This month, we learned some real numbers about AI search instead of relying on guesses, including how often ads actually show up and which platforms cite you when they do. OpenAI stayed busy too, adding a stack of features to its ad platform and putting Codex on every desktop for free. Here are the five most impactful stories for online advertisers in July.
1. Ads Now Appear on Nearly 30% of AI Mode Commercial Queries, But Paying Doesn’t Buy a Citation
Source: SE Ranking | July 15, 2026
SE Ranking looked at just over 50,000 commercial keywords and found ads inside Google’s AI Mode answers on nearly 30% of them. The more a keyword costs, the more likely an ad is to appear. Ads showed up on about a quarter of the cheap keywords and on more than half of the terms running $10 and up. It swings hard by industry, too, from 72% of pet searches down to under 3% in healthcare. But what I found most interesting was about citations. When an advertiser’s ad appeared in an AI Mode answer, that same advertiser’s site was named as a source only 11.5% of the time, and their page ranked organically for the query in just 2.3% of cases. Buying the ad slot and getting cited in the answer turn out to be two completely different things.
| JumpFly Takeaway There’s been a quiet assumption that buying your way into AI Mode would help you show up in the answer itself, that the two systems somehow talk to each other. They don’t. Paying for the slot buys the slot, and nothing else comes with it. What that means practically is you need two separate plans: a paid strategy for the placement and a content strategy for the citation. Neither one covers for the other, just as it has always been. The CPC correlation is worth noting. If your keywords are already expensive, ads are showing on more than half your AI Mode queries right now, whether or not anyone on your team has thought about it. So that is a win, assuming you are leveraging AI Max or Performance Max. You do not need to do anything special to qualify for this placement. I am still holding out hope that we will get to see this breakout in the future: traditional search placement vs AI Mode/AI Overview placement. |
2. Semrush Studied 126 Million AI Prompts: ChatGPT Cites 15 Sources per Answer, Gemini Cites 3
Source: Semrush | June 26, 2026
Semrush released an expanded AI Visibility Index built on 126 million US AI search prompts collected between January and April 2026, covering ChatGPT, Gemini, Google AI Mode, and AI Overviews across more than 1,200 brands in 22 verticals. The headline finding is how differently these platforms behave when they answer: ChatGPT averages 15.4 sources per response while Gemini averages just 3.3. Only 36 brands held top-100 visibility across all four platforms every month of the study, a list dominated by YouTube, Google, Reddit, Amazon, Facebook, and Apple. Overlap between being mentioned and being cited ranges from 64% on AI Overviews down to 30% on Gemini, meaning a brand can be discussed in an answer without ever being linked in it. Semrush also found that 45% of marketing leaders cannot accurately measure their brand’s visibility in AI answers, and only 9% can track every metric that matters.
| JumpFly Takeaway The big five-to-one gap in how often ChatGPT and Gemini cite sources reveals a simple truth: there is no “one-size-fits-all” approach to winning in AI search. When a platform like ChatGPT pulls 15 sources into an answer, your goal is visibility. You need to show up in as many credible places as possible. But when Gemini pulls only three, you’re in a high-stakes authority game where the most trusted voices win. If you try to use the same strategy for both, you’re likely setting yourself up to fail. The data backs this up: companies that treat SEO and AI visibility as one unified effort see an 81% success rate, compared to just 36% for those that keep them separate. As I have stated in my past posts, SEO is the engine that drives AI visibility (GEO). If you have two different teams chasing these goals, you’re paying twice for work that should be a single, coordinated mission. |
3. ChatGPT Images 2.0 in Practice: On-Brand Creative, Sized for Every Platform
Source: OpenAI | April 21, 2026
OpenAI released ChatGPT Images 2.0 in late April, and it’s the first image model that reasons before it generates. It can search the web for real-time reference, produce up to eight images from a single prompt, and check its own output before handing it back. For advertisers, two capabilities matter more than the rest. The model handles any aspect ratio from 3:1 ultra-wide to 1:3 ultra-tall natively, at resolutions up to 3,840 pixels on the long edge, with no post-processing and no manual cropping. And the edit endpoint accepts multiple reference images in one call, so you can feed it several real product shots and get back a composed scene without masking anything. Text rendering improved substantially, including non-Latin scripts, which means ad mockups come back with actual headlines instead of placeholder text. It’s available to every ChatGPT user and through the API as gpt-image-2, costing roughly half a cent to 21 cents per image, depending on quality.
