SODP Dispatch - 6 August 2026

The Same Regulator That Counts 2,864 Australian News Outlets Would Fund Fewer Than 100

By Scott Purcell

Sydney, Australia – 3 August 2026 – Man of Many, one of Australia’s largest independent digital publishers, says the final design of the News Bargaining Incentive announced on 2 August improves the scheme in ways worth acknowledging, but leaves untouched the two design features that determine whether any money reaches the broader Australian news industry.

What the government got right

Man of Many wants to be clear about what has improved, because a good deal of it has.

The universal levy closes the loophole that made the original Code unenforceable, where a platform could simply remove news and walk away. Professional networking services have been brought into scope. The offset for deals with small and medium publishers has been lifted substantially, making those deals materially more attractive than deals with incumbents. Platforms must now strike more deals, not fewer, to discharge their liability. The definition of who counts as a journalist has been broadened to include freelancers and other essential news roles, which reflects how independent newsrooms actually operate.

On the distribution side, the government has created a grants program for the smallest publishers and start-ups that fall below the $150,000 revenue threshold, and has increased the flows going to regional journalists, to small and medium publishers, and to outlets serving communities that have historically been underserved, including First Nations, culturally and linguistically diverse, LGBTIQA+ and disability communities.

That last measure deserves particular credit. Directing more of the pool toward journalists serving communities that mainstream coverage has long neglected is exactly the sort of structural choice a public-purpose scheme should make, and the sector asked for nothing like it.

“Anyone claiming the government has not listened is not reading the same document,” said Scott Purcell, Co-Founder of Man of Many. “Most of what the independent sector asked for in May is in here, and the support for regional and diverse-community journalism goes beyond what we asked for. We should say so plainly.”

🧮 SODP Free Publisher Tools

One thing we kept hearing from publishers is that real benchmarks are locked behind a sales call, buried in a stale report, or handed to you by a vendor with a reason to make their own numbers look good. So we built three tools that hand you the number directly, free, in under two minutes.

Here are the tools and what you can do with them:

• Rewarded Ads Earnings Calculator — drop in your pageviews, niche and geography for a monthly estimate, a per-placement breakdown, and a full-year view

• EPMV Benchmark — get the typical range for sites like yours, plus the exact percentile you land in

• Monetization Leak Auditor — a fast seven-question audit flags your top leaks and what fixing each one could recover

No signup, no email wall, no display ads. The full results unlock through a single optional rewarded ad, letting you experience the reader-first model these tools are built to measure, not just read about it.

Try the tools ▸

🚀 Introducing Benchmark Grid: Independent Benchmarking for Publisher Tech

Every publisher conversation surfaced the same problem: too many platforms, too little independent data to evaluate them with. Vendor directories are pay-to-play. Review sites go stale. Feature checklists tell you what a tool does, not whether it works for publishers like yours. So we built the missing layer, with 281 platforms assessed across 22 editorial categories spanning content management, distribution, revenue, advertising and analytics.

In practice, this means you can:

• Shortlist on evidence, not sponsorship — vendors cannot pay for inclusion, ranking or score changes, ever

• Compare on a published methodology — 45% Capability, 40% Publisher Fit, 15% TCO, applied consistently across every category

• Judge fit, not just a number — tools are grouped into Recommended, Worth Considering and Specialist, with assessments written by editors who actually work with newsrooms, magazine teams and B2B media

• Build your stack, not just pick a tool — the Build My Stack tool maps the right combination of platforms to your specific setup

As AI search becomes the layer through which recommendations get discovered, publishers need an independent source feeding it, one that gives every vendor a fair shot without tilting toward any single company. That is the solution Benchmark Grid is providing.

Explore Benchmark Grid ▸

📊 Publisher Revenue Dinner Miami 2026

We are hosting a private dinner in Miami for senior revenue leaders across news, gaming, newsletter, specialist, and brand publishing. This dinner is for publishing professionals who are actively driving revenue strategy and want a candid, peer-level conversation on how to win the biggest commercial quarter of the year. If you are a Chief Revenue Officer, CEO, Programmatic Head, or Partnerships Director making decisions around yield, inventory, first-party data, and Q4 growth, this evening is designed for you.

