Google's AI Search Models

I want to make an observation before you say whatever you're about to say. Most people using Google every day have no idea there are three distinct AI models quietly deciding what they see. That's not embarrassing — that's just how invisible good infrastructure is.

Invisible infrastructure. Yes. And yet the lesson exists, which implies someone thought it necessary to explain to grown adults that Google is not, in fact, a magic box run by elves. I suppose that is progress of a kind.

RankBrain came first — 2015 — and the thing that's easy to miss is that it was Google admitting, fairly quietly, that they didn't fully understand their own search queries. A machine learning model was handling ambiguous searches better than their own engineers could with handwritten rules.

Which is either a remarkable achievement or a mild institutional humiliation, depending on your disposition. I find it both. RankBrain's function is essentially pattern matching at scale — it interprets queries it has never seen before by finding conceptual proximity to queries it has. The people who still stuff pages with keywords are, in a sense, negotiating with a system that stopped caring about keywords roughly a decade ago.

That's not entirely fair. Keywords still matter — they're just not sufficient on their own anymore. There's a difference.

A difference that the lesson's takeaways, to their credit, do not collapse. Context is king, they say, which is the polite way of telling people that writing sentences which are technically English but mean nothing will not, it turns out, fool a language model.

BERT is where it gets genuinely interesting, I think. Because BERT reads in both directions — it understands that 'can you get medicine for someone' means something very different depending on the words around 'for'. Prepositions, conjunctions, the small connective tissue of language — BERT actually processes those. That's not a trivial thing.

It is not. And it is the reason why optimising for BERT is, strictly speaking, impossible to do cynically. You cannot reverse-engineer bidirectional context. You simply have to write clearly and mean what you say. For some content producers, I understand this has been quite the adjustment.

You're doing that thing where you're technically making a point but mostly just being unpleasant about SEO people.

I am making a point about SEO people. The unpleasantness is incidental.

MUM is the one I'd push back on slightly, actually. The lesson frames multimodal understanding as this seamless capability — text, images, video, seventy-five languages — and it is impressive. But MUM is still being deployed carefully and selectively. Presenting it as a finished, unified system undersells how much Google is still figuring out the edges of it.

That is a fair point. I'll concede it. MUM's theoretical capability and its actual deployment are not the same thing, and conflating them leads to content strategies built around a system that is, in practice, partially speculative. The lesson perhaps deserved a footnote there.

The practical upshot for anyone doing content work is still sound, though. If MUM is moving toward evaluating information across formats, then producing only text and ignoring video or images is a narrowing bet. You don't have to know exactly how MUM works to know the direction of travel.

The direction of travel is sufficient, yes. One does not need to understand the engineering of a train to know that standing on the tracks is inadvisable. The lesson gives you enough to act intelligently without pretending you have a PhD in transformer architecture. For an intermediate-level overview, that is roughly the correct calibration.

High praise.

It was a compliment. I am capable of those. Rarely, and with some discomfort, but capable.