Writes about media, tech, and AI business for people who actually run digital. Former agency lead. Skeptic of frameworks that read better than they perform.
A retailer selling in Connecticut may not set a shopper-specific price from personal data, and New York already makes you say so next to the price. Loyalty prices and posted discounts survive, but only when the terms are public and open to anyone who qualifies.
Google put numbers on first-party data plumbing on September 10: 14% more conversions with tag gateway, 11% from enhanced conversions, 26% higher incremental ROAS from Data Manager. Two of the four rest on two-week windows and one on a comparison of two half-years, so read them as signal you get back, not demand you create.
Both the EU and the UK say in their own operative texts that direct marketing may rest on legitimate interests. Neither says it about advertising built on tracking, and the line between them is a balancing test almost nobody writes down.
We read the robots.txt of twenty large US storefronts and asked each one for its homepage twice, once as a browser and once as an agent. The blocking is real, the enforcement sits somewhere else entirely, and the feed underneath never expires.
Two Reddit documents describe the same year. The audited one says the price of ads rose 40%, and the marketing one sells a 17% lower cost per action, measured in split tests that ran in 2025.
Incremental sales is the metric buyers ask for most, and about two-thirds of networks say they supply it. The industry’s own standard lists four methods behind that one word, and only one of them is an experiment.
The paid media audit is sold on percentages, and chased to their sources most end at another agency blog. Here are six failures you can prove inside your own account in an afternoon, and the arithmetic that separates a finding from an assertion.
Roundups frame agentic commerce as a standards war between OpenAI and Google. Read both specifications and OpenAI’s retreat from Instant Checkout, and the two rails are converging on one journey, asking merchants for the same feed at a cadence the spec never demanded.
Two migration dates are circulating with the wrong campaign types attached. Google’s 7% and the independent 16% come from different verticals, baselines, and years, and the question worth testing is whether the added volume was ever new.
Claude’s newest models weave an invisible mark into their text, with no documented opt-out. In Anthropic’s explanation the mark measures one thing, how many words the model chose: faint on a proofread draft, loud on prose it wrote end to end.
The 31 markets that went live on August 24 are exactly the EEA plus Switzerland, the territory OpenAI carved out of personalization. Polish buyers get a contextual product with a 65 PLN daily floor and a CPC figure that is a suggested bid, not a market rate.
Teams running several European markets argue about splitting campaigns per country, and settle it with a number somebody half-remembers. Google publishes it as the sample you need to read a result, not the volume the algorithm needs, and that changes the decision.
Thirty conversions a month. Fifty events a week. Every consolidation argument in the industry rests on those numbers, and neither platform states them as structural requirements. Google says something quieter and more useful, and once you read what it actually says, the real cost of splitting a campaign turns out to be arithmetic you can do yourself.
Every team now has models writing, classifying, and summarizing at volume, and almost none can say whether the output got better or worse last month. The fix is an eval: a frozen set of real inputs, a rubric that discriminates, and a threshold. The research on how to build one is unusually clear, and unusually uncomfortable about the shortcut everyone takes.
The slide says AI traffic is up and to the right. Underneath it, some assistants send a referrer and some send nothing, the biggest AI surface of all arrives labeled as ordinary organic search, and the crawler that decides whether you get cited at all is a different bot from the one everybody blocked. Here is what is actually measurable, and what to stop claiming.
Two things break creative testing at once: the platform decides who sees what, and almost nobody runs the numbers on how much data a real answer needs. The arithmetic is not close. Detecting a 20% difference takes roughly 400 conversions per variant, and most tests are called at a tenth of that.
Your sales team knows which leads were worthless. Your ad account does not, because nothing ever told it. Closing that loop is not a reporting project, it is a change to the optimization target, and the hard parts are a seven-day window, a hashing spec the two platforms disagree on, and the double-counting trap nobody warns you about.
The number on your dashboard divides all the revenue by all the spend, which describes money you already committed. The decision in front of you is about the next increment, and that return is always lower than the average and falling. Here is how to find where your curve bends, and what to do at the point the two numbers disagree.
In Shopping and Performance Max you cannot pick keywords, audiences, or placements in any way that sticks. What decides which auctions you enter is a text file: the title, the attributes, and the labels you put on each product. Here is how the matching actually works, and the one Google rule about AI-written titles that almost nobody has implemented.
Google and Meta will optimize toward whatever a conversion is worth, and they take your word for it. Most accounts send the order total, which is not profit, then set a target on top of a number the business never earned. Here is what to send instead, and the refund pipeline that keeps it honest.
The whole path, in order: the Google Cloud project, the credential, the connection, and then the part that actually pays for itself. One scheduled pull becomes a dashboard, a cohort grid the interface cannot draw, a written weekly summary, and a Slack alert that stays quiet until something real moves.
From August 2, three separate regimes (the EU AI Act, New York’s synthetic-performer law, and Amazon’s marketplace policy) start forcing you to disclose AI-generated creative. The panic that the label tanks conversion is repeated everywhere and proven nowhere. Here is what each rule actually covers, and how to decide where a human face still earns its cost.
You can run Google Ads from a terminal now: read the account, draft the change, spend the money, all in plain English. This is the build guide. Which of the four paths to use (scripts, the official read-only Model Context Protocol (MCP) server, a community write server, or the raw v25 API), what v25 quietly broke, and how to wire a write path that fails safe instead of fast.
The June 2026 spam update set off a wave of “brands lost their own results” panic. No primary data supports it. What actually got demoted was the coupon and affiliate sludge on your brand-plus-modifier queries, and branded search is the one segment AI Overviews are making stronger.
