The Plane from Sydney: When the Engineers Confirmed Everything
Thinking in Motion: Article 5 of 7 Category: Thinking in Motion (/thinking-in-motion/) Site: jasonbarnard.com
Beer Mats and a Barman’s Pen: What Gary Illyes Told Me About Confidence Scores
The night before the flight, in a hotel bar in Sydney, a barman had a pen but no paper, so I split beer mats in half and wrote on those, which is not the most professional way to document a breakthrough but is what you do when a Google engineer is explaining something foundational and you cannot let the words evaporate into the noise of the room.
I was at a table with Gary Illyes from Google and Brent D. Payne, who runs one of the most respected SEO consultancies in the world, and Gary described something that would change how I understood everything: Google’s annotation system works as a bidding mechanism, meaning that when Google encounters information about an entity (a person, a company, a brand), it does not simply store that information as fact but annotates it with a confidence score, and each piece of information from each source carries a weight, and these weights compete with each other, and the system resolves competing claims by evaluating which sources are most trustworthy, most consistent, and most corroborated.
A confidence score as a bid, which means the information does not just exist in the system but competes, and this is how I finally understood why contradictory signals are so catastrophically destructive: if one source says you are a digital marketing consultant and another says you are a cartoon blue dog, those claims are bidding against each other, and the system has to pick, and it picks based on the strength of the evidence and not on what you would prefer.
Brent added something I still quote: “Better to be a straight C student than three As and an F,” meaning that if your website is brilliant but your social profiles contradict it and your third-party references are inconsistent, the algorithm does not average the scores in the forgiving way a school report might, because the F destroys the As, since in algorithmic evaluation signals are not additive but multiplicative, and a zero in any factor zeros the product.
How Nagu Rangan at Microsoft Independently Described the Framework I Had Already Built
That Sydney trip was part of a year of conversations with engineers, and the most significant happened with Nagu Rangan, a Principal Programme Manager at Microsoft, whom I interviewed at SMX London in 2019 and who was the first Microsoft engineer to speak on the record about how Bing processes entity information.
What Nagu described was, functionally, the framework I had been building for years but had not yet named: he talked about Relevancy (does the algorithm understand what this entity is?), Quality (can the algorithm trust the information?), and Context (is the information appropriate for the user’s query?), and as I listened I was thinking that is Understandability, Credibility, and Deliverability, different words but the same structure and the same underlying logic.
I had built the UCD Framework from the outside, observing algorithm behaviour and inferring the underlying architecture, and Nagu described it from the inside, explaining the system’s actual design, and they matched, which is not confirmation bias but independent verification, because when an outsider’s empirical observations converge with an insider’s system knowledge you know the model is real.
Once Nagu had spoken on the record and the world had not ended, others at Microsoft followed, and I owe him for sticking his neck out first.
Nathan Chalmers Told Me That Bing’s Algorithm Is Literally Called “Darwin”
But the convergence that still astonishes me came from Nathan Chalmers, who led the Whole Page Algorithm team at Bing, and who told me almost in passing that one of Bing’s algorithms is literally called “Darwin.”
I had coined the phrase “Darwinism in Search” as a metaphor years earlier to describe how algorithms apply selection pressure to information, meaning the strongest, most consistent, most well-adapted signals survive while weak, contradictory, poorly corroborated signals are suppressed, like natural selection applied to data.
I thought it was a metaphor, and Bing thought it was an algorithm name, and neither of us knew about the other, and the convergence was so complete that when Nathan said the word I had to ask him to repeat it because I was not sure I had heard correctly.
Flying Home from Sydney with Convergent Proof from Google, Microsoft, and Bing
The flight home from Sydney is the longest haul you can do from Australia to Europe, and I had the worst jetlag of the trip (the wrong way round, landing before you took off), and I could not sleep, which meant I spent the hours doing what I always do in transit when something has been taken away: eyes closed, thinking, eyes open, writing desperately on whatever was available, eyes closed again.
What I carried onto that plane was not a single insight but a convergent evidence base from independent sources. Gary’s annotation bidding from Google, Brent’s multiplicative destruction, Nagu’s proto-UCD from Microsoft, Nathan’s Darwin from Bing. Each conversation had been a single piece, and the picture they formed together was something I could now say with confidence was not a theory I had invented but a description of how these systems actually work, confirmed from both sides of the wall by the people who built them and by someone who had spent years observing them from the outside.
The interviews continued through April 2020 in a dedicated Bing series: Fabrice Canel on crawling and indexing, Frédéric Dubut on ranking and relevancy, Nathan Chalmers on the whole page algorithm, and each one added depth and precision to a framework that was no longer speculative but was backed by convergent proof from the two largest search engines on earth.
I lost the beer mats from Sydney (I have looked, they are gone), but the principle Gary articulated (confidence as a bid) is now baked into everything we build, and the convergence between “Darwinism in Search” and an algorithm literally called Darwin is the kind of thing that only happens when you have accidentally described reality before anyone told you what reality looked like.
Next in the series: The Flight from Seattle. An air steward confiscated my Mac over the Atlantic because of a battery recall, leaving me with ten hours, no screen, and everything I had learned from a year of investigation crystallising into universal theory.
