There’s been a lot of talk and uncertainty about the pricing models, minimum spends, and expected results for OpenAI’s ChatGPT ads.

They launched reservation buying only with a $60 CPM because that’s more or less what Netflix did, and they wanted a big splashy talked about premium ad product launch.

On that maiden voyage, ChatGPT ads was well positioned as a product for careerists, not people looking for serious results; I discussed this and voiced my sincere admiration for their cynical pursuit of easy money, and my admiration for anyone willing to hit the “easy button” to get a nice resume bullet point by running in that alpha / beta.

Mo Bitar just released a video saying it better than I could:

One of my dear readers contacted me to tell me that I got a whole lot wrong in that post, but also to confirm (with better information that I had) that I was very right about one thing: OpenAI’s senior brain trust don’t like ads, don’t care about making a good ad product, and maybe even kind of want it to fail.

I think that was maybe the most important thing to get right because it explains how everything with ChatGPT ads has gone so far, whether you think it’s fumbling around in the dark, or a concerted effort to do the cheapest speedrun of remaking Meta Ads circa 2019.

In the spirit of feeling like I’m riding high off of telling you so, I’m going to look at what’s happened so far, explain the whats and whys, and tell you where it’s going to go.

Near the end I am going to suggest something I think is kind of novel, even though it’s an old idea, and volunteer to take it on for OpenAI in case Sam or Fidji or whoever are reading this.

Click here to skip right to my proposed solution-esque idea, which won’t hurt my feelings.

It’s both true, and a bit misleading, that the CPM model ChatGPT ads launched with is just natural to develop first because you essentially have to build it into your ad serving tech and commercial model before anything else.

If you were the kind of rogue founder who declared you never even wanted to see the word “impression” your team would have a really hard time showing you a flowchart of the user’s ad experience.

There’d be a node in there weirdly titled “full ad render” or “first MRC accreditable transaction point” or something and you would be proud of them for trying but you would know: that’s an impression.

Then your CFO asks you to hang back for a minute with them after the meeting, and it’s so that they can tell you that the investors admire your ambition to do something different, but they’re afraid you’re trying to reinvent the wheel, or worse, not meeting your customers where they are, nor speaking their language, and they think impression buying should be on the table, and just like that you have a CPM model.

It’s a much more reasonable version of this kind of convo, which is how 90% of ad tech sales led GTM strategy forms:

The thing that’s misleading about CPM buying being a necessary and simple first step is that it isn’t as bad as some people out there in ad land think it is. That’s because CPM is cousin to Cost Per Person Reached, or just reach and frequency if you’re nasty.

Despite what a lot of direct response and related disciplines (e-com, DTC, GTM, growth marketing, etc.) bros would all scream in unison at me, these metrics are Very Good, Actually, and even fundamental to understanding ad performance and the nature of what you’re even doing with your dollars.

For longer than any of these digital channels have existed, there have been mountains of evidence that advertising is basically the act of showing as many people your ad as possible, and probably keeping the number of times each person sees that ad from being too high in a short time period, and not really a whole lot else.

While I will later argue this is probably a worse metric for LLM chatbot ads than any digital ad medium before it, people will just want to reach as many people as possible via ChatGPT, even if this is a much messier way to look at a campaign than if it were digital video or anything else.

If you’re wondering what I think about better-than-impression metrics for reach and human attention, oh, don’t worry, as one of the longest tenured advisors to Adelaide, you can be assured I’m going to strip down to my little shill skivvies and shake my stuff at you about all that later in the piece.

If I had to do that kind of shimmying regularly, here is what I think my song would be as a bonus secondary soundtrack piece for all of you:

The point I am making, though, is is that many people on the buy side shouldn’t be trying to race beyond CPM buying as an option.

The funny thing is that, excluding one channel I will touch on in a moment, CPM ends up being the foundational commercial unit for modern digital channels even when they’re heavily optimized to deep performance metrics.

As someone running all kind of complex event driven and cohort growth oriented software campaigns for pieces of software with rich analytics layers ported directly into media buying platforms, I am paying almost exclusively on a CPM.

You used to be able to toggle “pay on CPC” in Meta Ads, for instance, but that’s just an intermediary step between impression and your ultimate conversion, and this tends to just be a tool for people gaming attribution, either their own or someone else’s, for the purpose of accruing hollow credit for actions they didn’t cause.

Even mighty Google Ads, largest and nearly last surviving scion of the Cost Per Click economics, used to have a deeply buried CPM campaign option, presumably for big CPG types who were just pumping it into an MMM that rewarded high impression volume.

There’s even a lesson to be had there-if there’s easy money to be had in peddling pounds of impressions, even a bit cynically, I wouldn’t run away from that before I was a profitable company.

This is all to say I expect CPM to be around for good as the ad offering evolves and is better tailored to the needs of all advertisers, reach oriented and “performance” alike.

