This Week

The Current

Issue 06 14 September 2026 7 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 is about a draft Australian law that would let people switch the algorithm off, and what happens to a customer's trust the moment they work out a post was made by a machine. No sign-up yet, just this week's issue.

Correction

In issue 04 we put Meta's settlement with the US state attorneys general at up to $18 billion, which we took from news coverage rather than from the source. The attorneys general put it at up to $17.1 billion, with about $12.2 billion of that paid over ten years and the rest conditional on other things happening, so $17.1 billion is the figure to use and we'll be going to the primary release first from here on.

Worth Knowing

Two things that keep working while everything else changes

What other people say about you does more work than anything you say about yourself

A business telling people it's good is the least persuasive sentence on the internet, because everyone expects you to say that, so it barely registers. A customer saying it, in their own slightly awkward words, with their real name on it, lands completely differently, and I think that's the single most underused asset a small business has sitting there already. So I'd treat reviews and customer photos as content rather than as admin, which means asking for them properly at the moment someone is happiest, screenshotting the good ones and posting them, and reposting the customer who filmed your product without being asked. It costs nothing, it's the one thing a bigger competitor can't just outspend you on, and it's getting more valuable rather than less as more of what people scroll past turns out to be made by a machine.

Turning up every week beats turning up brilliantly once

Most small businesses I talk to aren't failing at social because their posts aren't good enough, they're failing because there's a great week, then three quiet ones, then a guilty burst, and the algorithm and the customer both read that the same way, which is that there's nothing here worth checking back for. A modest post every week for three months will do more for you than one excellent post a quarter, partly because you get better at making them, and partly because people need to see you a few times before they think of you as a business that exists rather than a thing that appeared in their feed once. So I'd pick a number you can actually hit on your worst week, not your best one, and protect that number, because the plan you can keep is worth more than the plan you'd like to have.

Australia wants to let people turn your algorithm off

On Tuesday 8 September the federal government released draft Digital Duty of Care legislation for targeted consultation, and buried in it is the part that should matter most to anyone who relies on getting in front of people who don't already follow them. It's called My Feed, My Way, and it would require platforms to notify their users, new and existing, and ask them a straight question: do you want the feed the algorithm builds for you, or do you only want to see the accounts you actually chose to follow.

The rest of the package puts a general duty on digital services to provide a safe online environment, sets minimum standards for the features and tools they use, expands what the eSafety Commissioner can do about harmful content, and carries penalties of up to $109.2 million for platforms that don't comply. It's a draft out for targeted consultation, with the government saying it intends to bring a bill to Parliament this year, so none of it is law and none of it has a date you can put in a calendar yet.

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

Being recommended to a stranger has always been the platform's decision. This would make it the stranger's.

Three days later, on 11 September, a separate piece of the same push actually became law. The Online Safety Amendment for strengthening enforcement of the social media minimum age got assent, which doubles the maximum penalty for breaching the under-16 rules to $99 million and gives eSafety the power to demand documents from platforms showing what they've actually done to keep under-16s off, and to ask third parties like age assurance providers and app stores as well. So the enforcement half of this is real now, and the feed half is still a draft.

Here's why I think the feed part matters to a small business more than the headline suggests. Nearly everything you get for free on social right now comes from being recommended to people who've never heard of you, and that decision has always belonged to the platform. This would hand a version of it to the person scrolling. If a meaningful number of Australians pick the follow-only feed, then for those people you simply stop appearing unless they followed you, and every follower you already have gets quietly more valuable.

What nobody can tell you is how many will pick it, and I'd be sceptical of anyone who claims to know. Defaults usually win by a mile when the setting is buried, but this wouldn't be buried, it would be a prompt put in front of you, and prompted choices behave differently. Platforms also have a long history of complying with the letter of something like this in the least useful way available, so I'd wait and watch what actually ships rather than what's announced.

