Ads grew 26%. Units grew 17%.

Happy Friday,

Amazon told a federal court it was defending a trusted shopping experience.

Amazon told its shareholders something more specific three weeks earlier.

Amazon's Q2: advertising revenue $19.8 billion, up from $15.7 billion. That's 26% growth, roughly 1.7 times faster than online stores. Paid units grew 17%.

Two more numbers, and I'm going to hedge these properly because everyone else won't. Amazon does not report advertising revenue per unit, and it does not report third-party unit counts at all. Both figures below are trade estimates built on derived numbers, not line items:

Ad revenue per paid unit: up roughly 8%.

Seller services revenue per third-party unit: up roughly 1%.

Treat the ratio as directional rather than precise. Directionally, it says Amazon's incremental money per unit is coming from selling attention rather than from selling the marketplace — and that is not a side business that grew well. That is the business.

Now hold that next to the lawsuit.

Amazon sued Perplexity because Perplexity's Comet browser lets a customer instruct software to buy something on Amazon — the customer's own account, the customer's own password, the customer's own explicit permission. Amazon's claim was unauthorized access to a protected computer under the federal anti-hacking statute. It won a preliminary injunction in March. Amazon said the injunction protected "a trusted shopping experience."

On August 4 the Ninth Circuit vacated it. Amazon.com Services, LLC v. Perplexity AI, No. 26-1444. The holding: it is the user who accesses Amazon's computers, with the help of the assistant.

Everyone covered that as a hacking case. It is a hacking case. It's also an advertising problem, and here's the mechanic.

One piece of bookkeeping first, because it matters and nobody else will bother with it. That $19.8 billion is a single consolidated line. Sponsored ads, display, video, DSP, retail media — all of it, in one number. Amazon does not break out the split, and live sports sold out this quarter across Thursday Night Football, the NBA, the WNBA and NASCAR, so some real and undisclosed share of that growth is television.

What Amazon did say, in Jassy's words on the call, is that Sponsored Products "continues to be our largest offering and a key driver of growth."

That's the part this is about. Not the whole nineteen-eight. The largest piece of it.

Sponsored Products is sold by the impression. The impression requires a surface. The surface is a search results page, viewed, by a person, with eyes, who scrolls past the paid rows on the way to the organic ones. Every dollar of the largest advertising business Amazon owns rests on the physical fact of a human being looking at a screen for a few seconds longer than they needed to.

An agent does not do that.

An agent doesn't scroll. It doesn't linger. It doesn't glance at position two on its way to position seven. It does not experience a sponsored badge as anything except a word. There is no meaningful sense in which an agent can be shown an ad, because being shown something is what eyes do.

So run it forward. Every purchase that routes through an agent is a purchase that generated a fraction of the ad inventory it would have generated as a human session. Not zero — models read text, and paid placement is text. But nothing like a person browsing a category for four minutes.

You cannot sell an impression to something that does not have an impression. That is the entire case.

Now line the two behaviors up, because this is where it gets funny.

In January, Amazon's own AI programs — Shop Direct and Buy for Me — were reading third-party retailers' websites, listing their products on Amazon, and routing orders to them. Without asking.

Retailers found out when the orders arrived.

One of them, a paper goods company, received dozens of orders for a stress ball. It does not sell a stress ball. The orders came from an address reading buyforme.amazon.

A founder at another brand said she'd been "forced to be dropshippers on a platform that we have made a conscious decision not to be part of." A third found her entire catalog listed, and put it best: "There's a level of autonomy and consent that's being violated."

Amazon's description of the programs: they help businesses "reach new customers and drive incremental sales."

The opt-out was an email address.

So software walking into a store uninvited is incremental sales when Amazon's software does it, and unauthorized access to a protected computer when it walks into Amazon.

The difference isn't the conduct. The difference is which one has an ad business attached.

When Amazon's agent goes to your website, Amazon loses nothing — it wasn't earning ad revenue on your website. When a customer's agent comes to Amazon, Amazon loses the most profitable moment in its entire operation: the seconds a person spends looking at a page it rents by the impression.

The behavior Amazon called theft is the one that costs it money. The behavior it called incremental is the one that doesn't.

