If you’re new, welcome! If you’re a regular, thanks for being here as we navigate this AI shift in eCommerce. I share what I’m testing and what’s actually moving the needle. Where’s the vault you ask? Every previous edition is saved here.

In this edition (shortcuts):
Stop Paying Your AI to Read💥
Claude Watermarks Everything It Writes 🤓
5 Claude Skills for PPC with Orange Klik 🎤
News and Events Worth Following

The price of a token fell by about two thirds this year. Almost everyone's bill went up anyway.

That is not a contradiction and it is not a billing error. It is the most important thing happening in AI economics right now, and once you see the shape of it, the way you work with these tools changes permanently.

Here are the two numbers side by side.

Blended inference costs dropped from roughly $18.40 to $6.07 per million tokens between the first quarter of 2025 and the first quarter of 2026. That is a 67% cut in twelve months. Zoom out three years and per token prices are down something like a thousandfold.

Over that same period, industry analyses put the rise in enterprise AI spending at around 320%.

So the thing got radically cheaper and everyone started paying radically more.

The reason is agents. When you asked a chatbot a question in 2024, that was one model call. When you hand an agent a task in 2026, it fires 10 to 20 model calls to complete a single action. It reads a file, thinks, calls a tool, reads the result, thinks again, writes something, checks its work.

Average prompt tokens per request grew roughly 4x in 13 months. Completion tokens nearly tripled. Goldman Sachs has modeled agentic demand pushing total token consumption up as much as 24x.

You did not get more expensive. You got more automated.

Here is why this matters: every instinct people have about managing AI cost is aimed at the wrong number. Hunting for a cheaper per token rate is optimizing the variable that has been falling on its own the entire time. The variable that is actually moving is how much you consume, and that one has no ceiling and nobody is managing it.

This is my read rather than a reported fact, so take it as a working theory.

Nobody is charging you what this actually costs to build. The training runs, the data centers, the power contracts, the researchers: none of that is priced into what you pay per month ($20~$100). The gap is being covered by investor capital, and it is being covered on purpose.

What the labs are buying with that gap is not your subscription revenue. It is your habits.

Every workflow you build at today's prices is a workflow you will still be running at tomorrow's prices, because by then it will be load bearing. That is the trade. Cheap tokens now in exchange for a business process that cannot be unwound later.

There is a part of this that is not a theory, and most people reading this have never checked it.

On the consumer plans, your conversations are used to train the model. Not as a hidden practice, but as a setting, and since August 2025 that setting defaults to on. If you leave it on, retention runs to five years.

The picture is genuinely different by surface, and the distinction matters:

🔶 Consumer plans (the Free, Pro and Max style tiers): training may be on
🔶 API and business tiers: carved out. Not used for training by default.
🔶 One caveat worth knowing: conversations flagged for safety review can still be used regardless of your setting, per a policy update that took effect in July 2026.
🔶 Incognito style modes: never used for training at all.

So on a consumer plan you are paying twice. Once in dollars, once in training signal. Only one of those appears on your invoice.

I am not making a villain out of this. It is a reasonable deal and the setting is right there. But "reasonable deal" and "a deal you agreed to on purpose" are two different things, and the second one requires actually opening the page.

Do this before you read the rest: open your AI tool's privacy settings and look at the training toggle. Decide what you want it to say. It takes forty seconds and most people have never done it.

Now the practical half.

Once you accept that consumption is the number that matters, the question stops being "which model is cheapest" and becomes "why is this model reading so much." Because that is where the tokens go. Not into brilliant reasoning. Into reading.

There are four kinds of reading you are paying for right now, and three of them are optional.

This is the big one and it is the one almost nobody gets right.

When you paste a 50,000 row search term report into a chat window and ask for the top wasted spend, the model has to read all 50,000 rows. You pay for every one of them. Then it does arithmetic across them in its head, which is the thing language models are worst at, and it may hand you a number that is subtly wrong.

Ask it to write the code instead.

Ten lines of code. The model never reads a single row of your data. The script does, on your machine, for free, and it gets the arithmetic exactly right because arithmetic is what code is for.

Cheaper and more accurate at the same time is a rare combination in this business. This one delivers both, and it scales the right way: the script costs the same whether your file has 500 rows or five million.

The same move works for SQL. If your data is in a warehouse, do not export it and paste it. Ask for the query.

Every single turn in a conversation resends the entire conversation.

