We Asked 5 AI Engines for Recommendations in 15 UK Markets

• No one dominates. Only 2% of the 1,448 recommended brands were named by all five engines, and those few are household names with years of independent coverage.
• One engine tells you almost nothing. Check only ChatGPT and you would miss 79% of the brands the five engines collectively recommend in your category.
• The fastest win is one engine, not all five. Most recommendations come from a single engine that read a source the others did not, so earning one more engine is a content-and-coverage job, not a miracle.
• Check the list. Every brand from the study, with how many engines named it, is searchable here. If you are not in it and you trade in one of these categories, that is your starting line.

This is the first run of the Skavora AI Recommendation Index, a study we will repeat monthly on a fixed category list so the numbers become a trend rather than a snapshot. This first piece explains what we measured, what we found, and, most importantly, what the findings mean for a business deciding where to spend its effort. If the term is new, AI visibility is how often AI assistants mention or recommend your brand; this study measures how differently five of them do it.

What did we measure, and how?

For each of 15 UK categories, from women's activewear to fulfilment providers to electricians, we generated 25 realistic buyer questions and put the identical set to five engines: ChatGPT, Claude, Gemini, Google AI Overviews and Google AI Mode, each answering live with web access rather than from training data alone. We recorded every brand each engine named. All figures in this article come from the roughly 20 non-branded questions per category, the ones that name no brand, because those are the questions that decide who wins a customer that nobody owns yet. The full method and limits are at the end, and they matter: read them before quoting the numbers.

How much do the five engines actually agree?

Very little. Across all 15 categories the engines named 1,448 distinct brands. 29 of those brands, 2%, were recommended by every engine. 1,099, or 76%, were recommended by exactly one engine and never appeared in the other four. The average overlap between any two engines' recommendation lists was 14%: pick two engines at random, and for every seven brands one of them recommends, the other agrees on about one.

The spread matters as much as the average. Trades and services sit at the fragmented end: bathroom fitters produced 92 recommended brands and not one of them was named by all five engines. Established software categories sit at the consensual end: accounting for sole traders had six brands, FreeAgent, Xero, QuickBooks among them, that every engine named. Both ends carry a lesson, and they are different lessons.

What does this mean for your business?

The first practical consequence: checking one engine tells you very little about the other four. In our data, a brand owner who checked only ChatGPT would have seen none of 79% of the brands the five engines collectively recommended in their category. Gemini, the broadest single engine in this run, still missed 63%. If you have ever asked ChatGPT about your own category, seen a competitor, and drawn a conclusion, the conclusion was about one engine, not about AI.

The second consequence is more hopeful, and it is the one we would act on. 76% of every recommendation in this study came from a single engine, which means most of these markets have no incumbent answer. Nobody owns the question. An engine that recommends a brand the other four ignore is an engine that found that brand on sources the others do not read, and earning a place on one more engine's sources is a much smaller job than displacing a household name everywhere at once. Fragmentation is not noise; it is the opening.

The third consequence is about the consensual categories. Where all five engines agree on a brand, that brand is invariably one with years of third-party evidence behind it: review coverage, comparison articles, retailer listings, community mentions. Consensus is what entrenchment looks like from the inside of an AI answer. If your category has consensus brands, the realistic first goal is not to join them but to win the engines one at a time, starting with whichever one your buyers actually use.

How does a brand win while it is still up for grabs?

Start from your situation, because the two ends of the table above need different plans. Find your category, or the one most like yours, and read the row: a high named-by-only-one figure means a fragmented market, a cluster of all-five brands means an entrenched one.

If your market is fragmented (bathroom fitters, electricians, agencies, most trades and services): nobody owns the answer, so the job is to become the best-evidenced option any single engine can find. Concretely: measure where you stand on each engine, pick the one where you are closest to appearing, and look at which sources its answers in your category actually cite, because that list is your outreach plan. In our data those are typically directories, review platforms and local editorial for trades, and comparison articles for services. Then publish direct answers to the questions you were absent from: a page that plainly answers "how much does a bathroom fitter cost in Leeds" is retrievable evidence, and retrievable evidence is what a grounded engine recommends. One engine recommending you is the realistic 90-day goal, and in markets where more than eight in ten recommendations are single-engine, it puts you level with almost everyone else being recommended at all.

