In partnership with
Good morning. Here's what's worth acting on today.
BREAKING STORY
ChatGPT ads now run in 89 countries
OpenAI has added 26 countries to ChatGPT ads, bringing the total to 89, including 14 more Latin American markets such as Argentina, Chile, Colombia and Peru, plus new markets across Europe, the Middle East, Africa and Asia. Ads appear only on the Free and Go plans; Plus, Pro and Enterprise stay ad-free. OpenAI says more than 100,000 advertisers have run ads on ChatGPT since launch, and advertisers can buy through technology partners or directly in OpenAI's Ads Manager. Reuters ties the push to building revenue ahead of a potential IPO. The $1bn annualised run rate OpenAI cites again was first disclosed in August.
My take: ChatGPT is now a self-serve ad channel with real reach. If you sell a course, template or SaaS tool, run a small test budget now while the auction is still thin, and benchmark your cost per customer against your Meta CAC. To check those sales wouldn't have happened anyway, measure the test with a holdout (section 03). Ads only reach Free and Go users, so your organic citation work still matters for the paying users who never see them.
MARKET TRENDS
Employers want people who build with AI
Workday's Global Workforce Report, drawn from more than 550 employers using Workday Recruiting, found demand for hands-on "build with AI" skills climbed 51% between September 2025 and July 2026. Demand for basic prompting skills peaked in January 2026 and then fell 25%. The AI Exchange reads "building" as writing down a process an AI can follow, which needs no code. The data covers only Workday's own customers.
How to make money from it: Prompt packs are losing value, and buyers will pay for proof that a workflow runs. Package one role-specific playbook that turns a process into an agent, such as a month-end close kit for bookkeepers with the SOP, a skill file and an eval checklist. Or run a cohort where every student ships one working automation by the final week, and sell on that outcome.
Find what’s slowing your GTM.
Most GTM problems aren’t one big failure. They’re one or two weak spots quietly dragging down the rest.
The free GTM Audit scores your go-to-market across four dimensions, then sequences a 90-day roadmap for your stage.
Drop your URL. Get your score in about 10 minutes.
GROWTH HACK / DIY
Run your own holdout test
The idea: Attribution tells you who touched a sale. Incrementality tells you whether the sale would have happened anyway. To find out, you withhold a campaign from a random slice of your audience and compare the two groups. Google sells this as Conversion Lift, which lists Search and Performance Max among supported campaigns but needs a $5,000 USD minimum campaign budget, at least 1,000 observed conversions and, per Google's page, a call to your account rep. You can run the same test on your email list for free.
Do it today:
Randomise your list. Export it to a spreadsheet, add
=RAND()in one column and=IF(B2<0.2,"holdout","send")in the next, then paste the random column as values so it stops reshuffling.Send your next promo to the "send" group only. Keep the holdout out of every other campaign for a fixed window (say 7–14 days), and don't stop early.
Compare purchase rates. Lift = send rate minus holdout rate. Multiply by the number sent to get incremental buyers, then divide by the sales your email tool attributed. That's your incrementality factor. Made-up example: 2.0% vs 1.2% across 16,000 sent gives 128 incremental buyers against 200 attributed, a factor of 0.64.
What you'll need: A list big enough to see a difference. Evan Miller's rule of thumb (n = 16 × σ² ÷ δ² per group) means detecting a 1-point lift on a 2% base takes about 3,100 people in the holdout. Smaller list? Hold out more, or pool several sends.
AGENT LAB
Let the change decide how many humans review it
The pattern: Anthropic's Claude Code team has machines read every pull request first. A classifier then decides how many people must approve it: none for low risk once checks are green, one for medium, two for high. Any CI workflow file always needs a person. The team reports that merged PRs more than doubled from June to September while the revert share fell by more than 40%. Those numbers are self-reported with no control group, and the period overlaps new Claude model releases, so some of the gain may come from the models.
How it works: PR → CI, AI bug review and a security pass → a classifier sets the tier → merge queue and fast rollback. The AI review loop stops after three rounds and asks the author to rethink. After that change, the team says the share of PRs passing after one round rose from 65% to 78%, and the share needing five or more rounds fell from 8.9% to 1.5%.
Try it: Write a short rule script: workflow, auth, billing or migration files score high; over 400 lines changed, medium; docs or tests under 100 lines, low. Run it on your last 30 merged PRs. Every one that caused trouble must score medium or high. Then label each tier no-look, skim or read, and cap the AI reviewer at three rounds.
