How Do AI Engines Decide Which Brands to Recommend?

They recommend whoever they can describe confidently: a brand whose details are readable in plain HTML, stated as facts, and phrased the same way in several places. What differs between engines is the index behind them and what each one rewards on top of that.

First a distinction that changes everything below it: being cited is not the same as being recommended.

Cited means your page was used as a source for an answer. Recommended means your name appears in the answer as an option to choose. You can be cited in an answer that recommends somebody else. It happens constantly on “top ten agencies” listicles: you wrote the list, and the engine uses it to recommend the people on it.

If you want the short, business-facing version, we wrote it separately in how to get ChatGPT to recommend your brand. This piece is the technical half: what is common to all the engines, what differs, and how to test without fooling yourself.

The foundation, common to all of them

The content has to exist in the HTML. Googlebot renders JavaScript, late and with no guarantee. The crawlers used by AI assistants generally do not. Text that only appears once a script has started may not exist for them. This is the single biggest practical difference from classic SEO, and it is invisible if you only look at your site in a browser.

Facts beat adjectives. A model cannot quote an impression. It can quote a sentence containing a fact: what you do, who for, since when, at what price or at least in what range, in which markets. “Creative solutions with impact” cannot be extracted from because there is nothing in it.

A paragraph has to survive being lifted out. An answer is assembled from pieces taken from sources. A sentence that loses its meaning without the three paragraphs above it will never be taken.

The same description everywhere. This is the part most brands miss. If you describe yourself one way on the site, another on your Google listing, another on LinkedIn, and another in an old interview, the model sees four versions that contradict each other and picks someone whose story is consistent. It is not about volume, it is about agreement.

What each engine rewards

The engines do not share an index and do not weigh things the same way. What follows comes from research on each platform rather than from our own measurements, so treat it as a map, not as law.

ChatGPT leans on the Bing index plus what it absorbed in training. It rewards content shaped like an answer: written plainly and conversationally rather than promotionally. Freshness counts for more here than elsewhere.

Perplexity combines its own index with search results and rescores them. It rewards question-and-answer structure, paragraphs that stand on their own, and a steady publishing rhythm. Downloadable public documents do well.

Google AI Overviews and Gemini lean on the Google index and the knowledge graph. What matters most here is that the entity is unambiguous: who you are, where you are, what you do, confirmed in several places. Structured data helps directly.

Claude searches through Brave. It cites selectively and prefers factual density: numbers, dates, named sources.

Copilot runs on Bing. Presence on professional platforms weighs more here than in other places.

The practical conclusion: you do not optimise separately for each one. You get the foundation right, then adjust the form. You do not make five sites.

The files people talk about too much

robots.txt is the only one with a guaranteed effect, and the effect is negative if you get it wrong: you can block AI crawlers by accident. On this site GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot and Google-Extended are explicitly allowed. You can open robots.txt and check.

llms.txt is a text file at the root describing briefly what you do. It is not required, nobody guarantees it does anything, and it is not a visibility switch. It is useful as a clean source about you, and that is all. Ours is at llms.txt, you can read it.

What neither of them does: put you in answers when the rest is missing.

How to test, and the trap we walked into

The method is simple. Open the engines and ask the questions a buyer would ask, not your company name. “Which packaging design studio would you recommend for a coffee brand launching in the EU.” Note whether you appear, where in the list you sit, and which sources get cited. Repeat monthly with exactly the same questions.

Now the trap. We ran this test with our own account signed in, and Perplexity opened its answer with a reference to our own site, because it knew who was asking. The engine personalises by account.

If you test your visibility while logged in, you get an answer about yourself and walk away convinced you are visible. The test is run from a private window, with no account. It is an easy mistake to make and we have not seen it written down anywhere.

One more caution: if you go looking for advice on the subject, do not search for how to check whether you appear in ChatGPT. You get AI text detectors, an entirely different subject.

What does not work

Adding yourself to somebody else’s “best agencies” list. If the list recommends ten firms, the engine can use it to recommend any of them, not you.

Asking for fabricated reviews. Consensus is assembled from sources that confirm each other, and false sources contradict each other faster than you expect.

Expecting results in weeks. A new mention has to be found first, then repeated often enough to count.

How we think about it

We run SEO and GEO as one job, because the foundation is shared. The part strictly about how engines understand you as an entity and attribute you correctly has its own page.

Everything written above is applied on this site, and the files we are talking about are public and you can open them right now. It is the only honest way to sell this.

FAQ

Do AI assistants recommend UK brands differently? They use the same signals, but they check local sources: UK reviews, directories and press decide whether a British brand gets named.

Is being cited by an AI the same as being recommended? No. Cited means your page was used as a source. Recommended means your name appears in the answer as an option to choose. You can be cited in an answer that recommends somebody else, which happens constantly on top-ten listicles.

Do I need an llms.txt file? It is not required and it guarantees nothing. It is a text file at the root of the site describing briefly what you do, useful as a clean source about you. We have one, you can open it at /llms.txt, but do not treat it as a visibility switch.

How do I test whether I appear in AI answers properly? From a private window, with no account logged in. If you are signed in, the engine can personalise the answer around who you are and show you your own brand, which makes you think you are visible. Ask what a buyer would ask, not your company name, and repeat the same questions monthly.

Does it matter that my site uses JavaScript? Yes, more than it does for Google. Googlebot renders JavaScript, late. The crawlers used by AI assistants generally do not. Text that only appears after a script starts may not exist for them.

How long before an AI engine cites you? Months, not weeks, and it depends on the engine. The sources they lean on refresh at different rates, and a new mention has to be found first, then repeated in enough places to count as a signal.