Atlas launched in October 2025. It is being retired about ten months later, having never shipped beyond macOS. OpenAI has said the browsing features move into products people already use, which in practice means the ChatGPT desktop app, a Chrome integration, ChatGPT Work and Codex. Conversation history lives in your ChatGPT account and is unaffected.

The bit worth paying attention to

Losing a browser is an inconvenience. The useful question is what it says about how quickly the ground moves under AI tooling, and how much of your working process you should let sit on top of it.

Consider a three-person agency that spent a fortnight last spring building a competitive research routine inside Atlas. Tabs organised by client, agent instructions saved per project, a browsing workflow the junior strategist was trained on. None of that structure survives the shutdown. The underlying skill does, and the ChatGPT account does, but the specific arrangement of it is gone, and someone has to rebuild it in a different tool during a normal working week.

That is not an argument against adopting new AI tools quickly. Moving early is often the whole advantage. It is an argument for being deliberate about what you build on top of them.

Three questions before the next tool goes into your workflow

Can you get your work out? Not “is there an export button” but “if this closed on Sunday, what would I lose and how long would recreating it take”. Atlas users are finding out that the answer for bookmarks and saved logins is manual and fiddly.

Is this the vendor core product or a side bet? A browser was never how OpenAI makes money. Assistants, ads and enterprise seats are. Side products from companies moving this fast get folded back in, and the reported security problems around prompt injection in agentic browsing did not help the case for keeping it alive.

How many people depend on it? A tool one person uses is a preference. A tool four people were trained on is a process, and processes need somewhere stable to live. Keep the process written down somewhere the tool cannot take with it.

What can go wrong if you overcorrect

The wrong lesson here is to wait for AI tools to prove themselves for a year before touching them. That caution has its own cost, and the businesses that got useful at prompting in 2023 did it by using things that later changed shape. The risk is not adopting early. The risk is quietly making a fragile tool load-bearing without noticing.

This also may not be worth your attention at all. If you tried Atlas twice and went back to Chrome, there is nothing to do here. The people who need to act are the ones who made it their default browser, and that is a small group, which is arguably why it is closing.

Editor’s note: This area changes quickly, so check the latest platform policy before making compliance decisions.

Before Sunday, open Atlas, export your bookmarks to HTML, and write down anywhere your saved logins existed only there. Then spend twenty minutes writing your research routine into a plain document rather than a tool. The next time a product gets folded into something else, and there will be a next time, that document is the thing that survives.

The post OpenAI Is Shutting Down Its Own Browser on Sunday. Check What You Built On It. appeared first on Mark8ng.com.

On 24 July, OpenAI added a Conversions objective to ChatGPT Ads with optimised cost-per-click bidding, so campaigns push delivery toward the clicks more likely to produce a conversion while still charging on a click basis. The same release brought geographic exclusions, average daily budgets calculated over a rolling seven day window, intraday budget pacing, AppsFlyer and Adjust integrations for app installs, automatic advanced matching using hashed customer data, and asynchronous bulk campaign creation through the Ads API.

Read as a feature list, it is housekeeping. Read as a business decision, it is the moment ChatGPT Ads becomes something a finance lead can put on the same spreadsheet row as Google and Meta. That matters more than any single feature in the release.

Why your first month will flatter the channel

Here is the trap. When a new channel gains conversion tracking, the first month almost always looks excellent, and the reason is rarely channel quality.

Three things inflate early numbers. The audience self-selects, because people clicking an ad inside a ChatGPT answer in the first weeks are curious, high intent, and unusually engaged, and that population does not scale. Automatic advanced matching improves attribution, which means conversions that were already happening now get credited to the channel, so it appears to create demand it merely observed. And small budgets in thin auctions face little competition, so early cost per acquisition sits well below where it lands at ten times the spend.

None of that means the channel is bad. It means the first result you get is not the result you should plan around.

A test design that survives contact with reality

Picture a home services company spending roughly $4,000 a month on Google Ads and producing about 60 leads from it. A sensible test looks like this.

Run four weeks, not two. Optimised bidding needs conversion events to learn from, and below roughly 30 conversions in the learning window the system is guessing confidently. Budget 10 to 15 percent of existing paid spend: enough to generate signal, small enough that a bad month is survivable.

Exclude your strongest existing geography using the new geo exclusions. This is the least obvious step and the most useful one. If you let a new channel run in the area where your brand is already known, you will measure brand recall rather than incremental demand.

Then write down, before you see any data, the number you would need in order to fund a second month. This sounds trivial. It is the step most teams skip, which is why so many channel tests end in an argument about whether the result was good instead of a decision.

Compare against that threshold, not against your Google account average. Your account average includes branded search, which converts at a rate no new channel will ever match.