TRENDING AI RESEARCH
RAG tricks don't stack, so test the whole pipeline on your own docs
What it says: IBM researchers Ido Levy, Segev Shlomov and Leshem Choshen ran 450 RAG configurations across eight benchmarks and corpora of 5K–200K documents. Three retrieval techniques (cross-encoder reranking, turning off MMR, document-awareness prefixes) that should have added up to about +47% delivered +17.2% together. Across 17 combinations, a median 41% of the expected gain went missing. Retrieval quality and answer faithfulness were essentially uncorrelated (r = -0.01), and the best setup changed with size: a simpler stack won at 5K–15K documents, a heavier one from 25K.
Why it matters: If you run a docs bot or "chat with my course" product, adding up benchmark gains overestimates what you will get, and a bot that finds the right page can still make things up.
How you can use it: Write 40–60 real questions from your own docs, each with the source that answers it. Score retrieval (right source in the top five?) and faithfulness (every claim backed by the retrieved text?) separately. Add one technique at a time, then test the full stack against the sum. Re-run when your corpus doubles. The authors' open-source harness installs with pip install lasagna-rag.
TOOL REVIEW
Radaric: an AI feed reader for curators
What it does: Radaric lets you write your editorial rules in plain words, then scores every post from RSS, Telegram (including private channels), YouTube transcripts, Substack, Bluesky, Mastodon, GitHub, arXiv and Product Hunt against them, with a reason for each decision. Reject a post with one line of explanation and, the site says, it learns. Posts about the same event are merged into one story with every source linked.
Price (as of 24 October 2026): Lite is $9/month or $90/year. Pro is $59/month or $499/year and adds X/Twitter. There's no free tier, but the maker says sign-ups through the Product Hunt link get 7 days of Lite free with no card. The terms say the 30-day money-back guarantee covers annual plans.
Who it's for: Newsletter writers and anyone who triages hundreds of sources a week.
Verdict: Worth a 7-day trial if you run a curated newsletter; skip Pro unless X is central to your sourcing.
What Got Funded Yesterday
Four of these rounds were reported on 23 Oct and one (CubiFab) on 22 Oct.
Higgsfield (higgsfield.ai) · AI video and creator tools · US. Co-founded by Alex Mashrabov, Yerzat Dulat and Mahi De Silva, Higgsfield sells an AI suite for making professional video, images and audio. It says it passed $1bn in annualised revenue 18 months after launch. It added an undisclosed Series B extension from new investors TCV and NVIDIA; DST Global led the original Series B in August.
Intuitivo (intuitivo.pt) · edtech · Portugal. Founded by João Guimarães, Duarte Oliveira and Diogo Silva, Intuitivo is a digital assessment platform where teachers create, schedule and grade tests, from primary school to university. It says the product was born from a teacher's request and is still built with teachers. It raised a €1.35m seed round led by Caixa Capital and the Impact Innovation Fund, with Techstars, COREangels and Angels Way.
CubiFab (no website found) · maker hardware · China. (22 Oct) Led by chief executive Wang Wentao, the Shenzhen startup makes the C1, a desktop CNC workstation for machining metal. It says the C1 simplifies work that usually needs separate design, machining and setup tools; its team includes ex-Apple, DJI and Bambu Lab staff. It raised tens of millions of yuan in angel funding from Nanshan Strategic Emerging Industries Investment and Longhua Capital.
Swaay (swaay.io) · packaging tech · Ireland. Founded by chief executive Ciarán Gorman with Sébastien Baril, Mark O'Donnell and Jean-François Gaudreault, the Mullingar company prints artwork straight onto drinks cans, replacing plastic labels and shrink sleeves. It says it pairs short-run digital can printing with fillers and distributors across North America. It raised €2m led by Dublin's BVP, with members of the BVP Investor Club.
Phononic Vibes (phononic-vibes.com) · noise-control materials · Italy. Founded by chief executive Luca D'Alessandro, Giovanni Capellari and Stefano Caverni, the Milan company makes noise and vibration-control products for rail, energy and mobility. It says its metamaterial designs weigh up to 30% less than conventional products for the same acoustic performance. It secured a €25m financing round led by Jolt Capital, with Eureka! Venture, CDP Venture Capital and Simest.
QUICK HITS
OpenAI's Decisions API is now generally available: up to 400 classification questions per request, billed on input only at $0.10 per million tokens ($0.01 cached). (22 Oct)
Anthropic's Agent Skills and OpenAI's Open Responses have joined the Linux Foundation's Agentic AI Foundation, which now hosts 10 projects. (22 Oct)
PHP2Shell: two critical CVEs (CVSS 9.2) in PHP itself allow unauthenticated takeover of WordPress, Drupal and Joomla sites; the NSA advises patching PHP or taking unpatched servers offline. (21 Oct)
That's today. Hit reply and tell me what you're building. I read every email.
Dev