This Week

The Current

Issue 08 28 September 2026 8 minute read Past editions
A laptop screen showing marketing analytics
Photo by ThisIsEngineering via Pexels.

If social media for your business feels like a moving target, this is for you. Plain, honest notes each week on what's actually working, from two people who run it for a living. This week YouTube offered to test your thumbnail for you, Google made a still photo into a one-tap ad on Shorts, and two different ad platforms started quietly grading how samey your creative is. No sign-up yet, just this week's issue.

Worth Knowing

Two basics this whole issue is quietly about

The first frame is doing more work than the rest of the video

Whatever image sits on top of your video, the thumbnail on YouTube or the frame someone lands on when a Reel starts, is deciding whether the rest of it gets watched at all, and I think most small businesses treat it as the last five seconds of the job rather than the first. You can spend a whole afternoon on a good video and lose it to a blurry frame of somebody mid-blink, because the video isn't competing on quality, it's competing for one second of somebody's attention against everything else on the screen. So pick the frame deliberately, before you edit if you can: one clear subject, close enough to read on a phone, and something in it that makes a person want to know what happens next.

Shoot a batch in one session, not a post at a time

The reason posting falls over for most businesses isn't motivation, it's that every single post needs a separate decision, a separate setup and a separate hour you didn't have, so one busy week takes the whole habit down with it. If instead you block out one session and shoot five or six things in it, you come out with a few weeks of material and the posting becomes admin rather than a creative act, which is a much easier thing to keep doing. I think the extra benefit people miss is that batching is also how you get genuine variety, because you're already set up, so moving to a different corner of the shop or a different distance costs you two minutes instead of another day.

YouTube will test your thumbnail for you now

On Wednesday 23 September YouTube ran Made on YouTube, its annual event where it lays out what's coming, and most of it was AI features, which is normal for 2026 and not especially interesting on its own. Two things in there are worth your time though, because they're a platform finally offering an answer to the question every one of us has been guessing at: which image, and which opening.

The first is that A/B testing is expanding from titles and thumbnails to whole videos, so you can put up more than one version and let YouTube work out which one holds people. It's going to select creators for now, with YouTube saying it'll expand globally next year, and from 2027 you'll be able to upload and test three different cuts of a Short, which in practice means three different openings of the same video. YouTube's own figure for how much of this is already happening on the smaller version of the feature is more than 40 million tests run on titles and thumbnails.

A person checking a social media grid on their phone next to a laptop
Photo by Plann via Pexels.

The second is dynamic thumbnails, where you make three options and YouTube shows each one to a different slice of your audience, and then later this year Studio starts watching how a thumbnail is performing on its own and swaps it if it isn't working. That's a small sentence with a big consequence, because it turns the thumbnail from a decision you make once, badly, at 11pm, into something that keeps getting corrected after you've stopped thinking about the video.

You have been guessing at the first two seconds for years. This is the first time the platform has offered to tell you.

I want to be straight about who can actually use this, because the headline and the reality are a fair way apart. The full video test is select creators with a global rollout described as next year, and the three-cuts-of-a-Short version is 2027, so for most people reading this the good part is an announcement rather than a button. What did start rolling out on the day is Shorts series, which lets you group Shorts into seasons and episodes with their own thumbnails and sequential playback, and there's a Gemini-powered editing assistant coming to Shorts where you describe an edit in words, trimming speech, syncing music, adding a text hook, rather than doing it by hand.

The other half of the announcement isn't about creators at all, and I think it's the part to file away. YouTube is launching Custom Feeds, where a viewer builds their own homepage feed by describing it in plain language, and YouTube's own examples are things like witty podcasts for a thirty minute commute, or help me find new DIY creators. It's going to viewers in the US later this year on web, mobile and TV. A fortnight after Australia published a draft law that would force platforms to offer people a choice about their feed, which we covered in issue 06, the biggest video platform in the world shipped a version of it voluntarily. Nobody knows yet how many people will bother building one, and that's the whole question, so I'd hold off on concluding anything about it.