What can go wrong

Attribution inside a conversational surface is not the same as attribution after a search click, even when the reported metric carries the same name. Someone may ask ChatGPT about a category, see your ad, ignore it, then search for your brand by name two days later. Advanced matching may connect those events. Your Google branded campaign may also claim them. Both platforms report a conversion and you have paid twice for one customer. The fix is not clever modelling, it is a holdout region and an honest look at blended cost per acquisition across everything.

There is a second risk that has nothing to do with bidding. Product feed ads are now showing updated cards with pricing and star ratings. If your feed pricing is stale or your review counts are thin, that unit will display the weakness more prominently than a text ad ever did.

When this is not worth your time

If you generate fewer than about 30 conversions a month in total, skip this entirely. Not because ChatGPT Ads is unsuitable, but because no bidding algorithm on any platform can learn from that volume. You would be paying for automation with nothing to automate. The same logic applied when OpenAI cut its API pricing: cheaper inputs only matter once you have volume running through them.

If your sales cycle runs longer than 60 days, a four week test tells you almost nothing about revenue. Either extend the window or judge the channel on lead quality reviewed by hand rather than on conversion counts.

The number nobody checks

The metric most likely to mislead you is not cost per acquisition. It is the conversion count sitting underneath it. Everyone checks whether CPA looks acceptable, and almost nobody checks whether the system had enough events to optimise on in the first place. A campaign reporting a $22 CPA on nine conversions has proven nothing. A campaign reporting $61 on 140 conversions has proven a great deal. Look at the denominator first, every time, on every platform.

If you want the platform’s own framing before planning a test, OpenAI set out its bidding and measurement approach in its post on new ways to buy ChatGPT ads, and the settings themselves are covered in detail in this breakdown of the July release. Start with a budget you would be relaxed about losing in full, and set your success threshold before the first impression serves.

Editor’s note: This area changes quickly, so check the latest platform policy before making compliance decisions.

Update, 6 August 2026: measurement arrived faster than expected

OpenAI has integrated AppsFlyer to bring in-app measurement to ChatGPT ads, with Grubhub and roughly forty other brands running the first tests. That closes part of the gap this post described, because advertisers can now connect a ChatGPT ad to what happened inside an app rather than guessing at the far end of the funnel.

The recommendation here does not change, but the timeline does. The original argument was to wait for measurement before moving real budget, and measurement is arriving in pieces. In-app conversions are covered. Web conversions, view-through effects and anything involving a long consideration window are not, and a closed-loop number supplied by the platform selling the media still deserves a sceptical read.

If you deferred a test on measurement grounds and you have an app, this is the point to revisit it with a small budget and your own analytics running alongside.

The post ChatGPT Ads Just Got Conversion Bidding. Here Is How to Test It Without Wasting Budget. appeared first on Mark8ng.com.

What actually shipped

Klaviyo moved Composer into public beta at the end of June. You describe an outcome in plain language and it builds a launch-ready campaign, including the audience segments and the messaging across channels. The company is explicit about the direction it is heading: brands define outcomes, agents execute them, and the execution layer moves from people to software.

AnyMind Group has been shipping the same idea in pieces across a different part of the stack, with agent-driven products for media buying, video production and connecting online activity to offline retail sales. Solitics released an agentic engagement suite aimed at retail banking. Three companies, three sectors, one bet.

None of this is a launch story worth writing about on its own. What makes it worth your attention is that agentic execution now has a price and a contract attached, which means somebody in your business will ask about it within the quarter.

The delegation question, answered honestly

The useful frame is not “can AI do this.” It is “what does this cost me when it goes wrong, and how fast would I notice?” Sort your work by that and the answer stops being ideological.

Reasonable to hand over

• Segment building. A wrong segment sends a decent email to slightly the wrong people. Recoverable, and you find out from the open rate within hours.
• Send-time and frequency decisions. Genuinely better handled by a system with more data than you have. Low blast radius.
• Variant generation for testing. Six subject lines instead of two. You still pick.
• Lifecycle flows with a fixed shape. Abandoned cart, post-purchase, winback. The structure is well understood and the copy is not where the money is.

Keep on a human

• Anything that makes a claim. Pricing, availability, guarantees, comparative statements about competitors. An agent that invents a discount you are not running creates a real liability, and you may not spot it until a customer holds you to it.
• First contact with a new audience. The agent optimises against your existing customers. It has no view on the people who are not there yet, which is exactly the group most small businesses need.
• Anything sent to your top accounts. Not because the output is bad. Because the cost of an odd note to the client who funds a fifth of your revenue is not worth the twenty minutes you saved.
• Brand voice at the start. These systems learn from what you have already sent. If your existing library is mediocre, autonomous execution will produce more mediocrity, faster and more consistently.

