We asked AI platforms which UK cybersecurity recruiter to use. On a third of the questions, they couldn't agree on a single name.

What we found

Five findings stood out enough to build the rest of this study around. None of them are about which recruiter is "best." They're about how AI platforms actually behave when a real buyer asks a real question.

Where the three platforms actually agreed

For each query we checked whether any single firm appeared in all three platforms' answers. This is the cleanest, fully-verified number in the whole study: no estimation, just a straight count.

Full agreement on 10 of 16 queries; zero agreement on 6. The zero-agreement queries cluster at the senior/executive end (CISO, London, both re-run "validation" queries) and on the one query where a normally-dominant firm had a specific blind spot (SOC).

Same need, different wording, different answer

Three times in this study we asked about the identical underlying hiring need, worded two different ways, and got materially different results. This isn't a one-off quirk. It happened every time we tested for it.

Six ways to compete for the same buyer

Ask an AI platform who to hire, and it doesn't just answer with recruiters. Every one of these business models showed up answering some version of "who should I use."

Methodology

16 queries were run across Google AI Overview, ChatGPT and Perplexity: two broad "who's best" questions, five technical specialisms (CISO, GRC, SOC, penetration testing, security architecture), contract vs permanent hiring, three UK cities (Manchester, Bristol, London), two of the specialism queries re-run with validation-stage wording to isolate the effect of phrasing, and two further queries built from modifier lists the platforms themselves surfaced mid-study (IR35 compliance, fractional CISO). Four additional control queries confirmed the paid-search finding.

Every answer was recorded in full, including which firms appeared on which platform, any "Missing: [term]" annotations Google showed against a firm's own page, People Also Ask questions, related searches, and any clarifying question a platform asked before narrowing its answer. All 16 queries have since been re-checked directly against the original screenshots, not just the working notes taken at the time.

Every one of the roughly 120 named agencies, consultancies and individuals across the 15 recruiter-focused queries was tallied by platform. "Google" counts appearing in either the AI Overview text or the page-1 organic listing. The table below shows the top of that ranking; the full list, down to firms that appeared on a single query, is in the downloadable data.

Robert Walters appeared on 14 of the 15 queries, more than any other firm, but two-thirds of its mentions are on Google; Perplexity named it only twice. La Fosse shows the same shape against ChatGPT. LT Harper and Lorien both show a genuine zero or near-zero on one platform (ChatGPT) while performing solidly on Google. Barclay Simpson has the most even spread across all three, and the highest total, but still drops out entirely on the SOC-analyst query. Full rankings for all ~120 entities, including the long tail of single-query appearances, are in the downloadable dataset.

The "missing word" pattern

Google flags when a result's own page doesn't contain a word from the query, showing a "Missing: [word]" note next to the listing. Across the study this happened often enough, on genuinely relevant words, to be a real and checkable pattern rather than noise.

ARM Recruitment and LT Harper each carry the gap on more than one word, across more than one query. Robert Walters' is the sharpest single example: its own GRC recruitment page doesn't contain the word "GRC."

Frequently asked questions

Why we didn't just automate this

A growing number of GEO/AEO tracking platforms offer exactly this kind of study as a one-click, API-driven product: run a few hundred or a few thousand prompts automatically, tally mention rates, publish a report. One competing UK recruitment AI-visibility study we came across claims 1,200 prompts across 12 sectors and 8 cities, run three times each for statistical reliability, a genuinely large dataset.

What that kind of study can't easily show, because no one is reading the individual answers, is why a firm was included or excluded, or what happens when the same question is asked a different way. Nothing in an automated mention-rate table would have caught ChatGPT reversing its own advice between two phrasings of the same hiring question, or a search result stating a company's own page as "Missing" the exact word a buyer searched for, or Google's AI Overview naming a completely different firm depending on which of three cities you asked about. Those are the findings that actually explain the mechanism, not just measure its size, and they only surface when a person reads what the AI actually said and checks it against what's really on the page.

It's also worth asking what a mention-rate number is actually measuring. A flat "appeared in 23% of prompts" figure doesn't say whether that 23% came from broad awareness-stage searches, specific comparison-stage searches, or the moment right before someone actually picks up the phone, and those are very different commercial outcomes for a real business. This study was built around that buyer journey deliberately: broad questions, five technical specialisms, contract versus permanent hiring, three cities, and two pairs of queries re-run with different wording specifically to test whether phrasing alone changes the answer. A single-click prompt-and-tally approach, however large the number of prompts, doesn't naturally produce that structure, because it's built to be run the same generic way for every industry.