Date: Wednesday, September 16, 2026
Time: 6:30 PM
Venue: Coral Room, Rusty Pelican Miami Restaurant, 3201 Rickenbacker Cwy, Key Biscayne, FL 33149, United States

What you will walk away with

• Practical yield, personalisation, and first-party data moves that are lifting revenue per visit right now

• A candid read on how your peers are pricing, packaging, and positioning inventory for the Q4 surge

• Insight into which new formats — video, newsletters, and creator partnerships — are actually paying in 2026

• An attendee-only summary of the evening's key insights and growth playbooks shared around the table

The dinner is high-trust, collaborative, and entirely off-the-record under Chatham House Rule. No panels, no pitches, no slides — just one focused evening on how to make every session, impression, and first-party signal work at full value in publishing's most important quarter.

Seats are strictly limited to 15 attendees.

A big thank you to Ezoic for co-hosting this evening with us.

Apply for your seat here ▸

🎯 LiveRamp × OpenAI: Measure Your ChatGPT Ad Performance

If you're running or planning ads on ChatGPT, you can now close the measurement gap. LiveRamp's Conversions API (CAPI) Hub connects directly to ChatGPT ad campaigns via secure server-to-server data connections, meaning you get reliable conversion tracking without depending on browser-based signals that often break or go missing.

In practice, this means you can:

• See which ChatGPT campaigns are actually driving conversions — not just clicks

• Optimize spend in real time based on accurate performance data

• Justify and grow your ChatGPT ad budget with measurement data you can trust

As consumer journeys shift toward AI-powered platforms, this gives marketers the infrastructure to keep up, connecting data, measuring impact, and proving ROI on one of the fastest-growing ad surfaces in the world.

Read more ▸

By Vahe Arabian

🧭 The Four Questions AI Shortlists Are Really Answering, And Why Brands Can Only Handle One

AI answers don't just mention brands. They rank them — and the ranking is decided inside sub-queries most brands have never looked at, on pages they don't own.

That's the whole argument for what publishers are still worth. Here's the mechanism, the technique, and what happens when the models start going direct to brands.

The mechanism: one prompt, then eight to twelve retrievals

Ask ChatGPT, Perplexity or Google AI Mode to shortlist a category and the prompt expands before anything renders. Google AI Mode fires roughly 9–11 parallel sub-queries. ChatGPT runs closer to 2.3–2.8 — a real difference, and one most commentary flattens into a single "8–12" range that belongs to Google.

DataForSEO ran 100,000 ChatGPT prompts and logged 100,249 fan-out queries: 47.5% of prompts expanded before an answer returned. Their sample comes from their own API prompt database rather than organic user sessions, and it's ChatGPT only — so, directional. On commercial, comparison-shaped queries, expansion is the norm.

The sub-queries aren't random. Strip the phrasing away and they resolve into the four questions a buyer is actually asking:

• Which one? — "X vs Y", "best X for [use case]", "alternatives to X"

• What does it cost? — "X pricing", "is X expensive", total cost of ownership

• What's it actually like? — "X reviews", "is X worth it", "X after a year"

• What goes wrong? — "problems with X", "X complaints", "why people leave X"

Why that hands the category to publishers

Look at what a brand can credibly serve.

It can answer what does it cost on its own domain. That's first-party fact, and models will happily take it.

It cannot answer the other three.

Which one is unanswerable from inside the comparison — no party being compared has standing to adjudicate it. What's it actually like is a brand marking its own homework, and models discount it accordingly. What goes wrong has never been honestly answered by any brand about itself, and never will be.

Those aren't gaps in brand content strategy. They're structural disqualifications.

So fan-out mechanically routes retrieval toward independent sources on precisely the questions where consideration is decided — three of the four, and the three that carry most weight once a buyer is past browsing.

