Why I’m betting on AI Engine Optimisation

The companion post was the announcement. This one is the argument behind it - why, of all the things I could have spent the next chapter of my career on, I picked this one.

I’ve been writing a lot lately about operating points: the idea that there’s rarely a globally correct answer in a retrieval system, only the right answer for a particular workload, corpus, and constraint. The best embedding dimension depends on your latency budget. The best hybrid-search weighting depends on your query mix. The best quantisation depends on your hardware.

Here’s the operating-point shift that this whole series has been about, stated as plainly as I can:

For a growing number of businesses, the operating point that matters is no longer query latency, or index size, or even where they rank on a results page. It’s whether the model mentions them at all.

That’s a genuinely new question, and I think it’s a big one. Let me say why I find it worth a career bet rather than just a blog series.

The problem is structurally interesting

I’ve worked on a lot of search problems. Most of them, honestly, are variations on a theme: here is a corpus, here is a query, return the best results, make it fast. AI visibility is a different shape of problem, in three ways that I keep turning over.

It’s corpus-side, not query-side. Classical SEO and classical search both fundamentally work the query: you optimise a page, or a ranker, against the queries you expect. AI visibility flips it. You’re trying to influence whether your content survives a retrieval-and-generation pipeline you don’t own, for questions you’ll never see the exact wording of. The lever isn’t the query. It’s the corpus, and your content’s place in it.

It’s multi-provider by nature. For twenty years, "search visibility" effectively meant "Google". One engine, one set of rules, one game. AI visibility is fragmented across ChatGPT, Claude, Gemini, Perplexity, Google’s own AI surface, and whatever launches next quarter - each with a different index, different crawl behaviour, different citation logic, and a different rate of change. There is no single algorithm to reverse-engineer. There’s a moving set of them. That’s much harder, and much more interesting.

It’s measurement-first. This is the one I keep coming back to. Most optimisation disciplines start with tactics and bolt on measurement later. AI visibility can’t - because, as I argued in optimising for AI search, you cannot see the result. There’s no rank tracker for "are we in the answer". So the discipline has to begin with instrumentation: a reliable, multi-provider, noise-aware way to measure whether you’re showing up, before you can sensibly do anything about it. A field where measurement is the hard part and the first part is a field where engineering - not marketing - is the core competency.

Put those three together and you have a problem that is corpus-shaped, fragmented, and measurement-first. That is not a marketing problem wearing an engineering hat. It’s an engineering problem with a commercial outcome bolted directly onto it.

What fifteen years transfers - and what doesn’t

I want to be honest about this, because it’s easy to claim that everything you’ve ever done was secretly preparation for the thing you’re doing now. It wasn’t. But a real amount of it transfers.

What transfers: the retrieval pipeline is the retrieval pipeline. Knowing how candidate generation, reranking, and filtering actually behave under load; knowing why late interaction trades storage for fidelity; knowing how an embedding space distorts and where the operating points sit - all of that is directly load-bearing when the question is "why did this document get retrieved and that one didn’t". And the discipline of evaluation - building trustworthy measurement of a fuzzy, non-deterministic system - is something search engineers have been forced to be good at for years. That instinct is exactly what a measurement-first field needs.

What doesn’t transfer: the assumption of a single engine. The crawler-signal craft of classical SEO. The comfort of a deterministic results page you can screenshot and diff. A lot of hard-won intuition about Google specifically is now just one data point among many. I’ll have to hold the old playbook loosely.

That balance - most of the engineering transfers, most of the assumptions don’t - is, for me, the sweet spot. Familiar enough that I’m not starting from zero. Strange enough that it’s worth getting up for.

Why now

Timing is the part of any bet that’s easiest to get wrong, so let me be concrete about why I think this moment is the right one.

The consumer shift is already past its tipping point - that was pieces one and two of this series, and it’s not really arguable any more. The commercial consequences are landing right now - that was piece five, the traffic that went somewhere else. But the tooling and discipline to respond are still embryonic. Most businesses can see the problem in their dashboards and have no vocabulary for it, no measurement for it, and no credible playbook.

That gap - problem fully arrived, response barely begun - is exactly the window where it’s worth building. Too early and you’re evangelising a problem nobody feels yet. Too late and the category is settled. Right now, in 2026, the problem is acute and the answers are wide open. That doesn’t come around often.

Why this company

It’s one thing to believe a category exists. It’s another to believe the company you’ve joined is the right vehicle for it. The thing that closed the second question for me was, frankly, the traction.

Searchable passed $2M ARR within a couple of months of launching. That isn’t a category-creation curve - it’s a category-is-already-here curve. Customers aren’t buying because they’ve been convinced the problem will bite; they’re buying because it’s biting now. That matches what I was seeing in the data while writing this series, and it was the thing that pushed "interesting category" into "real company".

The funding picture points the same way. A $14M seed round, led by Headline, is a lot of conviction at this stage - and the kind of conviction that buys the time to build the measurement layer properly rather than racing to ship the obvious thing first. In a measurement-first field, that runway is the difference between getting it right and getting it shipped.

Put the two signals together - revenue that arrived faster than the category narrative, and backers willing to fund the deeper build - and what you have is a company whose constraints actually line up with the shape of the problem. That, more than any single conversation, is what made the decision feel less like a leap and more like the obvious next thing.

The actual reason

I’ll close on something less strategic and more honest.

I have spent my entire career on one question, wearing different clothes each time: how do people find things, and how do we build the systems that help them? Elasticsearch clusters, video discovery, recommendation pipelines - same question underneath.

The thing I find quietly thrilling about this next chapter is that the question has finally turned around. For fifteen years I worked on helping people find things. The problem now is helping things be found - in a world where the finding is done by a model, on someone’s behalf, inside a box. It’s the same question viewed from the other side of the glass, and almost everything I know still applies, and yet it’s new.

That’s the bet. I think it’s a good one. Thanks for reading the series - and let’s see.