Turn two sends turn one. Turn twenty sends turns one through nineteen. Turn forty sends everything that came before it, again, at full price. That long running thread you have kept open since Tuesday is charging you for its first message forty times over.

This is where that 4x growth in prompt tokens per request is coming from, and the fix costs nothing.

Start a new chat. Paste three lines of context summarizing where you got to. Carry on.

I know it feels wasteful to abandon a thread that has all the context in it. It is the opposite. The context is worth keeping, the transcript is not, and a three line summary carries the first and discards the second.

My rule of thumb: when a conversation stops being about the thing I originally asked, it is a new conversation. When I catch myself scrolling up to find something, it is a new conversation.

Don’t paste screenshots if you can paste text. Images and screenshots are expensive and the cost is invisible because nothing on screen tells you what an image just cost.

The rough formula:

Run that on the things people actually paste:

That last one is the number that should make you pause. A single unresized product photo can cost more than a long document.

Two habits fix most of it. Resize before you upload, because the model does not need 4000 pixels wide to tell you the background is not pure white. And when the content is text, send text. A screenshot of a table costs thousands of tokens and forces the model to read numbers off an image, which is exactly when it starts inventing them. The underlying CSV costs a fraction and is read correctly.

This is the same failure I keep running into from the other direction: ask a model to read text off a picture and it will confidently quote you words that are not there.

You already moved your brand guide into a skill. Good. But moving it did not make it free. It made it invisible.

Everything in your “always loaded layer” gets read on every request. Your instructions file. Your tool definitions. And the name and description of every skill you have installed, not just the one you called.

That last one grows with your library. Fifty skills means fifty descriptions riding along on every message, used or not.

The fix is one line: keep the skill’s front matter clear and the descriptions short. The body only loads when you invoke the skill. The description you pay for all day. Don’t make your skills read like SOPs, they are not!

Not everything is about reading. Three others worth knowing.

The obvious move when a bill climbs is to switch to a smaller model. I tried it recently and it did not work out the way I expected.

Recently we moved some listing optimization work from Opus 4.8 down to Sonnet 5. On paper it was an easy call. Sonnet 5 is a coding beast and it was running about 60% cheaper per token.

It was a mistake.

The coding reputation is deserved. Language is where it fell short. Sonnet 5 runs compact, and compact is exactly wrong for listing copy. The output came back clipped and flatter, stripped of the phrasing that makes a bullet actually work. Same prompt, same inputs, a different voice entirely. Required a bunch of expensive rework.

Moral of the story: Cheaper per token is not cheaper per finished listing.

Anything you do not need back in the next few seconds can run as a batch job at roughly half the standard rate.

For an e-commerce operator, a surprising amount of AI work qualifies. Rewriting a few hundred listings. Classifying a quarter of search terms. Running sentiment across a review backlog. Translating a catalogue for a new marketplace. None of that is interactive, and all of it is usually paid at full price because nobody thinks to ask for the cheaper lane.

The test is simple. If you would be perfectly happy getting the answer after lunch, you are paying a premium to have it now. So turn it into a script or a SQL query that will run on a schedule and not consume tokens on each run.

We run Codex and Claude alongside each other rather than picking one.

The thing that makes that affordable is not either tool. It is the document between them. When work moves from one to the other, the first one writes a handoff: what it did, what it decided, what is still open, and what the next one needs to know. The second tool reads that document. It never reads the conversation that produced it.

Which is the same move as starting a fresh chat, applied across tools instead of inside one. You are not paying the second tool to sit through forty turns of the first tool's reasoning. You are paying it to read a page.

It has a second payoff that has nothing to do with cost. Once the state of a task lives in a document instead of in a thread, the work becomes portable. You can hand it to a different tool, a different model, a teammate, or yourself next Tuesday, and none of them need the history.

Every tactic in this article saves money today. That is not really why I think they matter.

They matter because we are all learning how to work with these tools during the one period when the meter is running slow. Prices are falling, capital is covering the difference, and the habits being formed right now are being formed under conditions that will not last.

The people who get caught out will not be the ones paying more per token. Per token prices may well keep falling. They will be the ones who learned to work in a way that consumes ten times more than it needs to, at a moment when that did not hurt, and then found out what it costs when it does.

The discipline is the asset. Not the savings.