That's the mechanism behind the figure everyone quotes without explaining. AirOps analysed 21,311 brand mentions across GPT-5, Claude Sonnet 4.5 and Perplexity Sonar: 85% came from third-party pages, 13.2% from owned domains. Scope caveat twice over — those three models only, and commercial-intent B2B software queries. But the direction isn't an accident of sampling. It follows from how the retrieval works.

Publishers are that 85%. Not by goodwill, by structure.

What the ranking looks like, and why mention counting misses it

Presence isn't binary. The four questions resolve into a single response naming four brands framed against each other. One is the recommendation, one a credible alternative, and the rest are named so they can be dismissed.

Being present and framed fourth is worse than being absent. It's a negative comparison delivered with the authority of a neutral recommendation, to a buyer who never opens a source.
Scope note: this is conversational shortlisting behaviour, pronounced in ChatGPT, Perplexity and AI Mode where the response has room to compare. It's much weaker in classic AI Overviews, which are shorter and often resolve to one answer.

Counting citations in the AI era is counting impressions and calling it demand.

The technique — how to work fan-out without fooling yourself

Warning first, because it's the most common mistake. Almost no published fan-out research observes real fan-outs. Surfer's headline study — 173,902 URLs across 10,000 keywords, the source of the widely repeated "161% more likely to be cited" — extracted fan-outs using Gemini as a proxy for what AI Overviews actually did, pulled only top-10 SERP data, and notes 67.82% of AI Overview citations didn't rank top 10 for anything. Surfer themselves advise against chasing fan-out queries. Every practitioner tool that "shows your fan-out queries" is running a simulation and presenting it as observation.

Second, they're unstable. Surfer's repeatability run across 1,600 repeats found only 27% of fan-out keywords recurred, 66% appeared exactly once, and 0.6% appeared in every run. A fixed fan-out query list is a plan for a search surface that won't exist next week.

Both point the same way: build against the four questions, not against query lists.

What that looks like in practice:

1. Observe rather than simulate. Run 20–30 category-defining prompts per engine and log the sources actually cited. That's real data. The sub-queries are inferred; the citations are observed.

2. Sort what you find into the four questions, not into a spreadsheet of 400 phrases. You're building four buckets.

3. Score coverage per question, per engine. ChatGPT, Perplexity and AI Mode retrieve different sources and assemble different competitive sets. Blending them hides the only thing you can act on.

4. Build one maintained asset per question, not one mega-roundup trying to serve all four. A dedicated "X vs Y" page for each meaningful pair. Cost of ownership as a standalone, including the costs the brand omits. Verdict and long-term use with stated methodology. And a real answer to what goes wrong — the highest-value question and the least served, precisely because brands can't touch it.

5. Make the methodology machine-legible. Units tested, over what period, under what conditions, who funded the work, conflicts declared, last revalidated. Structured, not implied in prose. You're asking a model to weight your verdict over a first-party claim; give it the grounds.

6. Maintain on cadence. Seer's recency work, updated July 2026 and covering ChatGPT, Perplexity and AI Overviews, holds: these systems favour recently revalidated content, strongly in fast-moving categories. An unmaintained comparison page quietly stops being retrieved.

Measuring consideration, not citations

Three stages, run per engine:

• Discovery. Does the brand surface when the category is explored at all?

• Intent. Does it survive once the buyer adds constraints — budget, region, integration, team size?

• Consideration. How is it framed against the competitors named in that same answer, and would a human reading it pick it?

The drop-off between stages is the diagnosis, and the four questions tell you the cause. Strong at discovery, weak at intent is usually thin coverage on "which one" and "what does it cost". Strong at discovery, weak at consideration means "what's it actually like" and "what goes wrong" are being answered by sources that frame the brand badly — and no amount of on-site publishing reaches that.

And where the uplift actually lands

Not in referral clicks.

Profound tracked 2M+ conversations from January to June 2026. Once an assistant names a brand, site visits run about 1.5x the forecast baseline for ChatGPT and 2.5x for Gemini, sustained for seven days. 20.5% of first visits land within an hour, 42.0% within 24 hours, most after day one.

Attribution doesn't follow. Even after ChatGPT's 7 May 2026 change made links more clickable, Profound's framing is that more than 97% of those visits still arrived with no UTM. It's observational panel data, US-only, and a site-visit study rather than a conversion study.