| JumpFly Takeaway I am a big fan of Images 2.0. Its results are consistent, and it follows instructions much better. Exactly what OpenAI said it would do. From personal use to professional workflow, I’ve been using this one heavily. It definitely solves a problem that has quietly eaten more hours than anyone wants to admit: making one idea work across every platform’s specs. A square for the feed, a 9:16 for Stories, a wide banner for display, a 4:5 for Meta. That used to be a designer’s afternoon or a compromise where you crop the hero image badly and hope nobody notices. Images 2.0 generates the ratio natively instead of cropping into it, so the composition is built for the frame rather than squeezed into it. And I do not like the look of a squeeze. Feeding it your real product photography as reference is what makes it usable for actual brand work. The biggest unlock has been taking a produced image and having it resized. In fact, there are preset settings available to have it resized to 1:1, 9:16, 4:5, 16:9, etc. Very rarely does any content actually change, and it truly does resize in a matter of minutes. This saves me a few emails and saves the designer time to resize the images. But as always, review everything before it ships. |
4. OpenAI Ad Platform Updates – The Weekly Norm
Source: Email received from OpenAI on Jul 24, 2026
OpenAI pushed eight updates to ChatGPT Ads in a single release, announced directly to advertisers by email. Advertisers can now build conversion-optimized campaigns by selecting a Conversions objective, which shifts delivery toward clicks more likely to convert while still billing on valid clicks. Daily budgets are moving from a fixed daily cap to an average daily budget calculated across a rolling seven days, with automatic pacing throughout the day. The release also added geographic exclusions, Automatic Advanced Matching that improves web conversion measurement using hashed customer information, a bulk API for creating and updating campaigns and ad groups asynchronously, and a refreshed product feed card showing price and star ratings.
| JumpFly Takeaway Back in January, when ads in ChatGPT were first announced, we wrote that the platforms that deliver results are the ones with mature targeting, measurement, and optimization capabilities, and that ChatGPT ads wouldn’t have them at launch. The advice then was to be patient, watch what ad formats emerged, and be ready to test. This is the release worth testing on. Look at the list again: conversion bidding, average daily budgets, spend pacing, geo exclusions, enhanced conversions, a bulk API, and shopping cards with star ratings. That’s Google Ads. OpenAI has rebuilt the platform every advertiser already knows how to use and has done it in about seven months. Two things are worth acting on this week. The budget change happens automatically with no action required, so expect daily spend to fluctuate where it used to sit flat. Automatic Advanced Matching does not turn itself on, so go to Tools, then Conversions, then Data Source, and select “Enable for all Web data sources.” Better conversion matching is the difference between a channel you can optimize and one you’re guessing at. Being patient was the right call in January. And I think being patient even now with more capabilities is still the right call. This ad platform is so young that we are all still learning what it takes to be successful here. |
5. Codex Moved to Every Desktop, and the Skill File Is What Makes It Useful for Ad Copy
Source: 9to5Mac | July 9, 2026
On July 9, OpenAI released GPT-5.6 and folded Codex into a unified ChatGPT desktop app, free on Mac and Windows to every tier, including the free plan. Existing Codex installations updated into it automatically. The app added Computer Use, which lets it work across your local files, applications, and websites, along with a built-in browser and scheduled tasks. The piece that matters most for marketing work is skills. OpenAI documents them as a SKILL.md manifest plus optional scripts, references, and assets, saved in an .agents/skills folder, alongside AGENTS.md files that hold persistent rules the agent follows every time. OpenAI’s own framing is that this delivers deterministic reliability and consistent routing. The model handles interpretation and judgment, while the saved instructions handle everything that shouldn’t vary from one run to the next.
| JumpFly Takeaway The interesting part isn’t that Codex writes code. It’s that a saved instruction file turns a general-purpose model into something that reliably produces work you can actually use. The most useful one I have built is a simple ad copy skill that knows the formats, character limits, what works in a responsive search ad headline versus a Performance Max asset, and the conventions each platform expects. It isn’t complicated to build, and that’s rather the point. The difference between that and a one-off chat is consistency. In a chat, you’re re-explaining the same context every time, and the output shifts depending on how you happen to phrase things in the moment. With the skill file, the constraints are already in place, so you get on-brand, format-correct variations on the first pass instead of the fourth. This is the same argument we made back in February: if every advertiser is running the same AI tools with the same default settings, differentiation has to come from somewhere else. Skill files are one of the clearest places it can come from. The model everyone has access to is identical. What you’ve taught it about your brand and your formats isn’t. And if I don’t mention this often, review the work before you ship! |
Looking Ahead
There’s a lot happening, but the useful takeaways this month are pretty simple: Ads and citations in AI Mode don’t come as a package, so you need a plan for each. Every AI platform picks its sources differently, which means one visibility strategy won’t work everywhere. And the creative tools got fast enough that production isn’t the reason a campaign sits stuck anymore.
None of that changes the fundamentals. Clean data, good content, and someone actually reviewing the work before it goes live are still what separate the accounts that do well from the ones that don’t.
Questions about how these AI trends affect your campaigns? Let’s talk.
