A nurse thinking about a move asks ChatGPT which care providers near her are good to work for. A graduate asks Gemini which consultancies have the best training. A warehouse supervisor asks Perplexity who pays fairly for night shifts.
None of them gets a list of ten links. Each gets one paragraph that names three to five employers, with a line or two on why. That paragraph is the shortlist. If your organisation isn’t in it, you weren’t ranked lower. You weren’t there at all.
For twenty years, employer-brand discovery meant Google, Glassdoor and your careers page. Most of what we learned there still matters. But a growing share of candidate research now happens inside an assistant, and the assistant plays by different rules.
From ranking to being named
Search engine optimisation was about position. Ten results shared a page; come fourth and you still got clicks.
An answer engine works differently. The model reads what it can reach, decides who is worth naming, and writes a single answer. There is no page two. There isn’t really a page one. There is the answer, and there is everyone else.
That changes the question employer-brand teams need to ask. “Do we rank?” is no longer enough. The question is “Are we named, and what is said about us when we are?”
It also changes who writes your first impression. When an assistant answers a question about your pay, shifts or culture and your own site has nothing quotable on the subject, it answers anyway, using whatever it can find: a Glassdoor page, a two-year-old Reddit thread, a local news story about a restructure. A candidate who then clicks through to your careers site arrives with expectations set by strangers.
What the data shows: known, but rarely recommended
At SetpointHQ we score 100 UK employer brands across 14 sectors every month for The SetpointHQ Index, a free public dataset. Each brand is probed with candidate-style research questions across four assistants (ChatGPT, Claude, Gemini and Perplexity) and scored from 0 to 100 on five dimensions: citation likelihood, schema completeness, knowledge-graph presence, crawlability and social signals. The rubric is published and versioned. The figures below come from rubric 2026.06.5, in the cohort run dated 1 October 2026.
One pattern stands out above the rest.
- The median knowledge-graph score is 89.3. Machines almost always know who these organisations are.
- The median citation score is 26.3. When a candidate asks an open question such as “who are good employers in my field?”, these same brands are rarely the ones named.
- By my count from the published table, only 17 of the 100 brands score 50 or more on citation.
Put simply, nearly every brand is recognised, and roughly one in four is actually recommended. Being known and being recommended often get treated as the same achievement internally. The systems doing the recommending measure them separately.
The technical layers sit between those two figures. Median crawlability is 75.0, and no brand scored zero, so most careers content is at least partly reachable by AI crawlers. Schema is weaker: the median is 35.0, and 19% of brands scored zero, meaning there’s no structured data an assistant can read to learn who they are, where they hire or what a given job is.
The brand-level data also shows that no single signal is enough. In the same run, Deloitte had one of the higher citation scores in the cohort (71.3) despite a crawlability score of 15. Octopus Energy scored 100 on crawlability and 85 on schema but only 20 on citation. Technical hygiene makes citation possible. It doesn’t guarantee it. (Inclusion in the Index is not an endorsement or a client relationship; these are public scores for well-known employers.)
A caution on method, because practitioners should be suspicious of neat numbers. These are point-in-time readings, and LLM responses vary from run to run. The social dimension currently draws on Glassdoor only; Indeed, LinkedIn and Reddit are planned for a later version. The longer-term value is in the trend, measured the same way each month, not in any single score.
Why a good careers site and solid SEO aren’t enough
Most employer-brand teams I speak to assume a strong Google presence carries over to AI assistants. Almost every brand in the cohort ranks for its own name, yet median citation is still in the mid-twenties. The gap usually comes from a few places.
Jobs live where the machines can’t, or won’t, read them. Many careers sites hand job content off to an ATS subdomain: Workday, Greenhouse, Phenom and the like. That content often renders client-side in JavaScript, sits on the vendor’s domain and sometimes carries a noindex tag. Any authority it earns goes to the vendor, not to you. You don’t need to rip out your ATS, but you do need indexable, server-rendered job content on a domain you control.
