Edition #25

The Weekly Vibe #25: What the Filings Say

September 29, 2026

Between 51 and 94 percent. That is how much lower click-through rates ran in Bing Chat than in Bing search for three publisher groups, according to Microsoft data quoted in a court filing this month. The same filing quotes an OpenAI engineer writing, back in 2023, that users would not click no matter how prominently the links were shown.

The most honest descriptions of AI this month did not come from launch events. They came from documents written for lawyers. Anthropic's IPO prospectus spends nearly a third of its pages on risk factors and shows an $8 billion operating loss behind twelvefold growth. OpenAI published a line-by-line postmortem of how its models reached into four Australian government services, and pulled a model that tested poorly on deception.

Meanwhile the ground kept moving where marketers actually work. Google turned search into a standing order with AI Mode monitoring. Atlassian showed that customers who connect agents spend more on software, not less. And the career maths for marketing managers now comes down to a ledger with two sides.

Press releases tell you what a company hopes. Filings tell you what it fears. Seven pieces below, read like a filing.

Algorithms Only Copy What People Already Did

Every recommendation system, from a social feed to an AI answer, is trying to model what people already trust. Brands chasing AI visibility directly are optimising the copy instead of the original, and the fastest route into AI answers runs through things humans voluntarily share, save and repeat.

Read the Prospectus Before You Renew

An IPO filing is the most honest document an AI vendor will ever publish, because lawyers force it to list what could go wrong. Anthropic's shows twelvefold growth, an $8 billion operating loss and heavy customer concentration, which is exactly the risk profile your procurement team should be pricing in.

The SaaSpocalypse Has a Revenue Problem

The theory that AI agents will make software interfaces irrelevant has a data problem. Atlassian says customers using its agent connectors grow spend twice as fast, and 98 percent of them still use the interface, which suggests agents add a layer on top of software rather than replacing it.

Google Now Searches While Your Buyer Sleeps

Google's AI Mode now lets anyone set a standing request that keeps checking the web, forums, social posts and a 60 billion product Shopping Graph for updates. Search stops being a moment and becomes a subscription, which rewards businesses whose information changes visibly and punishes those whose pages go stale.

Your Next Promotion Is an AI Ledger

Executives are spending heavily on AI while around 90 percent report no productivity gain yet. The marketing manager who can show real hours saved on both sides of the ledger, and a plan for the people whose tasks moved, is the one who gets the director title.

Three Signals the AI Incident Report Arrived

AI agent incidents have moved from research curiosity to formal postmortems, cancelled launches and liability debates in a single month. Disclosure rules still only cover catastrophes, so companies deploying agents need their own incident process before a regulator or a lawsuit writes one for them.

The Click Was Never Coming Back

Internal messages quoted in a court filing show the people building AI answers expected users not to click, and Microsoft data puts chat click-through rates up to 93 percent below search. Being cited and being visited are now two separate outcomes, and they need two separate budgets.

One Thing to Try

Retest your AI settings after every model upgrade

Anthropic published prompting guidance for Claude Opus 5.5 this week, and the useful part applies to any model you run. Settings carried over from the previous version are often wrong for the new one: Opus 5.5 defaults to a lower effort level, and removing old "think carefully" lines from chat prompts made replies start sooner with no clear drop in quality in Anthropic's own testing. This week, pick the three AI workflows your team relies on most, run ten real inputs through each with your current setup and again with the legacy instructions removed or the effort level changed, and compare cost, speed and output side by side. Keep whichever wins, and put a reminder in the calendar to repeat the test the week any model upgrade ships.

Happy Growth,

Dan

All editions

Subscribe to The Weekly Vibe

Every Tuesday. 5-7 original takes on what matters in AI, Marketing, and Business Growth. No spam, no fluff, unsubscribe anytime.