Do this week

Open YouTube Studio and check whether video A/B testing has turned up for you, since it costs a minute to know. Then do the manual version regardless: pull your two best performing videos from the last three months, look only at the thumbnail and the first two seconds, and write down what they have in common. That's your thumbnail rule and your hook rule until a platform is willing to test it properly for you.

A photo can be an ad on Shorts now, and it clicks through in one tap

On Thursday 24 September Google published its September Demand Gen drop, which is the monthly list of what's changed in the campaign type that buys YouTube, Shorts, Gmail and Discover, and the useful bit is small and practical. You can upload a 9:16 full-bleed image ad, so a single still shaped like a phone screen, and on Shorts and in Gmail somebody can get to your landing page with one tap instead of the usual fiddle.

I think that matters more for a small business than it sounds, because the reason most small businesses never advertise on a video platform isn't the money, it's that they haven't got video they're confident in, or they've got three clips and no appetite for making more. A good still photograph of the actual thing you sell, taken on a phone, cropped tall, is a much lower bar to clear than a video that has to survive being watched, and it puts you on a surface you'd otherwise have skipped entirely.

Google's own number in the announcement is that advertisers who add Gmail to their Demand Gen campaigns see their image creatives get around 40 percent more conversions at the same return. I'd read that carefully rather than repeat it, because it's Google's figure about Google's placement, published to get you to switch the placement on, and there's no independent version of it I can point you at. So treat it as a reason to run a small test, not as a result you should expect to get.

There's one other thing in the drop worth noticing even though it's not for you yet. Google is putting a conversational business agent alongside YouTube video ads that run with a product feed, so a viewer can ask a question about what they're seeing and get an answer without leaving the video. That's the same shape as the sponsored agents OpenAI started testing inside ChatGPT this month, and the direction is consistent: the ad stops being a picture and becomes a conversation, and what that conversation says about you gets pulled from your own product feed and your own website. Which makes the boring work on your product data the thing that decides how you get described.

Do this week

Pick your three best photos of your product or your place, crop them to 9:16 full bleed with the important thing kept out of the bottom third where the interface sits, and put them into a small Demand Gen test with Shorts and Gmail turned on. Keep the budget at roughly what a week of coffees costs you, and judge it on cost per enquiry rather than on views, because views on a still image will look flattering and mean nothing.

Two ad platforms started grading your creative, and they're grading the same thing

On Monday 21 September Pinterest added a campaign optimisation score to its Ads Manager. It grades a campaign across three things, how it's set up, the mix of creative inside it, and signal strength, which is Pinterest's way of asking whether your tracking is actually feeding it data, and then it hands you specific recommendations about what to change.

On its own that's a modest feature on a platform plenty of readers don't use, and I wouldn't have written about it. What makes it worth a few paragraphs is that Meta did the same thing a month earlier. A creative diversity rating appeared in Meta Ads Manager on 26 August, marked in development, and it rates each ad set Low, Medium or High on how visually similar the images and videos inside it are to each other. You find it under Columns, then Customize columns, then Creative diversity.

Two different platforms, within a month of each other, decided the thing worth telling you was that your ads look too much like each other.

The thing to understand is what's being measured, because it isn't quality and it isn't quantity. Reported behaviour on Meta's version is that an ad set mixing stills, video, customer footage and carousels can still come back Low, because what's being read is structural sameness across the assets rather than which formats you ticked. Ten versions of the same shot with different text over the top is, as far as the system is concerned, one piece of creative. I should say plainly that Meta hasn't stated the rating affects who your ads get shown to, and it's still labelled in development, so the honest reading is that it's a readout of how the platform sees your work rather than a rule you're being marked against.

Why both of them landed on this at once isn't mysterious once you look at it from the platform's side. Both of them are choosing, ad by ad and person by person, which of your creatives to put in front of somebody, and to do that well they need options that are genuinely different from each other. Handing them six variations of one idea is asking them to pick between near-identical things, and the result is that one of them gets nearly all the spend and your cost creeps up as the audience gets tired of it. So this is less a new rule and more the platforms telling you out loud what they've always needed.