CPC, I do not expect to survive…so long as the second C in that is Click and not Chat.

This is just a signal / lazy roll over into appealing to “performance” advertisers.

Real performance advertisers that try to deliver serious incremental gains aren’t clamoring for this-this is to appease some sort of already active advertiser base that is panicking or, at best, trying to glean some information that I don’t think will shed any light on this ad product’s real value in the long term.

The burning question they’re trying to answer is: what REALLY happens when a user hits a website from ChatGPT? I am sympathetic to people who want to know this and can see how we got here, which is by following in some smart footsteps.

The behavior of users referred by ChatGPT is a major concern for two groups that are much further along in really understanding LLMs and than advertisers are: publishers and SEO / content / discoverability people.

For publishers this is a legitimate area of concern and research for a very simple reason-the only viable way for them to make money for all of history and as far as anyone can tell, in the near future, is for people to actually go to their websites.

They have seen search engine traffic collapse, even long before LLM mania due to Google’s various zero-click SERP widgets, and they need to know if what juice is coming in through LLMs is worth the squeeze.

The SEO-cum-GEO-plus-content people are in a somewhat similar boat; while they have other options and many of the smartest ones are starting to embrace them, everything started and ended with website visits for a long time in their world.

There’s a healthy debate about the relative value of showing up on zero-click searches for their clients, since there can be value there for commercial enterprises…well, except for web publishers.

As advertisers, it is a great blessing to us that we have many options to derive value from user journeys that do not even involve a click on an ad that goes to a website, and although they’re scary, perhaps we can take an example from the greater courage our publisher and SEO brethren have had facing a much deeper crisis, and fully embrace those.

As we’re thinking about things besides clicks to our websites, I would not like to be so laser focused on one particular area as the most “websites are dead” oriented marketers have been: social shopping, and “everything app” commerce.

Here is where I would normally hate on “agentic commerce,” but in a piece about ads in an LLM product, that seems silly.

Social shopping and “everything app” commerce have been vastly over-promoted in the west because of lazy analysis that goes like “it’s big and China and so it will inevitably be big here.”

Actual China understanders have been saying for a long time that this isn’t true, and that the current Chinese e-commerce environment formed under unique-to-China circumstances from the 2000’s and 2010’s that are absolutely not going to replicate themselves anywhere else, especially not anywhere that developed its internet ecosystem before they did.

People have been saying this since at least 2018, they’ve been right for long enough, let’s put this to bed.

I think when we consider non-click actions, we don’t want to limit ourselves to opportunity that requires commerce shifting to entirely new spaces; we want to try to change hearts and minds enough to drive impact clearly visible everywhere our goods and services are purchased.

What excites a lot of people about LLMs is that they have an incredible persuasive power that we’ve never seen before, and I have read way too many accounts of “AI Psychosis” to debate that.

This is a factor of the amount of attention and trust people pay to them, and those are extremely powerful, however misplaced they may be.

Why am I not going to be super bullish on the attention measurement family of methodologies here?

Because what would concern me about the attention economy, and definitely did concern OpenAI, is that their product fundamentally speaks a dying language (text) and struggles to sling the new slang.

I am talking, of course, about short form video.

OpenAI already tried to tap into the media format that rules the world currently, and it did not go well.

I’m talking, of course, about Sora. Breaking down what happened here would be its own whole article so for now let’s just all accept that OpenAI has left short form video and isn’t coming back anytime soon.

This is a huge issue because short form video has eaten the world, to the extent that even people who lament this are coming to terms with it.

Political pundit Jamelle Bouie is on BlueSky begging people to get on TikTok and YouTube if they want to actually persuade people of anything, and has taken his own medicine on that one.

Ryan Broderick, too, has both lamented that this is probably not good for the world but it is a necessity for any kind of serious effort at mass communication.

So the principal method for mass influence has slipped through OpenAI’s grasp, which means two things could be true:

If you think there’s a way forward here, they’ve got to use digital media to build brands via text. That’s going to be tough sledding, so they would need to make their “performance” ad business sizzle at scale quickly.

If you are rolling your eyes seeing that this is my big idea, believe me, I am right there with you.

Now, get ready to roll your eyes again, even harder, when I tell you why I think it could be different this time: the power of agentic LLMs.

First, let me tell you from extensive experience as a buyer and seller of CPA / revenue deals and “affiliate” marketing why this has never worked enough to completely take over the industry like the most fervent Incentives Perverts believe it should.

The issues are measurement governance, deal structure, and operations; the first two problems are exacerbated deeply by the third.

First, you have to ink a deal that everyone expects will need to be changed periodically, which is already a problem.

