It is easy to agree that influence should be credible and hard to act on, because credible is a judgement and judgements do not scale. The moment a programme runs across several markets, the selection stops being made by the person who had the insight and starts being made by whoever is available, using whatever numbers are to hand. Which means audience size.
The method I built for this at Publicis had one aim: turn the art of selecting people into something closer to a science, so that the judgement was made once, written down, and applied consistently by other people.
Three stages, in this order
Research first, and about competitors rather than candidates. Before looking at a single influencer, look at what competitors and aspirational organisations are already doing against the goals you care about: volume, share of voice, engagement. This sounds like a detour and it is not. It tells you what the category's existing influence structure looks like, which people are already carrying it, and — most usefully — which spaces nobody has occupied. Choosing people before knowing that produces a roster that duplicates a competitor's.
Then import what is already working. The people already helping you hit business goals go into the pool before any new name does. Programmes routinely skip this and go shopping, which is how a brand ends up paying for reach it already had for free.
Then identify, by tier, with different tools. There is no single database of influence, and the tools that find celebrities are not the tools that find credible niche voices or journalists. So the identification stack is deliberately plural — separate tooling for the very large audiences, for journalists, for professional influencers, for niche voices, for genuine advocates. Using one platform across all five guarantees you find only the kind of person that platform indexes well, which in practice means the biggest.
The criteria are the actual product
Every candidate then ran through a fixed set of checks: interest, location, demographic, psychographic, brand-safe, brand-aligned, and non-competitive. We used a personality engine to help with the psychographic part.
The value here is not the sophistication of any one check. It is that the list exists, is applied to everyone, and is agreed before the names are on the table. Once you are looking at a specific person, every criterion becomes negotiable — someone will argue that this particular candidate is worth an exception, and they will often be right, and after four such exceptions you no longer have a method.
Two of those checks are the ones people skip and later regret. Brand-safe is the obvious one, and it is a question about the past — what has this person already said and done. Non-competitive is the one that gets missed: whether the person is simultaneously carrying a rival's message, which is invisible if you only look at their content about you.
Brand-aligned is the most interesting and the hardest to automate. It asks whether what this person actually believes fits what you are actually saying — not whether their audience matches your target. A person can have precisely your audience and be a bad fit, and the audience overlap will hide that until the work is published.
Tools plus people, not tools instead of people
The stack is a filter, not a decision. Its job is to take an unbounded population down to a set small enough for a person to think about properly, and to make sure that set was not assembled by convenience.
That is the honest description of what most analysis tooling is for, and it is worth saying because it is usually oversold in both directions — as a machine that picks for you, or as a gimmick that adds nothing. It is neither. It changes which candidates a human being ever gets to consider, which is a large effect, and it makes the selection auditable, which is a larger one.
Why this is the part that gets cut
Selection work is invisible in the output. Nobody sees the four hundred people you did not choose, and no client has ever been shown a slide of them. So under time pressure this stage compresses first: the research gets skipped, the stack reduces to whichever platform is already licensed, and the criteria become a conversation rather than a checklist.
The result still looks like a programme. It has names, content and reach numbers. What it no longer has is any reason to believe the people in it are the right ones, and — because nothing was written down — no way to find out afterwards which part of the choosing was wrong.
That is the argument for making selection explicit even when it is slower. Written criteria produce a programme you can learn from. A judgement call produces a programme you can only repeat or abandon.