The failure mode people are not talking about

The obvious worry is a bad send. The real risk is quieter: agents optimise against the metric they were given, and marketing metrics are easy to satisfy in destructive ways. A system told to maximise revenue per send will find that emailing your most engaged segment more often works, right up until the list burns out. That damage shows up in month four, not week one, and by then it is expensive to reverse.

So the guardrail that matters is not a content review. It is a frequency cap and a suppression rule that the agent cannot override, set before you turn anything on.

When this is not worth it

If you send fewer than four campaigns a month, autonomous execution solves a problem you do not have. The setup, the guardrail configuration and the review cycle will cost more hours than the sending currently does. The same applies if your list is under a few thousand people, where the data is too thin for the optimisation to beat your own judgement.

It also depends heavily on how clean your customer data is. Retailers keep discovering that the blocker is not the agent, it is that purchase history, service tickets and email engagement live in three systems that disagree with each other. An agent on top of contradictory data is a faster way to be wrong.

If the underlying concept is still fuzzy, our plain guide to agentic AI marketing covers what these systems are before you get to whether you want one.

A reasonable first move

Pick one flow you already trust, cap the frequency, and let an agent run it for six weeks against your current version. Not because the test will be scientific, it will not be. Because the argument in your business will be settled by somebody watching the numbers rather than by whoever read the most vendor material. Give it one flow, one metric and a hard stop date, and you will know more than the case studies can tell you.

Editor’s note: This area changes quickly, so check the latest platform policy before making compliance decisions.

The post Marketing Agents Can Run the Campaign Now. The Question Is What You Hand Over. appeared first on Mark8ng.com.

The tool was not lying. It was doing exactly what it was built to do with a data set far too small to support the conclusion drawn from it.

Attribution software is having a moment again, rebranded around AI, and the pitch is reasonable: last-click is obviously wrong, so replace it with something smarter. The first half of that is true. The second half depends on numbers most small and mid-sized teams do not have.

What last-click actually gets wrong

B2B buyers commonly touch a brand somewhere between eight and twelve times before they convert. Last-click hands 100% of the credit to whatever happened immediately before the form submission, which is very often a direct visit or a branded search. Both of those are effects, not causes. Somebody typed your name into Google because of something that happened earlier.

So last-click systematically overvalues the bottom of the funnel and undervalues everything that created demand. If you optimise against it for long enough, you end up cutting the things that made the branded searches happen, then watching the branded searches decline.

That failure is real and well documented. It is the reason attribution vendors exist.

What AI attribution does differently

Data-driven attribution, which is what most of these tools are running under the branding, assigns credit by comparing paths that converted with paths that did not. Google’s own version is built into Ads and GA4. Google’s GA4 attribution documentation describes it as measuring how each touchpoint changes the estimated probability of a key event, using signals like timing between the interaction and the conversion.

That is a genuinely better question to ask than “what was last?”. And when it works, it works well.

The condition is volume. The model learns from patterns across many converting and non-converting paths. Vendor guidance commonly points at something in the region of 3,000 monthly conversions before a data-driven model is producing output you should act on. Treat that as a rule of thumb rather than a hard line, because it varies by how many channels you run and how varied your paths are. The direction is what matters: fewer conversions, less reliable output.

Here is the uncomfortable part. Below that volume the model does not refuse to answer. It produces a number, formatted identically, with the same confident decimal places. Nothing in the interface tells you that it is essentially interpolating.

The 2026 problem that no attribution model solves

There is a newer issue that makes all of this harder, and it is the one worth understanding before you spend anything.

A growing share of commercial research now resolves inside AI answers. Someone asks ChatGPT or Google’s AI Mode for a shortlist, reads a summary that mentions you, does not click, and arrives at your site three days later by typing your name. That is the mechanic behind getting cited in AI answers, and it is invisible to every measurement system you own. Your analytics records direct traffic. Your attribution model, AI-powered or otherwise, distributes credit among the touchpoints it can see, and that particular touchpoint produced no click, no referrer and no session.

So the model does what it is designed to do: it hands the credit to something else. Usually whatever paid channel was running at the time.

This is the takeaway most vendor content will not give you. AI attribution redistributes credit among the channels you already measure. It does not discover channels you do not. If your largest blind spot is discovery happening inside AI answers, a more sophisticated model does not shrink that blind spot. It makes it less visible, by confidently allocating the missing credit to something you can see. The dashboard gets cleaner while the picture gets less accurate.

What to do instead at small volume

If you are under a few hundred conversions a month, four cheap methods beat any model.

Ask people. A required “How did you hear about us?” field on your form, free text rather than a dropdown, is crude, biased toward recency, and still the single most informative thing most small teams can add. When someone types “saw you mentioned in ChatGPT”, no attribution product on the market would have told you that.

Run holdouts. Turn a channel off completely for three or four weeks and watch total pipeline, not that channel’s reported conversions. It is the only method that measures incrementality rather than correlation. It costs you real volume during the test, which is why almost nobody does it.

