How we measure AI visibility.
They score what your content looks like. We measure what AI actually says.
SetpointHQ Pro is the scoring and retainer engine behind The SetpointHQ Index, a longitudinal dataset of UK employer AI visibility. This page documents the methodology in full: the dimensions we score, how they roll up to a composite, and the discipline that keeps every published score consistent.
- Dimensions scored
- 5
- LLM providers probed
- 4
- Probe cadence
- Per audit
AI visibility is not SEO.
SEO measures whether Google ranks your page. AI visibility measures whether ChatGPT, Claude, Perplexity and Gemini cite your company when a candidate asks them about employers. The mechanisms diverge: search engines crawl and rank pages by intent matching; LLMs absorb a corpus and synthesise answers.
A strong position, on a page the candidate may never open.
A mention in the answer they actually read.
A company can rank #1 on Google for its careers page and never be named when a candidate asks an AI best fintechs in London for software engineers. We measure the second of those, which is increasingly the question candidates ask.
Existing tools score whether content looks citable: passage length, statistical density, formatting heuristics. Pro scores whether the company is actually cited when an AI is asked. The two are correlated but not identical, and only one is what candidates see.
Empirical, not heuristic.
Citation likelihood is the centrepiece of the rubric. We probe four major LLMs (Anthropic Claude, OpenAI GPT, Perplexity Sonar, and Google Gemini) with candidate-research queries and observe whether the client is named in the response. This is empirical, not heuristic: we are not scoring whether the content looks citation-ready, we are measuring whether AI cites it.
Multi-engine probing is deliberate. AI visibility means visibility across engines, not against a single model. Each provider has different training data, different retrieval surfaces, and different citation behaviours; an audit that probes only one of them measures one model’s opinion, not AI visibility in any meaningful sense. We run the same query set across all four and aggregate per-provider results into a cross-provider mean.
We do not measure raw search traffic: that is SEO. We do not measure LinkedIn engagement metrics: that is social listening. We measure whether AI cites you when candidates ask AI about employers, which is a question the existing tooling stack does not answer.
Five dimensions, one composite.
Each dimension scores 0 to 100 against an empirical or structural signal. The composite is a weighted average, using the weights below for every published score.
Citation likelihood
Empirical: do four major LLMs cite the client when asked about employers?
We run a standardised query set across Anthropic Claude, OpenAI GPT, Perplexity Sonar, and Google Gemini. We parse each response for mentions of the client by canonical name and known aliases. Position matters: citations at the top of an answer score higher than citations at the end. We average per-provider scores across providers that responded. Providers that errored on every probe drop from the mean rather than score as zero. Citation is the IP centrepiece of the rubric.
Schema completeness
Structural: does the careers site expose the structured data AI crawlers expect?
We check JSON-LD presence and validity for Organization, JobPosting, Person, ProfilePage, BreadcrumbList, and WebSite types against schema.org. Each type's sub-signals weight by which fields it declares. Missing required fields cost more than missing optional ones.
Knowledge-graph presence
External: is the company present in the structured knowledge sources LLMs draw on?
We score Wikipedia entry, Wikidata Q-number, and the completeness of sameAs cross-references in Organization schema. Knowledge-graph presence drives baseline AI awareness independent of the careers site itself. Companies with strong KG presence appear in LLM responses even when their careers content is sparse.
LLM crawlability
Access: can AI crawlers reach and read the content at all?
We score robots.txt permissiveness against a fourteen-crawler list (GPTBot, ClaudeBot, PerplexityBot, Google-Extended, and equivalents). We check for /llms.txt. Render-readiness matters too: whether the careers content sits in raw HTML or needs JavaScript to materialise. AI crawlers that cannot render never see the content regardless of how good it is. Same for crawlers we block.
Social-signal density
Distributed: is the employer brand present where LLMs absorb employer signal?
We score the employer brand on Glassdoor via web-search-mediated probing: review count, recency and average rating.
