If you've ever asked ChatGPT "best accountant for landlords in Bristol" and watched it confidently name three firms, you've probably wondered: how did it pick those three? Why them and not the dozen other perfectly good practices nearby? This guide explains, honestly, what's known about how the main assistants choose which firms to name, and what isn't known. No one outside the AI companies has the exact recipe. But the broad mechanics are well understood, and almost all the signals that matter are within your control.
Where do AI assistants get the names from in the first place?
They don't invent firm names; they retrieve them. Most assistants read across many web sources at the moment you ask, then synthesise a short answer naming the firms those sources mention most credibly for your question.
It helps to picture the assistant doing two jobs at once. First, retrieval: it gathers candidate sources such as directories, your website, review platforms, local listings, and news and industry media. Second, synthesis: it reads those sources and writes an answer, quoting and naming the firms that appear most relevant and trustworthy. If your firm isn't present in the sources it retrieves, you can't be named; you were never in the running. This is the single most important thing to understand: AI visibility is mostly about being present, consistent and clear in the right places, rather than clever copy on your homepage.
Do the different assistants choose differently?
Yes, meaningfully. They lean on different source types, so being visible on one tells you little about the others. They have to be checked separately.
Independent industry analysis across 2025–26 found clear differences in what each engine cites:
ChatGPT leans heavily on third-party directories; around half of its citations come from them.
Gemini tends to favour brand-owned websites; roughly half its citations are the firm's own pages.
Perplexity leans on industry directories and tends to cite the most sources per answer.
Strikingly, only around one in ten cited domains shows up across more than one engine.
// Sources: UK/industry AI-search citation studies, 2025–26. Figures indicative and rounded; treat as direction, not precision.
The practical lesson: a firm can be named confidently by Perplexity and be completely absent from Gemini, simply because its strength is directory presence rather than its own website (or vice versa). If you only check ChatGPT, you've checked roughly a quarter of the picture.
What signals actually decide who gets named?
Six foundations come up again and again: directory presence, reviews, structured data, clear answer-shaped content, a consistent firm identity (your "entity"), and the citations others make about you. Get these right and you're in the running; neglect them and you're usually invisible.
Think of these as the groundwork. Getting them right won't guarantee a recommendation, but neglecting them is the most common reason a perfectly capable firm never gets named.
- 1Directory and professional listings. Because ChatGPT and Perplexity lean so heavily on directories, your ICAEW or ACCA listing, your Google Business Profile, and reputable local and sector directories carry real weight. The key is consistency: the same firm name, address and phone number everywhere. Mismatched details confuse the machines about whether two listings are even the same firm.
- 2Reviews. A healthy, recent body of genuine Google reviews is a trust signal assistants frequently cite. Freshness matters as much as volume; a wall of five-star reviews from three years ago says less than a steady trickle of recent ones. Never buy or fake them.
- 3Structured data (schema markup). Code on your site (your developer will know it as JSON-LD) that spells out, machine-readably, that you're an accountancy practice, where you're based, what you do and how to reach you. It's invisible to visitors but helps assistants read and describe you correctly.
- 4Clear, answer-shaped content. Pages that plainly state who you help and the questions you answer. Say "accountant for limited-company contractors in Leeds", not "bespoke financial solutions". Assistants quote text that directly answers the question, so write the way a client actually asks it.
- 5A consistent entity. "Entity" is just the machine's mental model of your firm: one clear, joined-up identity rather than a scatter of half-matching mentions. The cleaner and more consistent your name, location and specialisms are across the web, the more confidently an assistant can match you to a query and name you.
- 6Citations and earned mentions. When trusted third parties mention you (Accountancy Age, AccountingWEB-type media, local press, a partner's site), those are exactly the sources assistants retrieve and quote. A firm that's mentioned by credible others is easier to recommend than one that only talks about itself.
The signal people forget: can the AI crawlers even read you?
None of the above matters if the assistants' crawlers are blocked from your site. The robots file (robots.txt, at yourfirm.co.uk/robots.txt) can quietly tell bots such as GPTBot, ClaudeBot and PerplexityBot to go away, sometimes added by a previous developer or a security plugin without anyone realising. It's easy to overlook and easy to fix. Practising what we preach, this very site publishes an llms.txt file: a short, plain summary that helps assistants understand what we do.
How much of this can a firm actually control?
Most of it. You can't control the assistant's algorithm or guarantee an outcome, but every foundation above is something you (or your web person) can directly improve.
This is the encouraging part. The signals that decide AI recommendations aren't mysterious black-box magic. They're the things a diligent firm can put right: complete the Google Business Profile, tidy the directory listings, gather genuine reviews, add the schema, write the clear pages, unblock the crawlers, and earn the occasional credible mention. None of it requires gaming anything. It's the same honest groundwork that makes a firm easy for a human to find and trust, which is rather the point.
So why can't anyone guarantee a recommendation?
Because the assistants are third-party tools, run by companies that change their behaviour without notice, and their answers vary run to run. You can stack the odds heavily in your favour; you cannot promise the outcome.
Two honest realities sit underneath this. First, non-determinism: the same question can return different firms on different runs, and independent testing puts the rate at which assistants get facts wrong or inconsistent at roughly 9%. Second, change without warning: an engine can adjust what it cites overnight, so any check is a snapshot of today, not a permanent verdict. That's why the responsible promise is never "we'll get you recommended," but "we'll get the well-understood foundations right, measure properly, and give you the best chance."