Then, you’ll need to determine a contracted metric, such as leads or revenue dollars, and a system of record for measuring it. This system will need to be maintained as flawlessly as possible in terms of uptime and also the specific parameters it’s using to measure things, and if you’ve been deep in this business, you know how big a part of the issue this is.

Ask any affiliate program manager if they’ve ever experienced website conversion tracking breaking, or even a whole e-com site going down, and how this makes all their partners feel.

At the end of the day, you can sum up the reason less CPA / rev share deals are done at mass scale pretty simply: the juice isn’t worth the squeeze because it’s hard to sustain profitable, fair deals built on the ever-shifting sands of the internet.

That squeeze, though, is a lot of contract work, re-crossing of t’s and re-dotting of lower case j’s, over and over again, tediously, with lots of opportunities for miscommunication.

If LLM agents in advertising cannot bring us a solution to this problem, than I ask you truly: what the fuck can they bring us?

Lo and behold, there is one platform that has more reason than any in the world to make agentic LLMs more relevant in advertising, and it’s OpenAI.

It’s not just that they should be invested in agent success-they also need to throw themselves wholly at purely performance based advertising because the industry is going to take a long time to really get value out of their ad product if they don’t.

That’s the real best reason for OpenAI or any company to sell value, as they define it, instead of metrics-when other people will struggle to properly derive the value themselves.

I weirdly find myself thinking a previously useless technology maybe has a moment here: how could one best ensure that these mercurial, dynamic LLM based deals are properly tracked and accounted for and value and payment are fair?

(BIGGEST SIGH IN THE UNIVERSE)

Possibly a blockchain.

I’m not going to say anything else except that blockchains were the most overrated technology in terms of real utility in my lifetime, generally have no place in advertising, and I’m shocked I’m even saying it’s POSSIBLE they have a utility.

If this is not the time to give them one more at-bat, I do not know what is.

We need the verification of contracts and changes to contracts with precise time stamps and details, in an environment that is very slow moving compared to real-time ad auctions.

Again, if this is not something this technology can bring advertising-what can it even do?

OpenAI can’t even rely on their most fervent boosters to carry them in the long run, because when those people talk about LLM ads, it’s like:

“Search CAPTURED intent for single transactions. LLM ads will FREE humans to agentically generate infinite abundance via product and service discovery unparalleled in all of human history.”

This is vague and meaningless and while it gets good LinkedIn engagement right now, it’s it’s not a load bearing body of thought in the long run.

People in the middle and on the skeptical side generally aren’t thinking that much more deeply about it, mostly posting things like “CPM too high / CPC too high / results not there / scale not there.”

A very smart thing Luigi Ferguson said on LinkedIn is nobody should be surprised or upset if an experimental / in beta ad platform doesn’t scale. Much worse things could happen, also, like lots of spend and no results.

I agree, which is to say smart buyers are actually probably happy to give OpenAI some time and breathing room to get this right on their own terms.

This is good, because with the “results not being there” complaints, the fact that OpenAI doesn’t really own a serious browser, nor has analytics software deeply installed on a huge portion of websites and deep within app SDKs, is going to make it tough to win over the instant results seekers.

The “I need to see results fast” people are always talking about deterministic digital attribution and you need to be deeply buried in many ecosystems to have a competitive offering there.

Getting to “CPM and CPC too high” complaints, I find myself a bit surprised to be the one to firmly say this:

This is probably the time and place to retire these as valid complaints forever.

Optimizing to impressions on LLM ads brings up questions like: if a brand is mentioned multiple times in a conversation is that multiple impressions? By the standards of all other advertising, yes, but do you want what people call “ad collision” when it happens on websites happening in your chat ads?

I think the answer is actually “probably,” but there’s going to be lots of different cases that need different approaches, where I don’t think “impressions” as every “ad unit” shown help.

Impressions are probably best rolled up to the first brand ad in any convo, which then makes them something likely a lot more valuable than the average paid media impression, to an extent I don’t trust most buyers to properly value them.

Buying CPC tends to be tied to website or app side conversions, and while I think CPC would be fine for e-com advertisers when you blend in certain other click types like Google Shopping has for their ads, it’s going to cause ChatGPT Ads to struggle to drive investment that helps work on “the messy middle,” a term Google loved to use for a while to highlight their own advertisers’ issues with spending on Google Ads for anything but cold hard ROAS.

So for ChatGPT ads, CPM is a uniquely trash upper funnel metric and CPC is a barely mediocre “lower funnel” metric here.

OpenAI need to run large scale experiments on making their top level economic units a Cost Per Chat Entered and Cost Per Chat Ad, the former being participation the first time in any given chat session, and the latter being total ad units + mentions of the brand in the whole chat.

I put session in there because I think they need to define that and prevent odd “in perpetuity” situations with ultra long conversations. Or maybe just charge for them differently, I don’t know, that’s why real large scale experiments are needed.