Every company draws its lines somewhere. Very few draw them this precisely along the revenue.

And notice who never got asked. Not the retailers whose catalogs got listed. Not the customer, who had already given permission and whose permission turned out not to be the one that counted. And not you — the seller who funds the largest piece of that line, whose ad costs are set by an auction whose supply of impressions is now a contested legal question, and who will find out how that resolves the same way you find out about everything else. Afterwards, in a fee schedule.

Now the honest part.

Amazon's stated reason isn't fake. Bot traffic is a real problem, marketplaces do eat the fraud, and "a trusted shopping experience" is a thing companies genuinely care about — not least because it's also worth money. And this ruling didn't end anything: contract claims, tort claims and IP claims are all still live on remand, and Amazon can still block agents technically. It just can't call it hacking.

Read that as cynicism if you like. But a company's real position is in its cost structure, not its filings — and Amazon just told its shareholders, in numbers it is legally obligated to get right, that the fastest-growing thing it owns is a business that requires you to keep looking.

Amazon's own AI was doing the thing it sued over. Shop Direct and Buy for Me read third-party retailers' sites and list their products on Amazon without permission. Retailers found out through fulfilment; the opt-out is an email address. Worth ten minutes if you sell anywhere off Amazon — go check whether your DTC catalog is already listed, because nobody is going to tell you.

The BSA landed Monday, and the words nobody quoted still aren't there. The assignment and pledging change took effect August 24 — Tuesday had the full audit. One update: I searched the redlined Business Solutions Agreement that Amazon publishes on its own servers, and the words pledge and collateral appear nowhere in the document. The change may well be real. Every write-up quoting that language is quoting a paraphrase of a May news report.

October 21 is the date your Q4 turns on. October 14 for AWD, October 21 for FBA with minimal splits, October 28 with optimized splits. Peak fees run October 15 to January 14 — about $0.32 a unit over non-peak, flat year over year, plus a 3.5% fuel and logistics surcharge. The fees didn't move. Your ad cost per unit did, and that's the one nobody has on a chart.

Forget Amazon's numbers for a minute. Yours are the ones that pay your salary, and there is one you almost certainly don't carry.

Most operators cannot tell you, in under a day, what it currently costs them in cash to sell one unit. It's the cleanest number in the business and it lives on no dashboard.

Delegate to whoever pulls reports. Forty-five minutes. The output is one number, twelve times.

For each of the last twelve months, get two figures. Total advertising spend for the month — the advertising console, all campaign types, not one portfolio. And total units sold for the month, from Business Reports — all units, organic included, not attributed units. Attributed units are the ads reporting on itself, and they will flatter you.

Divide. That's your ad cost per unit sold. One number per month, twelve rows.

That is deliberately cruder than TACoS and that's exactly why it works. TACoS moves when your price moves, so a price increase can hide a rising ad cost completely. Ad cost per unit cannot be flattered by pricing. It answers one question, which is the question: what does it now cost me, in cash, to move one unit off this platform.

Now read the shape of the line. Flat is a healthy account. Rising while units rise is usually fine — you bought growth. Rising while units are flat or falling is the one that matters. That's the auction taking margin, and it will not announce itself. It shows up as a slow drift you attribute to seasonality for three quarters.

Then do the version that hurts. Same twelve rows, but for your top five ASINs individually. The account average hides everything. Nearly every catalog has one product quietly carrying an ad cost per unit that would fail its own P&L if anyone looked at it alone, and it survives because it's averaged in with the winners.

Artifact: twelve numbers on one line, plus five product lines beneath it. No dashboard, no tool, no consultant. It costs one report pull and one division, and it's the single number I'd want in front of me before signing off on a Q4 ad budget in September.

You don't need anybody's benchmark to read it. Your own twelve months is the benchmark, and the slope is the answer.

Amazon told a court it was protecting a trusted shopping experience. Amazon told its shareholders that advertising grew 26% while paid units grew 17%, and that Sponsored Products is the largest thing in that line.

Both statements are true. Only one of them explains the lawsuit.

Go work out what a unit costs you. That number doesn't care what anything gets called.

See you Tuesday.

Any questions? Book a meeting with me here.

Dan Head

Founder, AMZ Elite