Ask AI about Pagefield. Read it, and then in another chat, ask again. The second answer won’t quite match the first. The language shifts. A source appears that wasn’t there before, while another quietly drops out. The emphasis lands elsewhere, an old stat comes in as new, comparisons change shape, claims strengthen and soften. Nothing is necessarily wrong. It’s just how these systems work. They are probabilistic systems, which means the same prompt can produce different answers depending on context, model behaviour and variables no user can fully see. Variation is simply a property of the medium.
For many of the things people ask AI for, that instability is harmless. But questions about organisations or issues are different, because those answers have serious implications: on the bottom line, on regulation and policy and on what’s impressed upon key stakeholders and decision-makers. Clients, investors, policymakers, prospective employees — they’re all increasingly likely to ask an AI before they ever ask you. By the time a conversation begins, a view of your reputation has often already been formed, a variation on an answer you’ll probably never see, shaped by sources you might not know exist.
Or you might have experienced the feeling of urgency already. Someone sends around a panicked screenshot showing an AI response about the company or a relevant issue. I understand the instinct completely. The answer looks like evidence. But one AI answer is not measurement; it is an observation, a single lottery draw from a whole distribution of possible answers.
The trap has two doors, and we’ve seen organisations walk through both. If you treat the screenshot as a universal truth, you may reorganise a communications strategy around an answer that only represents what a small number of people are actually seeing. If you dismiss it as a one-off, you may be waving away the visible edge of a pattern that repeats across hundreds or thousands of answers a day. From a single observation, it’s impossible to tell which. That is the bigger problem.
Why counting mentions isn’t enough
Getting past it means looking at many answers, and not just one. The market’s first instinct in this category has been to count — mentions, citations, rankings, share of voice — treating AI responses as purely an evolution of SEO. Counting has its place, but rankings and mention rates are only the residue of narrative, not narrative itself. Counting tells you how often you appeared. It doesn’t tell you anything about the characterisation of an answer: how you’re framed when the question turns hostile, whether an investor is hearing a different story from a customer, whether the machine volunteers your worst moment without being asked about it. Those are the questions the organisations we work with lose sleep over. It’s not always about visibility; more often, it’s about narrative, sometimes told only to a small and particular audience.
Stakeholders don’t only encounter your reputation by asking about you directly. They also turn to AI to summarise issues, policies, legislation, and live debates relevant to you, but in which AI may have decided your voice is not important. For an organisation with a legitimate stake in a debate, absence from AI summaries or sources risks distortion or misrepresentation of the debate itself. Decision-makers now use AI to get up to speed before they act. While it has always been possible to be absent from the room, it’s now increasingly possible to be absent from the briefing as well.
What AI makes measurable
Until very recently, almost none of this was measurable, and the reason it is now is the most misunderstood thing about AI. The dominant story is that AI’s gift to communications is generating things. There is some truth in that, and I’ve seen it already changing production. But the more consequential gift in my mind runs the other way. These models can read: narrative, framing, tone, the positioning of one organisation against another. Material and nuance that used to resist computation entirely can now be encoded at scale into structures a computer can count, compare, and track. And once meaning has been read into structure, the rest of the work can be handed to the kind of computing we have trusted for seventy years: the kind where the same inputs produce the same outputs and every figure traces back to its source.
Software engineering has a long-standing distinction between a discipline’s accidental labour (the work created by the tools around the task, like writing) and its essential difficulty, the part that belongs to the problem itself. Generative AI is stripping accidental labour out of communications at speed. The essence, judgment about what to say, to whom, when, and whether to say anything at all, remains largely unchanged and unmoved. The serious question was never which model will be best, but how to assign the roles, and when to hand over to expertise.
That’s how we build our systems at Pagefield Digital Lab. Machines take the work where precision and repeatability are required. Models do the one thing only they can do, which is automate reading meaning at scale. People keep what cannot be delegated: deciding what matters, strategy and creativity, honed intuition, and taking responsibility for the advice.
Measuring through uncertainty
The machines’ share is, reassuringly, old territory. Much of that approach was built in London almost one hundred years ago, just a few miles from Pagefield’s Soho office: Jerzy Neyman set out the modern theory of stratified sampling at the Royal Statistical Society in 1934 and introduced the confidence interval three years later. By 1945, the discipline had its first major public test, when survey researchers pointed to a clear Labour victory while most of the press and political class expected Churchill to return. If anything, the statisticians understated the landslide.
The discipline’s achievement was never the elimination of uncertainty but making it central to its methodology. That is the tradition our audits are built on: designed samples rather than ad hoc prompting, stratification across the audiences and question types that matter, and a margin of error that we can stand by. Without that, all you have are claims pretending to be evidence.
Pagefield GPS
There’s a cliched reputation in comms that technical people don’t fully get the essence of what matters. If you allow me one nerdy indulgence, I’d like to prove that wrong. Richard Hamming, one of the great applied mathematicians of twentieth-century computing, printed a motto at the beginning of his 1962 book on numerical methods: “The purpose of computing is insight, not numbers.” Sixty-four years later, I struggle to find a better north star. The discipline doesn’t exist to manufacture or sell data for the sake of it. It exists to make human judgement better.
Pagefield GPS was built on this principle. It treats each AI answer not as the answer but as one observation in a carefully designed sample. It asks hundreds of questions across major AI platforms (including ChatGPT, Claude, Gemini, Grok, Perplexity, CoPilot, and Google AI Overviews), stratified across stakeholder perspectives and question types, repeated until the signal separates from the noise. What arrives at the end is a score with clear confidence, a picture of where narrative is stable, where it fractures, where competing narratives are preferred, where old stories persist, and the part no platform can supply: a consultant’s judgement about what it means for this organisation, in this sector, at this moment, and what to do next.
GPS is the first public expression of Pagefield Digital Lab’s working philosophy, and the same discipline is being applied wherever communications work can now meet machine interpretation. But measuring AI reputation is the start, not the limit. The Lab’s remit is wider: taking the judgement that this firm runs on — how a crisis is read, how a stakeholder map is drawn, how advice earns the confidence it’s delivered with — and building systems that help that judgement accumulate rather than evaporate.
What we build next might look different. The way we build it will not.