If they want to charge separately for CPC, as in clicks to site or a native shopping experience PDP, that’s probably fine, but I expect that when they finally nail e-commerce, that metric ends up looking quite expensive and still seen as a useless intermediary to conversion that doesn’t correlate much better than with impressions.

That or it ends up like Meta Ads, where it’s an incentive trap to find the most click-happy and purchase-averse people on Earth.

This is going to drive standardization people crazy but look, Meta exited MRC audits last year and the spend didn’t dry up. If we’ve already lost the world I think it’s time to consider measuring new things differently in the new hellscape being born.

I will do a whole piece on this and link to it when it’s finished, but whether we’re talking about The Trade Desk or Criteo or Stackadapt, that just doesn’t end up being the story here.

It’s maybe part of the story, or I am totally wrong, and there’s definitely money to be made temporarily, but history indicates this is a really hairy beast to tame.

It requires inventing and reinventing some things and making people okay with that, which is hard.

They need a fast pivot to more chat centric billing and sensible, slow (but not glacial) development of metrics specific to LLM ads that make sense and are fair and can be surfaced in advance of being a currency.

Lucky for them, there is a whole area in which they definitively do NOT want to reinvent anything, and can leverage whole existing ecosystems.

Independent third party verification, of the outcomes that I am saying they should rush into the business of selling, is robust, and ready for them.

Whether it’s geographic and time structured incrementality testing systems, MMM, panel based lift studies, or anything else you care to mention, every major player in these spaces would love to partner with them. They’d be number one of everyone’s roadmap immediately.

This also leans them away from major areas of weakness in tracking I think are being vastly underplayed, which is web analytics implementation, commerce data conversion integration, and, even more so, app ecosystem ownership and SDK implementation.

They don’t have the ubiquitous Meta Ads web pixel, let alone Google Analytics.

Beyond that, every time I see people propose some date where their ad revenue surpassed Google and/or Meta, I wonder how much the person saying this knows about the ad revenue these companies get from app advertising, and how much lift all their advertising gets from owning the Android ecosystem or the world’s most ubiquitous mobile app SDK that isn’t Firebase (also owned by Google).

People aren’t going to use that browser, devs aren’t going to race to stick an SDK they make in everything, and this announcement that they might make a phone, is, uh…interesting.

I find it odd that I’m arguing against OpenAI building ad measurement infrastructure to grade their own homework, because this is the most successful strategy in digital ads history, bar none.

I just think it actually might finally be too late for anyone else to do that, barring truly seismic changes to the internet and the way people use it.

To be honest with you, I am quite bearish about OpenAI being The Thing in LLMs and chatbots for much longer.

I was bullish when most people were bearish on Google after those stupid Bard demos, because I thought having lots of products with massive user bases it would be easy to integrate an LLM into was probably worth more, and I think they’re going to win the mass consumer LLM money game.

There is a whole other side in B2B land, or productivity or enterprise or whatever you want to call it, where it seems like Anthropic and smaller players like Cursor have completely squeezed OpenAI out.

The way that I think OpenAI could be positive for the open internet is contingent on them being a different company going forward than I think they are now, and also, the internet and world continuing to get worse in certain specific ways.

I’m going to cover this phenomenon’s impact on the internet and internet tech soon, but for now I just need to nod to the fact that lots of people are running out of money.

The K shaped economy, Raj Chetty’s work, “the 1%” and all, you know what I’m talking about. While general economic indicators haven’t gone in the toilet 2008 style and some have even consistently improved for a long time, more and more people are struggling to afford their basic needs.

Here’s where the “working in digital ads is heroic” people come in and say “see, all these free services that anyone can get are the true miracle of digital ads!”

I am not so optimistic because I think putting ads in things can create really bad experiences for users, and incentives for the builders of products, and the more this is foisted off exclusively on poor people, the more we’re doing the whole K Shaped World thing I’m alluding to.

That being said, OpenAI is already the Coca-Cola of AI. The user stats are pretty staggering. If it makes sense for one LLM to stay free, and focus 100% on the experience of its free user base, it’s already ChatGPT.

This also assumes LLM chatbot use is a positive thing for most people, or gets there. That’s a whole other blog post, or 50, that I’m much less qualified to write than lots of people already doing it.

If you do assume access to LLMs is a net good in the long run, you would want access to be universal, at least in some basic form, and ideally with a tolerable experience.

Business outcome oriented ads are, if I am being honest with myself, probably one of the more tolerable ways to make something free that currently exists for software.

For that reason, I would try and “fix” ChatGPT ads in the way I have outlined above, hoping that this is the least bad way to keep the average person somewhat close to par in terms of access to LLMs as those who can pay for expensive subscriptions.

Oh, but I would absolutely never bring back Sora, nor try and double down into generative podcasts or music. Those are absolute poison.