Google has built a team of AI detectives to catch AI spam, and it would rather not say how well they work
This week Google has hired a team of AI detectives to catch AI spam, and is being oddly coy about how good they are, while Brian Balfour explains how Chegg lost 90% of its value in nine months and why your growth model might be standing on the same trapdoor. Elsewhere, ChatGPT ads have gone from no ads at all to twenty outside referees in a year, your email platform still has no idea what anyone did after the click, and a novelist wonders who's left to watch if everyone is performing. Here are your five.
1. Google has built a team of AI detectives to catch AI spam, and it would rather not say how well they work
Google has published a research paper on SAFE, the Scaled Abuse Forensics Examiner, a system built to hunt AI-generated spam the way a human investigator would rather than the way a filter would. It runs four AI agents: an orchestrator that hands out the work and makes the final call, one reading content for synthetic giveaways, one watching behaviour for burst publishing and suspiciously synchronised uploads, and one mapping the shared infrastructure that ties a network together. Its party trick is catching content that breaks the spirit of a policy without technically breaking any single rule. The paper runs to three pages and declines to share its test results, which for a research paper is a bit like a recipe that just says "bake until done". It does claim early deployment has significantly cut investigation time, and it's the second AI-slop hunter Google has been linked to this year.
Worth reading the small print before anyone panics: the paper leans heavily towards video and channels, and the link to September's spam update is Search Engine Journal's reasonable inference rather than anything Google has confirmed. The shape of it still matters. SAFE isn't asking "does this read like a robot wrote it", it's asking "does this behave like a network", and pages published in bursts across sites that share the same plumbing look suspicious however polished the sentences are. Writing more convincing slop was never a long-term plan; now there's a detective agency to prove it. Search Engine Journal
2. Chegg lost 90% of its value in nine months, and Brian Balfour thinks plenty of growth models are standing on the same trapdoor
A decade after first writing it, Balfour has rewritten his Four Fits framework for the AI era. The premise: reaching $100m at speed needs four things lined up at once, product with market, product with channel, channel with business model, and model with market. Product-market fit, for all its fame, is only one of the four. What's new is that AI is shaking all four at the same time. Chegg is the cautionary tale, valued at $1.2bn in January 2024 and $150m by October, half a million subscribers lighter, because ChatGPT did the homework instantly and for free. Its first AI features shipped 26 months after ChatGPT did, a response time that once would have looked perfectly respectable. Elsewhere, Replit has 40 million users and just 0.44% of them want to pay, and Lovable's reported gross margin is 35% before counting free users, against a traditional SaaS norm north of 70%.
The piece doubles as an advert for his Reforge course, so read it with that in mind, but the section on channels is the bit that sticks. Platforms run a reliable cycle: find a moat, throw the gates open with cheap distribution, then close them again to make money. Balfour reckons most of the old channels are well into the closing phase, and he named ChatGPT as the next big opening. He wrote that in September last year. Five months later, ChatGPT started selling ads. Brian Balfour
3. In May, OpenAI promised outside referees for ChatGPT ads with no names and no dates. It has since named twenty.
Back in May, OpenAI opened its self-serve ChatGPT ads manager to US advertisers of every size and dropped its $50,000 minimum spend, while promising cost-per-action bidding and third-party measurement with, as Digiday pointed out, no partners named and no timeline attached. The ad itself was a small favicon and a line of text, OpenAI made every delivery decision, and advertisers got aggregated numbers rather than the conversations behind them. The promises have since landed, more or less. Conversion-optimised bidding arrived over the summer, though you still pay per click rather than per action. Then on 5 October OpenAI named twenty outside measurement partners, from AppsFlyer and Adjust to Triple Whale, with DoubleVerify and IAS piloting brand-suitability checks, and ads have spread well beyond the US in the meantime.
The headline result OpenAI chose to show off, WeightWatchers paying 15.3% less per acquisition than on its blended paid search, comes from OpenAI's own showcase, so the usual squint applies. Incrementality testing is still in its early days too, and that's the bit that would tell you whether those customers would have turned up anyway. Still, the pace is something. A year ago there were no ads in ChatGPT at all; it now has an auction, a pixel, product carousels and a small army paid to check the maths. That's a remarkably quick trip from open gate to toll booth. Digiday
4. Your email platform knows who clicked. It has no idea what they did next
Designmodo's guide to tracking email traffic in GA4 starts from a fair premise: people pour hours into subject lines, templates and segments, then judge the whole effort on a report that stops at the click. The fix is gloriously unglamorous. Every link gets UTM parameters, with source, medium and campaign required and an optional content tag for telling two buttons in the same email apart. Skip them and GA4 will cheerfully lump your newsletter traffic in with "direct", which is analytics for "no idea, mate". Add them, build a session segment where medium exactly matches "email", save it as an exploration, and you can see engagement, landing pages and revenue campaign by campaign.
Designmodo sells an email builder, and the 42% of B2B marketers it says rate email their most effective channel is a figure it cites from its own blog, so file that one under "enthusiastic". The core point holds regardless. The email platform is the right place for deliverability and open trends; GA4 is where you find out whether anyone stuck around. Opens have been a shaky measure ever since Apple started pre-loading emails in 2021, which leaves the click and everything after it as the honest part. An email that gets opened and goes nowhere is just a beautifully designed notification. Designmodo
5. She has 400 followers and narrates her holiday like a reality star, and she isn't the odd one out
Frederick Joseph, a novelist with a few hundred thousand followers of his own, opens with a school friend he calls Miranda. She has roughly four hundred followers, nearly all from their hometown, yet her resort holiday arrived as room tours, teasers for tomorrow's excursion and captions pitched at a fanbase rather than at the people who sat next to her in biology. Joseph traces the drift from AIM away messages and early Facebook, through friend counts quietly becoming follower counts and the like button becoming a scoreboard, to TikTok, which doesn't much care who you know at all. His metaphor for where we've landed is a slot machine: every post a lottery ticket, every dinner a potential reel.
It's an essay rather than a study, and it came out last autumn, but it lands uncomfortably close to home for anyone in marketing, given that treating people as an audience is more or less the job description. His sharpest point is about who actually benefits: not Miranda, not her old classmates, but the platforms banking every second of attention. He signs off with a question worth sitting with: "if everyone is performing, who is left to be the audience?" His darker answer is that it might just be the algorithm, which after a week of AI detectives and ad auctions reads less like a provocation and more like a status update. In Retrospect
Small business shout-out
This week's shout-out goes to Rapture & Wright, who hand screen-print fabrics and wallpapers at their print works in the Cotswolds, on a 20-metre table, to order, so nothing sits in a warehouse waiting to go out of fashion. Peter and Rebecca Thwaites started the business in 2004 in a leaky barn and have taken the green side seriously ever since: water-based inks, linen woven in northern England, rooftop solar that has out-generated the whole studio every year since 2022 (they insist the main energy source is actually biscuits), and a "Living Water" system of ponds that cleans the print waste and has turned into a wildlife habitat on the side. The trophy cabinet includes a Homes & Gardens sustainability award and a place on The Times' list of the best British-made design brands. Know a small business we should feature? Reply and let us know.