So the useful question is not how to lower this month's bill. It is whether the way you are working right now would still make sense if everything cost five times more. If the answer is yes, you are fine at any price. If the answer is no, you have found your homework, and today is a considerably cheaper day to do it than next year.

Know anyone who wants to work with an agency that has AI and tech in their DNA and are pros at managing advertising and creatives for E-commerce? Forward them this newsletter and let them connect with us at [email protected].

Since August 2, newly released Claude models watermark their own text. Anthropic does not describe any way to switch it off.

If you read my prompt injection piece, your first instinct will be invisible characters. Mine was. That is the attacker's trick: zero width spaces and Unicode oddities sandwiched into a product review, invisible to you, perfectly readable to your agent. The orange juice trick, in text.

This is not that. Anthropic is explicit: "Nothing is added to the text and there are no hidden characters."

So where is it?

It is in the word choices. At many points while writing, several words would do equally well, and the model normally picks between them at random. The watermark swaps that randomness for a keyed pattern, so the path taken through all those equally good options becomes the signature. The mark is not hidden in the text. The mark is the text.

Which is why a hex editor finds nothing, and why find and replace cannot strip it.

🔶 It exists because of the EU, whose marking rules bit on August 2. Every other major developer signed the same code of practice, so this is where the industry went this month, not a Claude quirk.
🔶 It marks involvement, not authorship. A hit says Claude touched the text, not that it wrote the thing. A paragraph you wrote and asked Claude to tidy carries the same mark as one generated from nothing.
🔶 It survives copy, paste and light editing. A full rewrite removes it, and reportedly so can translation.
🔶 It takes a key to read, and Anthropic holds it. They say a detection API is coming, which is the moment this gets interesting.
🔶 Images are different. They get C2PA credentials in the file metadata, a label attached rather than baked in, so a screenshot strips it.

Here is the part that matters for listings.

Anthropic says detection "doesn't work well on small samples" and is "sparser on factual passages." Amazon copy is both. A 75 character title has almost no room for the word choices the mark lives in. So the shorter and more factual your copy, the less signal it carries. Long flowing brand copy carries the most.

To be clear about what this is not: nobody outside Anthropic can check any of this today, and no marketplace is doing anything with it. Anyone telling you Amazon can now detect your AI copy is inventing that.

But your archive is being signed as you build it. That is the quiet part. You are not choosing whether your copy carries a signature, only whether you know.

I joined Augustus Kligys of Orange Klik to walk through the five Claude skills we actually run at the agency for Amazon PPC: a bulk file builder, a bid adjustment engine, keyword gap analysis, negative keyword cleanup, and performance reporting. The honest half of the conversation was the other list, the things I would still not hand to AI, and why I refuse to give the skill files themselves away: they are business judgment rather than code, so mine would be close to useless in your account.

Watch it here: Use AI to Automate Your Amazon PPC Tasks

PPC Ninja is touching down at Amazon Accelerate 2026 in Seattle, September 22–24. I am thrilled to share that I’ll be joining the Amazon Ads team at their booth for a featured live interview! 🎙️✨ We will deep dive into the future of marketplace search, AI workflows, and next-gen ad performance. If you are there I would love to chat, just hit reply on this email and let me know.

🔶 Amazon now makes you label AI generated people in your images If a listing image or video contains a photorealistic AI generated person, you have to write contains-synthetic-performer into the file's metadata before you upload it, and Amazon shows a badge to shoppers. It follows New York's synthetic performer law, and it only bites on photorealistic people, not on real people you edited with AI.

🔶 Ad costs now come out of your proceeds The billing change that triggered a one day ad boycott in April landed on August 1. Sponsored Products, Brands and Display are deducted from retail proceeds instead of hitting a card, which removes the float a lot of sellers were quietly running on. You can still choose Pay by Invoice in your billing settings for 30 day terms, and there are click credits running through December.

🔶 Peak season FBA surcharge is set An average of 32 cents extra per unit on everything leaving a fulfilment centre between October 15 and January 14. Worth putting into your Q4 margin math now rather than finding it on an invoice.

🔶 Solution Provider access is changing Amazon replaced the old invite flow for granting agencies and tools access to Seller Central, starting August 10. If you onboard a new provider this quarter, the steps you remember are not the steps any more.

We hope you liked this edition of the AI for E-Commerce Newsletter! Hit reply and let us know what you think! Thank you for being a subscriber! Know anyone who might be interested to receive this newsletter? Share it with them and they will thank you for it! 😃 Ritu