Prompts are a measurement sample. Buyer questions are the strategy.

A prompt library assembled at a desk measures an imaginary market very precisely. The questions that decide whether somebody buys from you already exist somewhere in the organisation.

The obvious way for AI search to recreate the worst of SEO is already visible. Replace keywords with prompts, ranking trackers with visibility tools, build a list of a hundred questions, check whether the brand appears, optimise the pages, run the report again next month.

Some of that is genuinely useful work. The mistake sits underneath it, in treating the prompt as the strategic unit — and I think correcting that changes how the whole discipline is approached rather than adjusting it at the edges.

Traditional search offered marketers something wonderfully concrete. A keyword had a volume, a page had a rank, a competitor sat above or below you. The numbers were never as solid as they looked, but there was a clear object to optimise, and the appeal of that clarity is not something to be sniffy about.

AI search doesn’t offer it, and the reason isn’t that the tools are immature.

The query is not a stable object

Watch how somebody actually researches a purchase. They ask what the best accounting software is for a fifty-person professional services firm. Then which of those integrate with Salesforce. Then for a comparison of two specific products for a UK finance team. Then what the implementation risks are. Then — the question that actually decides it — which one they’d choose if reporting across multiple entities mattered more than anything else.

The query changes as the buyer learns. Context accumulates across the conversation, each constraint alters the answer, and the assistant may decompose any of those questions into searches the buyer never sees.

So a list of a hundred invented prompts creates precision where none exists. You can track them, and the numbers will move, and the movement will look like information. What you cannot do is mistake the tracking system for the market.

The questions already exist

The better starting point isn’t which prompts to rank for. It’s which questions determine whether somebody understands you, shortlists you or buys from you — and those questions are already being asked, in sales calls, customer interviews, support conversations, win/loss reviews, review sites, events and the things prospects keep challenging subject-matter experts about.

Most organisations already hold these occasions and treat them as raw material for assets. The webinar becomes a recording, the customer interview becomes a case study, the conference becomes six social clips. Treated instead as evidence-gathering, the same hour answers a different set of questions: what do customers keep worrying about, which requirement keeps changing the shortlist, what language do buyers use that never appears in the marketing, which claim do prospects keep challenging, and what do they need before they can defend the decision to somebody internally? That shift is worth a piece of its own.

Its application here is that those conversations are a far better source of test scenarios than somebody at a desk inventing a hundred prompts, because they carry provenance. The questions represent something observed rather than something imagined.

That changes what a bad result means. When an AI system gives a weak answer about you, the prompt-as-keyword model asks how to optimise for that prompt. With a sourced question, you can ask the more useful thing: why is this answer weak?

Sometimes the page is genuinely hard to retrieve. Sometimes the positioning is vague. Sometimes third-party sources describe you differently from your own site. And sometimes the answer is that the evidence doesn’t exist — which is the possibility I think GEO discussion consistently underrates.

The point where optimisation stops being the answer

Suppose buyers keep asking how a product performs during a difficult implementation. You test it, and competitors come back with case studies, reviews and detailed implementation material while your company appears as a generic product description.

You can restructure the page, add an FAQ, change the headings. That might help a little.

But it’s worth considering that nothing is being suppressed. Nobody has interviewed the implementation team. Customers have never been asked where projects became difficult. The expertise exists inside the business and has simply never been captured, which means there is nothing for a retrieval system to find because there is nothing there.

At that point the GEO problem has become a research brief, and the loop looks quite different from a monthly optimisation cycle: a buyer signal produces a question, the question gets tested, the test exposes a representation gap, the gap turns out to be an evidence gap, the evidence gets gathered, published and activated, and then tested again. AI visibility work stops sitting at the end of content production squeezing citations out of finished pages, and starts helping decide what the organisation needs to find out next.

I’ve reached the same place from the opposite direction before, arguing that the durable advantage in AI search won’t come from producing more machine-readable pages but from building evidence and a reputation worth retrieving. Arriving at it a second time, from measurement rather than from content strategy, is the main reason I trust it.

Measurement has the same problem

There’s a second reason prompts shouldn’t become the whole system, and it’s more awkward: most of the AI-mediated journey is invisible.

A click from ChatGPT to your site may show up as a referral. But somebody can spend twenty minutes researching you inside an assistant, remember the name, type it into the browser directly, mention it to a colleague, arrive via Google a week later, or call sales the next day. The AI interaction shaped the decision and left no trace in conventional attribution.

Which is why I’d run three kinds of measurement together rather than one. AI representation samples the systems that matter against your sourced questions: do you appear, how are you described, which sources are cited, where are competitors clearer. Observable behaviour covers referral traffic, landing pages and conversions where the data genuinely exists rather than where it can be modelled.

And the third is the one marketers have spent years undervaluing, which is to ask people. How did you research this? Did you use an AI assistant? What did you ask it? What came up? What did it say about us? What made you look further?

That isn’t attribution. It’s customer research that happens to arrive in the form of attribution — you’re discovering the language buyers use, the comparisons they make, and the public account of your organisation they encountered before anybody from your company spoke to them. Feed it back and the next round of questions is better sourced than the last.

The tools are not the problem

I don’t want this to become another argument that software is missing the point. Testing hundreds of questions across several systems manually is absurd at any real scale, and you need automation, baselines, trend data and competitor comparison to do this seriously. The platforms building it are doing necessary work.

The condition is only that the tool answers a question that came from somewhere. Without it, the old machine reassembles itself almost automatically: a dashboard produces a score, the score becomes a target, the target generates optimisation work, the work produces more content, and eventually the organisation is busy improving a number nobody has connected to a buying decision. We have done this before, and the risk of doing it again is the thing I find most plausible about the current moment.

Where this argument is weakest

Breadth has real value, and provenance is not free. A hundred invented prompts cover ground that a dozen well-sourced ones miss, particularly for a broad market or a category where the organisation’s existing conversations are unrepresentative — you only hear from people who already reached you. Sourcing questions from your own funnel systematically overweights the buyers you already win. The honest version is that you need both, and I’ve written this as though the sourced set is simply better.

Asking people what they searched is less reliable than it sounds. Self-reported research behaviour has been unreliable for as long as anyone has measured it. People compress a fortnight into “I looked it up”, misremember which tool they used, and increasingly cannot distinguish an AI assistant from an AI summary inside a search engine. It’s still worth doing, and it is still evidence rather than proof.

And this conclusion is extremely convenient for me. I’m arguing that the answer to a measurement problem is more upstream evidence work, which is exactly what the methodology I’m developing consists of. Somebody selling optimisation would reach a different conclusion from the same facts, and I can’t claim to have arrived at mine from a neutral position.

The shift in AI search isn’t that prompts have replaced keywords. It’s that the buyer’s research has become conversational, contextual and partly invisible — which makes understanding the buyer more important rather than less.

So before building a prompt library, build a question library. What are people trying to understand? What constraints change the answer? Where do you already have a strong answer, where are you bluffing, where does the knowledge sit inside the business without ever reaching the public, and where does nobody know the answer yet?

The prompts come afterwards, because prompts are a measurement sample. Buyer questions are the strategy.