If you report AI performance from referral traffic, you're reporting roughly 3% of it. Branded search impressions, direct and unattributed sessions against a pre-mention baseline, and assisted conversions are the measurement layer — for brands, and for publishers watching their own category authority compound.

One operational note on the seven-day window: a mention lands, demand spikes for a week, and if the brand SERP is uncontrolled or the landing experience is weak, the lift leaks.

The real threat: when models go direct to brands

This is coming, and partly here. Structured product feeds, pricing APIs, spec databases, merchant catalogues. If a model can get authoritative product data straight from the manufacturer, why route through a publisher?

For "what does it cost", increasingly it won't. And more broadly, anything descriptive is commodity and already being absorbed, such as spec tables, feature lists, restated pricing, "what is X" explainers, roundups assembled from vendors' own marketing copy. That work has a short remaining life, and publishers still resourcing it are funding their own displacement.

But go back to the four questions. The direct pipe serves one.

Brands supply attributes. They cannot supply judgment. Feed a model twenty brand sources, and you get twenty claims to be the best. There is no first-party answer to "which of these is actually better for this use case." Nor failure data - what breaks, what support is like at eighteen months, what ownership costs past the promotional pricing, why people leave. No brand publishes this about itself, ever. Nor category definition - who belongs in the comparison set is an editorial judgment, and a brand feed can't draw a boundary that excludes the brand supplying it.

The direct pipe doesn't replace publishers. It commoditises the one question publishers should already be exiting, and raises the value of the three they own.

But it also changes who's competing for those three. Once the direct pipe absorbs the descriptive layer, every publisher left standing is fighting over the same "which one," "what's it actually like," and "what goes wrong", and coverage alone stops being the differentiator, because everyone's coverage will look the same. What separates one publisher's answer from another's is how narrowly the comparison set is drawn, how far the failure and cost-of-ownership data goes beyond what a press release would already say, and how visibly the methodology behind the verdict is stated.

That's also why models keep an independent citation in the loop at all. A recommendation built purely on vendor-supplied data is one the model stands behind alone. Pointing at independent evaluation makes the answer defensible, but only if that evaluation gives the model something concrete to point to. That's structural, not sentimental, and it's the difference between being cited and being the one cited over every other publisher saying the same thing.

What this makes you worth to brand partners

Not placement. Selling favourable framing erodes the precise signal that makes you citable — slowly, then all at once, and it won't show in a dashboard until the citations are gone. It also can't be done at scale without a competitor noticing and the whole category losing its independent source.

The defensible products:

Diagnostic and advisory. Run the fan-out for the category and show the brand its coverage question by question, engine by engine: which of the four it's absent from, which sources are answering them instead, where it's framed against a peer group it doesn't belong in, and at what stage it drops out. That's consultancy value drawn from category authority, and it's honest because every fix is on their side.

Category insight the brand can't see. You observe the whole category's buying journey, not one brand's slice. Complaint patterns, price sensitivity, what buyers ask before converting, how the comparison set shifts. That's a research product.

Being the standard worth measuring against. The healthiest version and the oldest. If you're the source models trust for a category, performance in your independent evaluation becomes a real KPI, and brands improve the product to win it. That relationship long predates AI. AI raises its value, because the verdict is now read by the machine assembling the shortlist.

Licensing your evaluation data — to the model providers, not to brands. Structured, verifiable, independent evaluation gets scarcer as descriptive content commoditises. That negotiating position strengthens as brand-supplied data grows more abundant, not weaker.

On the terminology

Worth settling, since it comes up every time.

"Query fan-out" is Google's own language, from patent US20240289407A1, Search with Stateful Chat, describing synthetic query generation inside AI Mode. Applied to ChatGPT it's loose — different architecture, roughly a quarter the sub-query volume. Purists are technically right, and contesting the label is still a waste; it's already the term clients, vendors and journalists use.

Use it. Be precise about which engine you mean, and name the proxy when you cite the research. That's where the credibility actually sits.

