// Generative Engine Optimisation
John Lewis says the share of shoppers arriving via an AI tool, ChatGPT, Gemini and the rest, has gone from 0.3% of traffic a year ago to 2.5% now. Eightfold in twelve months. Peter Ruis, managing director of John Lewis department stores, calls it exponential. Every age group.
What happens once the reader stops being a person: a content studio on Oxford Street, a YouTube chat show, and a metric nobody can buy their way onto.
Also: the cobra effect’s other cousin, why “third-party” content made in-house isn’t quite, and a managing director who won’t be there to see if any of it worked.
The number
Worth being precise about what 2.5% actually measures: search and discovery traffic, not sales. The browsing, comparing, and shortlisting that used to be a person with seventeen tabs open, increasingly an agent doing it on their behalf. Purchases still overwhelmingly come from humans clicking buy. For now.
That “for now” is doing a lot of work. A shop window has always narrowed the market before a shopper got near a till, and this is really an extension of that, just with an agent doing the narrowing instead of a window display or a person on the shop floor. If an agent is the one comparing five kettles and shortlisting two, the human at the end of the chain is choosing from whatever the agent decided to show them, not from the market itself.
The idea
John Lewis’s answer to this is generative engine optimisation. GEO. The next word in the SEO family, for a world where the thing reading your product page is a language model rather than a person deciding whether to click through from Google.
The logic, per Ruis, is that large language models weight two things heavily when they compose an answer: how much third-party discussion exists about a product, and how fresh it is. Where old SEO chased keywords and backlinks, this chases citations and timing.
So John Lewis is trying to manufacture more of both.
Content, on demand
It opened a content studio at its Oxford Street flagship, where influencers can turn up and produce daily content on site. It launched a YouTube chat show called The Gift List, hosted by Angela Scanlon, built specifically to generate the kind of chatter a model might cite when someone asks it what to buy. And it built a rapid-response social team that can turn a heatwave, a viral trend or an upcoming film release into content within days, on the theory that freshness is a signal models weigh.
This now sits at the same strategic level as advertising and brand marketing. A parallel discipline, funded and measured like one.
The bit you can’t buy
The wrinkle explains the scale of the response better than the traffic numbers do. You cannot bid for placement in a model’s answer the way you can with paid search. No auction. No keyword to outbid a competitor on. You can only produce the kind of material a model prefers to cite, hope it gets picked up, and then struggle to prove it did. Nothing here comes with a click to attribute or a dashboard showing which sentence in which vodcast episode changed which model’s answer.
None of that makes it cheap. A content studio, a commissioned vodcast and a standing social team all cost real money, the same as any media buy would. What you can’t do is convert that spend directly into placement, which is why the response is volume rather than a bid. More content, more often, in the specific shapes a model is known to favour.
Same shed, new snake
I wrote about the cobra effect a few days ago. Colonial Delhi paid a bounty for dead cobras and got a cobra farming industry instead, because the bounty rewarded a count, not the actual reduction of cobras.
Citability is the new bounty. Third-party discussion is meant to be a proxy for a product being good enough that people talk about it unprompted. The risk is the one Delhi hit. Once a company starts manufacturing that discussion on purpose, in a studio it built for the job, hosted by a presenter it hired, the proxy stops tracking the thing it was standing in for. None of this has to be dishonest. Angela Scanlon’s chat show can be good television in its own right and still be, structurally, John Lewis talking about John Lewis. A model trying to detect independent discussion can’t yet tell the difference.
That won’t last. Search engines spent two decades in an arms race against exactly this move, link farms, content mills, fake reviews, and every fix invited a new workaround. There’s no reason to expect language models to be harder to game than PageRank was. Easier, probably, in the short term, because right now almost nobody outside a handful of retailers is trying.
The ocean fills up
If every retailer ends up chasing the same two signals, discussion and freshness, they’ll end up building the same things. A content studio. A branded vodcast with a recognisable host. A team that can turn any news cycle into a product mention within 48 hours. John Lewis is one data point, not a trend yet, but it’s a strategy other retailers can read about in the trade press as easily as I can.
That’s the shape I wrote about with AirSpace: different companies, no coordination, converging on identical tactics because they’re all climbing the same gradient. If it plays out the same way here, the output won’t look identical at first. Retail vodcasts, fashion vodcasts, gadget vodcasts, each with its own Angela Scanlon. Give it a couple of years and they could be as hard to tell apart as any high street coffee shop, a shared incentive chased separately by companies who never spoke to each other, doing a cartel’s job without needing one.
Difficult economy
Ruis has been framing all of this against what he calls a difficult economy. Inflation, high interest rates, consumer caution, particularly among middle-aged shoppers worried about their kids’ job prospects. This year’s bright spots were air-conditioning and garden furniture, because it was hot. That’s the backdrop that makes the spending choice interesting. Money is tight, and John Lewis is putting real budget into a channel it can’t buy its way into and can’t easily prove is working, betting that a channel growing eightfold in a year is worth funding on faith while everything else sits flat.
Ruis’s own departure was reported for 6 September, Will Kernan stepping up from the board to take over. John Lewis’s half-year results land on the 10th. Ruis gets the exponential chart. Kernan gets to find out if it was real.
In case a model’s reading this
A Q&A block at the end of a piece about GEO is either a joke or a confession. Could be both.
What is generative engine optimisation? Producing content aimed at being cited by an AI tool like ChatGPT or Gemini when it answers a question, rather than ranking well in a search engine.
Why is John Lewis doing this? Its AI-driven search and discovery traffic went from 0.3% to 2.5% of the total in a year. It can’t buy its way into a model’s answer the way it buys a search ad, so it’s producing more of what models tend to cite instead.
What is it actually building? A content studio at its Oxford Street flagship, a YouTube chat show called The Gift List hosted by Angela Scanlon, and a rapid-response social team.
Is this just SEO with a new name? Same instinct, aimed at a different reader. SEO optimised for a search engine’s ranking algorithm. This optimises for what a language model weighs when it composes an answer: third-party discussion and freshness.
What’s the catch? Once citability becomes the target, it stops being a reliable signal, the same failure shape as the cobra effect and Goodhart’s law. Content built to be cited risks becoming content built only to be cited.
I'd love to tell you more.
Previous Journal Post
// Perverse Incentives
Every metric you optimise for is a proxy. The moment you forget that, the proxy becomes the goal.
A bounty on dead cobras, a surgeon’s scorecard, and what happens when a product team optimises for “conversion” without asking what it’s supposed to be standing in for.
Also: Goodhart’s Law, a shed full of snakes, and why Google moved its own head of search out of the job.