything else rests on.

We work in layers and we begin with the measures the evidence supports most strongly. That is a different order from the one you see at most agencies, where the schema work comes first.

Access for the crawler

First we make sure the AI crawlers get into your site and take all of it in. Four things, in this order:

• allow OAI-SearchBot, GPTBot, PerplexityBot and the other AI crawlers in robots.txt, and check that your CDN or firewall is not blocking them anyway
• render your content server-side, because none of the big AI crawlers execute JavaScript
• make your server response fast, because these crawlers work with short timeouts and do not come back to wait
• register your site with Bing Webmaster Tools and switch on IndexNow, because ChatGPT Search draws largely on the Bing index

One clear entity

Next we make plain who you are: one canonical company name, one description that is the same everywhere and a sameAs network that adds up, from your LinkedIn to your chamber of commerce listing. This is the part where we see the most gain in practice, because a model that has doubts about your identity would rather not name you.

Clean structured data and llms.txt, as hygiene

We put schema markup in place properly, without a visibility promise attached to it. Cited pages carry JSON-LD more often than uncited ones, but the only controlled study that isolated the effect found no measurable rise in citations. It helps Google understand your page and it costs little, so we do it. We also publish an llms.txt, for the same reason: it costs almost nothing. Today there is no evidence that the platforms fetch it, because a measurement across 137,000 sites found that 97% of those files had never been requested. We would rather tell you that now than have you discover later that you paid for it. There is more about this under what we do not do.

Giving a real answer

We write your core pages so that the model can lift a ready-made answer out of them, with your name attached. That means answering the question directly at the top of the page and working out the rest underneath.

Building authority

We make sure you are described consistently everywhere and that your existing mentions count. On Growth we also arrange an article about you in the trade press every quarter: we approach the editors, write the piece and see that it gets placed. If you want one every month we scale up, because that takes more time and the cost depends on which titles you want to reach.

Measuring and adjusting

Because AI answers differ from day to day, we do not measure once. We measure your position across several models and several days, against your competitors, so that your advice rests on figures and not on one chance result. How we do that is set out in the Findori Visibility Framework.

AI visibility counts most for businesses where a customer does some research before choosing: B2B service providers, software companies and specialists with a clear niche. Think of an accountancy firm, a lawyer with a specialism or a software company that sells to other businesses. If your customer puts his first question to an AI, you want to be in that answer. For online shops and purely local services it works out differently: there, the largest academic study so far measured that AI platforms deliver less than 0.2% of sessions, so there we advise something narrow and inexpensive rather than a monthly programme.

What you can realistically expect

The technical improvements often come through within weeks. For the first thirty days we promise nothing, because in that period an improvement cannot be told apart from daily fluctuation. That month we measure. After that our target covers the ninety days that follow, and every month you see whether the line is heading the right way.