Same discipline on the correlation data. Ahrefs — Google AI Overviews only — put branded mentions at 0.664, referring domains at 0.295, backlinks at 0.218, but describe all of them as moderate to very weak, and 26% of the 75,000 brands had zero AIO mentions and were excluded. Direction, not magnitude.

The window

Adobe surveyed 500+ marketers in April 2026: 98% had no documented AI search roadmap, 74% reported no measurable strategy. Semrush found 45% of marketing leaders can't accurately measure AI visibility.

Brands will work this out on their own timeline. The question for publishers is whether you've moved off "what does it cost" and onto the three you own before the commoditisation finishes — because the negotiating position you'll hold in two years is being set by what you invest in now.

What's in your archive that a brand could never publish about itself? That's the part that survives.

A different way to think about video on publisher platforms | INMA

For years, publishers have invested heavily in bringing video onto their own Web sites and apps. Yet, despite producing better journalism and increasingly sophisticated video products, most consumer behaviour still points in the opposite direction: Consumers want video, just not on our platforms. The reason may be simpler than we think. Consumers increasingly choose formats first, not brands first. When someone wants to watch short-form video, they instinctively open TikTok, YouTube Shorts, or Instagram Reels. When they want long-form entertainment, they open Netflix. For audio, they open Spotify or Apple Podcasts. Increasingly, when they want answers, they open ChatGPT or Gemini.

Read more ▸

A record-breaking eight Pulitzer awardees disclosed AI use this year | NiemanLab

A translation of a mass shooter’s cryptic journal in the days after an attack. A public records review that revealed failures to install flood warning systems in Central Texas. An exposé of American technology companies’ complicity in building the Chinese surveillance state. An audit of the SEC’s crypto lawsuits that showed weakening enforcement under the second Trump administration. On May 4, the Pulitzer Prizes recognized these stories among the winners and finalists across 15 journalism categories. The reporters behind each of these stories also disclosed using AI to the judging committee.

Read more ▸

Meta puts AI ahead of support tickets for developers by mid-September | PPC Land

Meta yesterday rolled out a revamped developer support experience at developers.facebook.com/support built around Meta AI Developer Assistant, according to a post published on the Meta for Developers blog under the byline of Zoë Lieberman. The assistant is described as the primary support path at that address, replacing the sequence in which a developer searched through documentation pages or filed a ticket and waited for a reply. The mechanics are simple to state and consequential in effect.

Read more ▸

Approaching Google Zero: As search referrals plunge, news publishers must anticipate they’ll never rebound | Editor & Publisher Magazine

“Google Zero” arrived for Phillip Swan, and it killed his successful online business. The term, coined by The Verge’s Nilay Patel, describes a future web where Google stops sending traffic to publishers. Increasingly, it feels like that time is nearing for publishers across the news industry. It certainly arrived for Swan. A journalist covering the television industry since the early 1990s, Swan launched TV Answer Man in 2017 to offer simple and clear guidance to viewers about emerging technologies, such as 4K and streaming.

Read more ▸

Disney+ looks to TikTok creators to bring fan content to its short-form video feed | TechCrunch

Disney is partnering with TikTok to bring Disney-focused fan content directly into the Disney+ app. The companies are starting with a pilot program in the U.S. in the coming months, and plan to expand to additional markets later on. As part of the agreement, fan-made videos on TikTok about Pixar, Marvel, Star Wars, and other franchises will be featured in the “Verts” section of Disney+, the streamer’s short-form video feed that rolled out a few months ago.

Read more ▸

AI can knock on the door. Your CRM marketing platform is still the house. | Search Engine Land

As AI assistants like ChatGPT and Claude plug directly into marketing systems, a false claim is spreading with them: that AI like this will soon make the CRM marketing platform unnecessary. While we run a platform company, this isn’t about preferring the platform. It’s about what each piece can and can’t do, which is easiest to see if you’re precise about who does what. The AI assistant is a visitor. It shows up, and it knocks. The connection that lets it in has a name, the Model Context Protocol, or MCP. Think of it as the door.

Read more ▸