If your market has consensus brands (accounting software, payroll, ecommerce platforms): the all-five names are not your first target, and trying to outrank Xero everywhere at once is how a year disappears. Compete on the specific question instead of the category: engines answer "accounting software for a sole trader who invoices in euros" as its own question, not as a subset of "best accounting software", and the specific questions are where challengers appeared in our data. Pair that with presence on the one or two sources your target engine leans on, which your own measurement will name, and re-measure roughly monthly: answers churn, and the churn is your way in as much as anyone else's.

In both cases, the sequence is the same: measure across engines, pick the nearest win, earn the two or three sources that engine trusts in your category, answer the missed questions directly on your own site, then re-measure and compare. It is the measure, fix, re-measure loop, and the reason it works is exactly the fragmentation this study found: the bar for being recommended by ONE engine is low enough for a small brand to clear it this quarter.

Which brands did all five engines agree on?

The full consensus list across all fifteen categories is 29 brands, and reading it is instructive because none of them are surprises: Shopify, Wix, Squarespace, Etsy and BigCommerce in ecommerce platforms; FreeAgent, Xero and QuickBooks in sole-trader accounting; Bose and JBL in wireless headphones; ASICS and New Balance in running shoes; CeraVe and La Roche-Posay in sensitive skincare; Lululemon in women's activewear; Petplan in pet insurance; NICEIC in electricians. These are brands whose independent evidence, reviews, comparisons, editorial coverage, is so widespread that five engines reading five different slices of the web all reach the same answer. The full list of all 1,448 brands, with how many engines named each, is published alongside this article.

The absences are as telling as the names. In bathroom fitting, digital marketing and fulfilment, three real markets with real buyers, no brand at all cleared the five-engine bar in this run. Those answers are currently being assembled fresh each time from whatever each engine happens to retrieve, which is precisely the condition under which a well-evidenced challenger can appear in them.

How do I find out where my own brand stands?

You can replicate our protocol by hand: write 20 or so realistic buyer questions for your category, none of them naming you, ask them in fresh sessions across the engines your buyers use, and log every brand named. It works, and it takes an afternoon per engine. Skavora automates exactly this: the free check runs 25 research-derived questions through Gemini with live web access and returns a scored report in a couple of minutes, no email, no card. The complete AI Visibility Audit, £9.99 once, runs the same questions across all five engines from this study and shows you the per-engine picture: where you appear, who is recommended in your place, and which sources drove each answer. We publish a real one, run on ourselves with a poor score, as an unedited example.

Methodology and limits

• Run on 4 August 2026, one scan per category, 15 categories from a fixed list of 20 (five deferred to the next monthly run for budget reasons: robot vacuums, mattresses, returns management software, employment law solicitors, private dentists). The list is versioned and will not change between runs, so month-on-month comparisons stay honest.
• 25 questions per category, generated from live research, roughly 80% non-branded. All statistics in this article use the non-branded questions only, so a category's seed brand cannot inflate the consensus figures through questions that name it.
• ChatGPT, Claude and Gemini were asked all 25 questions; Claude answered 20 of 25 in one category, and scoring failures left ChatGPT counted on 20 of 25 in two categories, with the affected answers excluded rather than estimated. Google AI Overviews and Google AI Mode were measured on a deterministic, evenly spaced sample of 15 of the 25, because each lookup is billed; their brand lists therefore come from fewer questions, which slightly understates their breadth.
• Brand names were normalised (case and punctuation) before comparison, so "De'Longhi" and "DeLonghi" count once. Brand extraction is performed by the scoring model and carries some noise: a small number of entries are product lines or organisations rather than brands. The 76% figure moved by less than one percentage point when we tightened normalisation, which is the scale of error we believe applies.
• Engines sometimes answer off the category: a question about digital marketing agencies can draw a software platform in the answer, and a sustainable-fashion question a high-street product line. A brand is in the list because an engine recommended it: the call is the AI's, not ours. The question sets will be tightened in later runs to keep answers on the service being asked about.
• This is one run in one month in one country. Engine answers vary between runs, so treat every per-category number as an estimate with real variance, not a ranking. The monthly re-runs on the identical list are what will separate signal from churn, and we will publish those, including any month that contradicts this one.
• The full methodology is published separately: exact models and web-access setup per engine, question counts, the sampling rule, what was excluded and why, and what the study cannot see.

A scan is a snapshot, and no tool, ours included, controls what any engine says. What a study like this gives you is a direction: measure across engines rather than one, find the engine where you are closest to appearing, and earn coverage on the sources it already trusts. Where those sources are for your category is exactly what the engines' own citations reveal.