Do this week

Nothing changes yet, so don't rebuild anything. What I would do is start treating followers and email addresses as the thing you're building, since neither depends on a recommendation you don't control, so ask people to follow you rather than assuming a good video does it on its own, and if you've got any way at all of collecting an email or a phone number for bookings, use it this week.

Half of people say they'd buy less from you if they knew AI made the post

Clutch published research on 9 September, from a survey of 601 consumers run in August, and the headline number is that 53 percent say they're less likely to buy from a brand they know uses AI in its social media content. Eighty-six percent said they've come across AI-generated content, which is really just confirmation that this isn't a niche concern any more, and 64 percent said they'd trust a brand or creator less if a product had been misrepresented with AI content.

Seen it
86%
Trust less
64%
Buy less
53%
Educational
42%

Share of consumers who have encountered AI-generated content, who would trust a brand less if a product were misrepresented with it, who are less likely to buy from brands using AI in social content, and who view AI positively for educational content. Clutch, survey of 601 consumers conducted August 2026, published 9 September 2026.

The more interesting number is the last one, because 42 percent said they see AI positively when it's used for educational content, which is the opposite direction to the other three. Clutch's own analyst, Anna Peck, put it as consumers not necessarily being opposed to AI, but wanting to understand when it's being used and what it's changing, and that reading fits the data better than a flat "people hate AI" does.

It isn't the AI people are punishing. It's finding out.

I should be straight about the limits here, because we ran a piece back in issue 02 on research from DoubleVerify pointing the other way, that people don't dislike AI ads so much as bad ads, and I still think that's largely right. This is a survey of what 601 people say they'd do, not a record of what anyone actually did at the checkout, and stated preference and real behaviour come apart all the time, especially on questions where one answer sounds better than the other. Six hundred people is also a small sample to hang a strategy on.

Where I think the two findings meet is disclosure. Nobody in this survey is describing a boycott of AI, they're describing what happens when they feel they've been shown something that isn't real, and that's a much narrower and much more avoidable problem. An AI-made background or a tidied-up graphic is a production choice. An AI image of a product that doesn't look like that in the box, or a face that reads as a customer and isn't one, is a claim about reality, and that's the one that costs you.

Do this week

Go back through your last month of posts and find anything where AI was used to show the actual thing a customer is paying for, so the food, the room, the finished haircut, the product in a hand, and reshoot it on a phone this week even if the real one looks worse. Everything else, the backgrounds, the captions, the graphics, leave alone, because that isn't the part people are reacting to.

OpenAI's new image model is the one everything is about to look like

OpenAI released GPT Image 2.5 on 8 September, calling it its new state of the art image model, and it's rolling out inside ChatGPT as well as through two API versions, one called Flare that's built for speed and one called Sunburst that's built for precise editing. OpenAI's own description of what's better is more natural lighting and richer texture, better preservation of a subject from a reference photo you give it, more reliable follow-through when you ask for changes across several turns, and generation up to 50 percent faster than the previous version.

On Artificial Analysis's text to image arena, which ranks models by blind human preference votes, the two new variants are sitting first and second this week, ahead of the previous GPT Image 2 and Microsoft's MAI-Image-2.6, which was second in our last issue. The gaps at the top are small enough that I wouldn't treat the order as settled, and it's worth remembering that arena scores measure which picture people prefer the look of, not whether it sold anything.

The bit that's genuinely useful for a small business is the reference photo handling rather than the raw quality. Generating a beautiful invented product shot has never been the problem, and after the piece above it's arguably the last thing you want. Starting from a real photo of your real thing and asking for a cleaner background, better light or a different crop is a completely different job, it keeps the subject honest, and that's the direction these updates keep improving in.

Do this week

Take one ordinary photo of something you actually sell, on your actual phone, and see how far you can push it with edits rather than starting from a written description. You'll get something you can post that's still a picture of the real thing, and you'll find out fast whether this is worth any of your time, which is cheaper than reading another list of features.