Test by geography. If you operate in several regions, run a channel in some and not others. Slower than a holdout, less painful.

Change one thing at a time. Unglamorous, and it works when you have too few conversions for anything statistical.

What can go wrong

The failure mode is not that the software is bad. It is that a confident number invites a decisive action, and the decisiveness is the problem.

Specific risks worth naming. Attribution tools need clean tracking across your site, ad platforms and CRM, and that integration work is usually underestimated by weeks, not days. Consent rates directly limit what can be tracked at all, so in markets with high opt-out rates you are modelling on a partial sample before the model even starts. And if the tool becomes the shared source of truth in a company, it starts settling internal budget arguments, which means the errors in it become organisational decisions rather than analytical ones.

This is not worth buying if you run a single acquisition channel, if your monthly conversion count is in the dozens, or if your sales cycle is long enough that the paths finishing today reflect marketing you ran a year ago. In that last case the model is accurately describing a strategy you have already stopped running.

A more useful starting point

Before evaluating any attribution product, work out what you would actually change if the answer came back differently. If the honest response is that you would keep running all your current channels either way, the tool is buying you comfort, not decisions.

If there is a real budget question at stake, start by adding the self-reported field to your forms and running one four-week holdout on the channel you are least sure about. Do that for a quarter. You will know more about your marketing than any dashboard would have told you, and you will be in a far better position to judge whether the software is worth its licence fee.

The post AI Attribution Promises to Fix Last-Click. Read the Fine Print. appeared first on Mark8ng.com.

For most people using AI in marketing today, this changes nothing. If your team uses ChatGPT, Jasper, Copy.ai, or any consumer-facing AI writing tool, those products have their own pricing. Your subscription did not get cheaper overnight. API pricing is what developers and technical teams pay to build directly on the model, it does not flow through to consumer products automatically.

But for teams building custom AI workflows, or thinking about it, an 80% price drop on the budget tier is worth actually doing the math on.

What the numbers mean in practice

At the old Luna pricing ($1 input, $6 output per million tokens), running 10,000 short social media posts through the API, roughly 100 tokens in, 150 tokens out per post, would cost about $1 in input and $9 in output. Around $10 total.

At the new pricing, the same run costs roughly $0.20 in input and $1.80 in output. Under $2 total.

That is not a scenario where a big team saves money on one workflow. That is a scenario where a small business or solo founder can run an automated content pipeline, email personalisation system, or product description generator at a cost that is essentially negligible.

The workflows that were economically borderline before, where API cost was close to the value generated, are now clearly viable. The ones that were already profitable just got significantly more so.

What is still expensive

Sol, OpenAI’s best-performing model, did not move. It is still $5 per million input tokens and $30 per million output tokens. For tasks that genuinely require Sol’s reasoning capability, you are paying the same as before.

This matters because the price cuts apply to models that handle simpler, higher-volume tasks well, not the one you would use for complex strategic analysis or nuanced long-form writing. Test the task on Luna first, but do not assume the cheapest model will do every job. The cost difference between Luna and Sol is significant enough that getting the choice wrong at scale adds up fast.

The other thing worth noting: OpenAI’s engineering team cited a 20% reduction in serving costs thanks to infrastructure improvements, including using Sol to rewrite their own GPU kernels. The Luna cut goes far beyond a cost pass-through, it is a strategic price decision to keep developers on OpenAI’s platform as Chinese providers and Google compete aggressively on price.

Which marketing workflows benefit most

The tasks where Luna-level models typically perform well enough include: generating product descriptions at scale, first-draft social captions, bulk email subject line testing, FAQ generation from existing documentation, and translation of standard marketing copy.

The tasks where you generally still need a better model: anything requiring nuanced brand voice, complex reasoning about strategy, sensitive communications, and long-form content where quality needs to match strong human writing.

A useful approach for teams evaluating this: pick one high-volume task you currently do manually or with an expensive tool, run 50 examples through Luna at the new pricing, and score the output quality against your standard. The economics make testing easy now, you can run a meaningful sample for less than a dollar.

The broader picture

API pricing has been falling consistently across providers. A year ago, the compute cost for processing a million tokens was several times what it is today across all models. The ceiling on what is economically viable to automate keeps rising.

The question for marketing teams is not really about price anymore. It is about which tasks are actually worth automating, which require human judgment that AI reliably gets wrong, and how to build a workflow that handles both. The decision to build on the API instead of using off-the-shelf tools is still primarily about control and capability, price is now rarely the limiting factor.

If you want to understand which AI marketing workflows are worth building custom versus buying as a tool, that is the kind of practical systems question mark8ng.ai is built to help with.

The post OpenAI Cut GPT-5.6 Luna Pricing by 80%. Here Is What That Actually Changes. appeared first on Mark8ng.com.