Do this week

Put your live ads next to each other on one screen and squint at them, then count how many genuinely different pictures there are: different place, different distance, a person or no person, daylight or lamplight. If the honest answer is one or two, then your next shoot is three setups rather than one setup and more captions. If you advertise on Meta, add the Creative diversity column while you're in there and see what it says about the ad sets you already thought were varied.

You can advertise on Threads without an Instagram account

Reported on 23 September: Meta is dropping the requirement that a Threads advertiser have a matching Instagram username, and businesses with no Instagram profile at all will be able to run ads straight from a native Threads profile. It's rolling out gradually, which is Meta's usual way of saying some accounts have it and some don't, so don't take it as missing if it isn't there.

For scale, Threads passed 500 million monthly users in June according to Meta's own newsroom, and ads went worldwide on it back in January, so this isn't a new platform so much as one being untied from the platform it grew out of.

My honest recommendation is to spend nothing here. There's no published small business benchmark for Threads ads that I can point you at, it's mostly a text platform which is an awkward fit for most local businesses, and the reason the inventory is cheap right now is precisely that nobody has proven it works. What's worth noticing isn't the opportunity, it's the signal: Meta separating Threads out commercially means Meta intends it to be a real ad platform rather than a side room of Instagram, and on past form the cheap window on a new Meta surface tends to close faster than anyone expects.

Do this week

Nothing to spend. If you already post on Threads and it's actually getting you replies from real people, it's worth ten minutes seeing whether the ad option has turned up on your account. If you don't post there, this changes nothing for you this week and you can skip it with a clear conscience.

What's actually good right now

Disclosed upfront: this newsletter is partly written with Claude, so take our AI opinions here as disclosed rather than neutral.

Best all-round assistant

Busy Tuesday. On 22 September Anthropic released Claude Opus 5.5, saying it matches Fable 5.1 on most tasks while costing about 40 percent less to run on a typical workload than Opus 5 did, with output more than 30 percent faster. The same day OpenAI released GPT-6 Sol and GPT-6 Luna, saying Sol makes roughly half as many mistakes as the model it replaces. We're still not quoting an index score, third week running, for the same reason: the public trackers we can see don't agree with each other. Claude is what we use for planning and drafting and what part of this is written with.

Best for images

Still GPT Image 2.5 and its Flare and Sunburst variants, and that's a repeat of the last two weeks, which we'd rather say than invent a new winner every seven days. Given what's in this issue about thumbnails and creative sameness, the useful half of image AI for you is still editing and cleaning up a real photo of your real thing, not generating a convincing photo of something that doesn't exist.

Best for video

The thing we flagged twice has now happened: the Sora API was removed on Thursday 24 September, on the schedule OpenAI set back in March. If a video tool you pay for has gone strange this week, that's the likely reason, and it's worth an email to them rather than a guess. We're not naming a video leader this week, because the public arenas we can see disagree about who's on top, so pick on price, on the clip length you actually need, and on whether it can do your specific shot.

Worth flagging

Both of the big labs cut their prices in the same week, Anthropic putting Opus 5.5 at 20 percent below its predecessor and OpenAI pricing Luna at a fraction of Sol. That's not directly your problem, since you pay a monthly subscription rather than for tokens, but nearly every AI tool you pay a small monthly fee for is built on top of one of these models, so the cost of the engine inside your caption writer or your editing app just dropped noticeably. Whether any of that ever shows up on your invoice is a separate question, and I wouldn't hold your breath.

We have been guessing at thumbnails too

Nearly everything in this issue is a platform offering to test something on your behalf, and the fair thing to admit is that we've been doing the guessing version ourselves. Tom and I shoot and cut everything, so the thumbnail and the opening frame get decided at the end of a long edit, quickly, by whichever of us is still looking at it, and calling that a process would be generous. So we're changing it on our own channel before we go telling a client to change theirs: three setups per shoot instead of one, and the first frame chosen before the edit rather than scraped off the end of it. Our own numbers are still thin while the case studies build out, so please take that as what we're trying rather than something we've proven.