That audience number in your ads manager was never a headcount

On 10 September, in among coverage of Reddit having its strongest month of user growth this year, there was a second thing worth noticing. Analysts at Piper Sandler expect a change to the way Reddit's Ads Manager forecasts audiences to cut its reported audience-size estimates by roughly 20 percent from September. I want to be careful here, because that's an analyst expectation reported in financial coverage rather than something Reddit has announced in its own words, so treat the 20 percent as a forecast and not a fact.

What changesThe modelNot the people
What to trustReach estimateCost per result

Piper Sandler expects Reddit's Ads Manager forecasting change to reduce reported audience-size estimates by about 20 percent from September 2026, per financial coverage of Reddit's August user growth published 10 September 2026. Reddit has not published its own figure for the change.

Even as a forecast, it's a useful reminder of something that costs small advertisers real money. The estimated audience size that every ad platform shows you while you're building a campaign, the one that moves as you widen the radius or add an interest, is not a count of people. It's the output of a model estimating who could theoretically be reached, and when the platform changes the model, that number moves without a single person joining or leaving.

A number that can fall by a fifth because someone changed a formula is not a measure of opportunity.

So the mistake I'd watch for is using that number to decide anything. It's genuinely common to see someone widen a local campaign from five kilometres to twenty-five because the estimated reach looked too small, and end up paying to show ads to people who will never drive that far. The size of a theoretical audience tells you nothing about whether any of them wants what you sell, and the only number that answers that is what it cost you to get an actual result.

The other half of this, and the reason I'm not writing Reddit off over it, is that the underlying platform is doing fine. Reddit's audience grew 18 percent year on year in August, faster than it had been, which is the opposite of the story a shrinking audience estimate would appear to tell. Which is more or less the point: the dashboard number and the real world were never the same thing, and this week they're about to visibly disagree.

Do this week

Open whichever ad account you actually spend in, find your cost per result for the last thirty days, and write it on something you'll see again. That's your number, and from now on it's the one that decides whether a campaign is working, not how many people a platform reckons it could theoretically show your ad to.

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

Still Claude Fable 5.1 or GPT-6 Astra, and I'm not quoting an index score this week, because the public trackers we can see are reporting different leaders on different numbers and we'd rather say that than pick the one that suits us. Claude is what we use for planning and drafting and what part of this newsletter is written with. If you're already paying for one of them, that's your answer.

Best for images

GPT Image 2.5, released on 8 September, with its Flare and Sunburst variants now sitting first and second on Artificial Analysis's text to image arena, ahead of GPT Image 2 and Microsoft's MAI-Image-2.6. There's more on it above. The short version is that it's the best default right now, and the editing side is more useful to you than the generating side.

Best for video

Nothing changed at the top this week that should change what you pick, which is the second week running we've said that, and I think that's worth saying out loud rather than inventing a new winner. Choose on price, on how long a clip you actually need, and on whether the thing can do the specific shot, because none of the rankings will answer that for you.

Worth flagging

The Sora API shuts down on 24 September, which is now ten days away. We flagged this last week and we're flagging it again because the deadline is close enough to hurt now: if you pay for a video tool that quietly generates its clips through Sora, and a lot of smaller ones do, ask whoever you pay what happens on the 24th before you find out on a day you needed a video.

We use AI images on this very site, and we'll keep telling you where

Since we've just run a piece on what happens when customers find out AI made something, it'd be a bit rich not to say where we stand. We do use AI imagery, on this site and in client work, and we're not apologetic about it, because for a small brand it's the difference between having a decent visual and having nothing at all. What we don't do is use it for the thing someone is actually buying, so the food, the room, the product, the person, all of that gets shot properly by us, because the moment an image makes a promise the real thing can't keep, you've spent trust you can't buy back. On our own numbers we're still light, since the case studies are being built out of two years of Dr Crack Wax content and aren't finished yet, so when we tell you something works we'll keep being clear about which part is our data and which part is somebody else's.