How the dimensions add up.
The composite is the weighted average of the five dimension scores, on the same 0 to 100 scale. A composite of 50 means the average dimension scored at the midpoint after weighting; a composite of 80 means most dimensions are scoring strongly. The scale is linear, not logarithmic: moving from 70 to 80 represents the same absolute improvement as moving from 30 to 40.
Two design choices worth flagging. First, the citation dimension is weighted highest because it is the closest direct measure of AI visibility. Schema and crawl signals are necessary upstream conditions, but a score on those without a corresponding citation result tells you the plumbing works without telling you the audience hears anything. Second, no dimension can disappear from the composite even when its score is zero: a company that scores 0 on citations because no LLM names them does not get a partial pass for having good schema. The weights are fixed; the composite reflects the full rubric every time.
Composite scores are point-in-time. The same site scored next week may score differently, both because the site changed and because LLM responses are non-deterministic. The longitudinal value comes from running audits at a steady cadence and tracking the trend.
One methodology, applied the same way.
Every brand in a monthly run is scored with the same dimensions, the same weights and the same query set. Scores within a run compare like for like, and the ranking never mixes scoring approaches.
Each score is stored with the inputs that produced it, so any published number can be traced back to the run behind it and checked. That is what makes a published score mean something.
Why this all exists.
The methodology described above produces individual audit scores. The methodology compounds when the same rubric is run across a structured cohort: 101 UK employer brands in known sectors, audited at a steady cadence and scored the same way every month. That cohort is The SetpointHQ Index.
The Index is a public monthly dataset, re-run each month. The v0 cohort ran on 7 May 2026. The current cycle covers 101 UK employer brands from a pool of 103. Not measured this cycle: Capgemini UK (page load timed out), Sky (page load timed out). The full ranked table is browsable at /the-index. The methodology page documents how the score is built; /the-index surfaces what the score actually says about the cohort.
Retainer clients of SetpointHQ Pro are scored under the same methodology. The Index is the public reference cohort; client reports are the private deep work. Both run on the same engine, the same dimensions, the same methodology. Cohort comparisons in client reports (“you score in the top quartile for fintech”) draw from the Index dataset directly.
Questions we get asked.
Why four LLMs and not just one?
AI visibility means visibility across engines. Claude, GPT, Perplexity and Gemini have different training data, retrieval surfaces, and citation behaviours. A score against one of them is a score against one model’s opinion.
A score across all four is a score against AI as candidates use the term. We aggregate per-provider results into a cross-provider mean.
Providers that errored on every probe drop from that mean rather than score as zero. Auth failures are not citation signals.
How do you handle LLM non-determinism?
LLM responses vary run-to-run. We accept this rather than fight it. Single audits produce point-in-time scores; the value comes from running audits at a steady cadence and tracking the trend.
Day-to-day variance is real signal noise; week-over-week and month-over-month trends are the durable measurement. Every score is stored with the inputs that produced it, so when a number moves we can see why.
What’s measured versus what’s recommended?
The rubric measures empirical signals only: LLM probe results, structured data presence, knowledge-graph entries, crawl accessibility, social-platform presence. Recommendations are a separate layer. When an audit identifies a gap, a play library suggests UK-recruitment-specific remediation steps with effort tags and expected impact.
Measurements stay strictly empirical: mixing recommendation logic into scoring would compromise the signal. We enforce the boundary at the architectural level.
Can I see a sample report?
Yes. The sample Pro report walks through one real, anonymised UK employer brand end to end, and every brand in The SetpointHQ Index has its own public page with its dimension scores.
On a discovery call we walk through how the same reading applies to your sector, your current visibility and what you are trying to fix.
See how your employer brand scores against this methodology.
Discovery calls run thirty minutes. We’ll walk through how the methodology applies to your sector, what your current scores would look like, and where the leverage is. No pitch deck, no slides: the methodology is the pitch.