Pages are quietly ineligible to be quoted. Google’s guidance on AI Overviews and AI Mode is refreshingly plain: there are no special optimisations and no special schema required. Pages must be indexed and eligible to show a snippet. Careers sites fail that test more often than you’d expect, through nosnippet, data-nosnippet, restrictive max-snippet settings or a stray noindex. Anyone selling you “AI Overview schema” is selling something Google says doesn’t exist.
The questions candidates ask have no page to answer them. Pay bands. Shift patterns. What happens at the second interview. Whether anyone gets promoted after two years. Assistants answer these questions today, whether or not you’ve written anything. If you haven’t, they answer from someone else’s material.
Job markup is missing or stale. JobPosting and Organization structured data tell a machine exactly what a role is, where it is and who is hiring. Roughly one in five brands in the cohort publishes none, and where it exists it’s often stale: expired roles, missing salaries, vague locations.
The employer story can’t be quoted. A lot of EVP work ends up in a hero video, a PDF brochure or a carousel of adjectives: “passionate”, “dynamic”, “people-first”. None of it gives a model a sentence it can lift and attribute. “We offer a 37.5-hour week with paid breaks and a pay review every April” is quotable. “We value our people” isn’t.
Corroboration lives off your site. Assistants weight independent evidence more heavily than a brand’s claims about itself. Research on what AI summaries cite has found that Wikipedia, YouTube and Reddit alone supply roughly 15% of referenced links. Trade press, review platforms and forum threads where real employees talk all count.
Practical checks you can run this month
None of this needs a vendor. A web developer, someone with Search Console access and an afternoon will get you a long way.
- Ask the assistants yourself. Write down 10 questions a real candidate would ask, such as “best employers for [role] in [city]”, “what is it like to work at [you]” or “does [you] pay well for [role]”. Run each one in ChatGPT, Claude, Gemini and Perplexity. Record whether you’re named, where in the answer you appear and which sources are cited. Repeat monthly; one run is a sample, not a verdict.
- Check snippet eligibility and indexing. View the source of your top careers pages and job pages. Look for noindex, nosnippet, data-nosnippet and max-snippet. In Search Console, confirm job pages are indexed, remembering its AI reporting covers Google’s surfaces only, not ChatGPT, Claude or Perplexity.
- Find out where your jobs actually live. Click through from your careers page to a live vacancy. If you land on a vendor subdomain, disable JavaScript and reload. If the job disappears, many crawlers won’t see it either.
- Check your robots.txt against AI crawlers. Decide deliberately which ones you allow, rather than inheriting a template written years ago.
- Build one page per real candidate question. Start with the five questions your recruiters hear most. Answer each one in the first 40 words, in plain text, above any accordion or tab. Accordions hide content from extraction more often than people expect.
- Replace adjectives with numbers. Go through your EVP copy and swap every claim you can for a fact: pay range, hours, training days, internal promotion rate, retention. If you won’t publish a number, accept that the assistant will find someone else’s.
- Validate JobPosting and Organization markup. Run live job pages through Google’s Rich Results Test. Check salary, location, employment type and closing dates are present and current.
- Claim and tidy your entity. Check that your Wikidata entry, Wikipedia page (if you have one) and the sameAs links in your Organization schema all point to the same, correct organisation.
- Measure citation, not just rank. Add “named in X of Y candidate questions across four assistants” to your employer-brand reporting alongside traffic and applications.
The scoreboard is still mostly dark
You can’t dictate the words an assistant uses about your organisation. Anyone who says otherwise is overselling. You can strongly influence what the assistant has to work with: content it can reach, structure it can parse, an entity it can resolve, facts it can quote and corroboration it can trust.
Most employers have never measured any of this outside Google, and the median citation score suggests the answer seats are mostly empty. That’s the uncomfortable part. The encouraging part is that most of the fixes are unglamorous, cheap and within the reach of an in-house team.
If you would like a reference point, The SetpointHQ Index is free to browse, and the methodology is public. Look yourself up, or a competitor, then run the ten-question test above and compare. Either way, start measuring. The shortlist is being written whether or not you’re watching.
