How does AI decide which businesses to recommend?

Almost every statistic in this field is published by a company that sells a product the statistic happens to justify. That is not a conspiracy, it is just how the market has developed. It means you have to read the methodology before you read the headline.

Here is a worked example, because it is the pairing I see most often on slides.

One widely quoted figure is that 85% of AI citations come from earned media. The real source is Muck Rack, the number is 84%, and it comes from a solid dataset: more than 25 million links across ChatGPT, Claude and Gemini responses in 17 industries, published May 2026.

The problem is what "earned media" means there. It is a residual bucket, meaning everything that is neither owned by the brand nor paid for. Muck Rack's own breakdown reads:

Strip out the vendor framing and a consistent picture does emerge. Here is the range across the largest datasets published in the last year.

The single most useful finding I have come across is not about PR at all. AirOps tested over 500 commercial-intent B2B queries across GPT-5, Claude Sonnet 4.5 and Perplexity Sonar, and analysed 21,311 brand mentions. They found that nearly 90% of third-party mentions came from listicles, comparison pages and reviews, not from journalism.

Meltwater's May 2026 data points the same way from a different angle: across more than 8 million citations, YouTube was the single most-cited source, with Reddit second.

Sit with that for a moment, because it reframes the whole exercise. When a broker asks ChatGPT to recommend a claims management partner, the engine is far more likely to be reading a "top 10 claims management companies UK" listicle, a Reddit thread, a review platform or a YouTube transcript than it is to be reading a piece in the trade press.

That does not mean press coverage is worthless. It means the "how AI decides" diagrams you see in PR decks, the ones listing press coverage, founder authority and awards, are describing a real but partial picture, and the parts they leave out happen to be the parts nobody can sell you a retainer for.

I have yet to find a single study that isolates awards as a citation driver. It is plausible that awards help, via the coverage and the credibility signals they generate. It is a second-order effect being presented as a first-order one, and you should price it accordingly.

There is one decent piece of evidence that it correlates, and it is worth quoting properly rather than in the inflated form.

Hard Numbers, a UK B2B PR agency, worked with the media monitoring firm Onclusive on a study published in November 2025. They took 143 global brands, scored them on Onclusive's media influence measure, and ran consistent queries through ChatGPT, Perplexity and Gemini.

What they found: brands in the top 10% for media influence were recommended around 80% of the time, against under 50% for the bottom 10%. That is roughly 1.6x. Separately, in head-to-head comparison prompts ("which is better, A or B?"), brands in the top decile for positive sentiment won about three times as often as those in the bottom decile.

The version circulating on slides is "brands with positive earned media are 3x more likely to win AI recommendation". That is three distortions in one line. It collapses a top-decile-versus-bottom-decile contrast into a simple binary. It takes a finding about a narrow prompt type and presents it as general. And it turns an observational correlation into a lever you can pull.

The correlation is real and I would not dismiss it. Brands with high media influence also tend to be larger, older and better known, and all of those independently predict whether a language model has heard of you. The study cannot separate those effects, and it was produced by a PR agency and a monitoring firm, so read it as informed advocacy rather than neutral research.

Based on what the evidence supports rather than what is easiest to sell, here is where I would put the effort for a UK service business.

Ask the engines directly, but do it properly, because the obvious method gives you a flattering false answer.

If you ask ChatGPT what it knows about your company while logged into your own account, it has access to everything you have ever discussed with it. You are looking in a mirror. Use a fresh session, incognito, or a device that has never been used to talk about you. Better still, ask someone outside your business to run the same questions.

Then be systematic about it. Use the questions your buyers would genuinely type, not your brand name. Run them across the four engines that matter for UK service businesses, which are ChatGPT, Google AI Overviews, Perplexity and Claude. Record who gets named, which sources are cited, and whether the description of you is accurate. Repeat it on a schedule, because the answers move.

Every figure on this page comes from someone else's dataset, and I have told you who and why they might be motivated. That is the honest position today, but it is not a good enough position.

Over the next year Optimus is compiling its own citation-mix data from audits of UK service businesses, specifically insurance broking, motor claims, credit hire, leasing and legal. UK-only, service-sector only, real commercial queries rather than consumer search. It will not be the biggest dataset in the field. It will be the only one built on the questions your actual buyers ask, and it will be published with its methodology open so you can argue with it.

If that is useful to you, the audit is where it starts.