AI answer systems read the visible text of a page and skip content that exists only in structured data, an image or a PDF. In an Ahrefs test, five major AI systems ignored JSON-LD, hidden Microdata and hidden RDFa entirely. For Australian B2B marketers that makes getting cited a placement and format job, and it is testable in an afternoon with a browser and a blank document.
The standard advice for earning an AI citation has been to mark the page up properly. Add the schema, describe the entity, let the machine parse the structure. Sensible, widely repeated, and aimed at a different machine than the one doing the citing.
Ahrefs put five major AI systems in front of pages carrying JSON-LD, hidden Microdata and hidden RDFa. All five ignored the markup and read the visible HTML instead. Which means your pricing, your regional benchmark and your own survey result are citable only if they exist as text a reader could highlight with a mouse. For a growth-stage B2B team, that turns a year of speculation about AI visibility into a format problem you can check today.
The citation question finally has a mechanical answer
Australian sites are already being read at volume, whatever your AI strategy document says. One Australian business directory logged roughly 1.8 million automated crawler requests from search engines and AI platforms across its network in 30 days, according to SmartCompany's report on AI recommendations covering Perth agency Search Scope's September 2026 study. Google also switched AI Mode on in Australia, in English, back in October 2025. This is not a US preview you have a year to prepare for.
Prevalence on the queries that matter to you is high. Australian B2B services SERPs returned a Google AI Overview on 47.8% of commercial-intent queries in Digital Nomads HQ's analysis of 116,918 SERPs across 18 industries and six cities, with the B2B services cut covering 6,495 SERPs. Digital Nomads HQ is an Australian SEO agency and the report is a marketing asset, so attribute it by name. Prevalence figures also swing entirely with the query set, which is why the intent cut matters more than any headline number.
The bigger shift is that ranking and being cited have become largely separate populations of pages. A 2026 study of 11,500 queries found similarity scores of just 11% to 18% between the sources surfaced by Google Search, AI Overviews and Gemini, reported second hand in that same SmartCompany piece without the underlying research house named. Directional, then. It still lands the point that an extractability review is a different job from an SEO audit, and a sharper cut than ranking breadth against citation rate on its own.
The click cost is better measured in the northern hemisphere. UK searchers ended 69.5% of Google searches without a click between January and April 2026, the highest of the six countries on SparkToro's Similarweb panel, which excludes the Google mobile app and so probably understates it. France gives the cleanest before and after: across 963 domains tracked around the 22 July 2026 AI Overviews launch, median click loss reached 8.8% while impressions held steady or rose, with a 23.1% median CTR decline in nine days for the most exposed domains. Ahrefs is a vendor publishing research on its own category, and nine days reads as an immediate post-launch effect rather than a settled state. Australia appears in neither dataset.
What an AI system actually sees, and what it silently skips
Start with what the platforms say about themselves. Google says its AI search features still rely on its core search systems, and Microsoft describes Bing's index as evolving "from a list of pages into a source of facts" that AI systems can use to build answers, both quoted in Ahrefs' guide to AI search optimisation. A source of facts is a phrase worth sitting with. It describes a system that lifts statements, not pages.
Which is where the markup finding bites. In the same guide, five major AI systems ignored JSON-LD, hidden Microdata and hidden RDFa and relied on visible HTML. The conclusion is narrower than "schema is dead", and more useful: Ahrefs' own recommendation is that important information such as pricing or product features should appear as text on the page as well as in structured data.
Australia has a live demonstration of what legible text buys you. The plain Australian business directory in that Search Scope study appeared 199 times across 2,400 AI responses covering 20 industries, third most observed after only Google and Reddit, ahead of Word of Mouth, ProductReview and hipages, with separate June testing by Victorian agency Dossis Digital across 12 trades on four AI platforms pointing the same way. The operator's prescription is unglamorous. Make sites easy for machines to understand, with clear information about services, pricing and previous work.
Now the awkward part for most B2B sites, which is that images are where facts go to disappear. The 2026 WebAIM Million found 16.2% of home page images across the top one million home pages carried no alternative text at all, on an average of 66.6 images per page. The HTTP Archive's 2025 accessibility chapter puts around 13% to 14% of images with no alt attribute and another 30% with an empty one. That chapter also states it covers HTML only and excludes PDFs and other office documents, which neatly explains why nobody can tell you how much B2B research is locked inside a download.
What the data does carry is that the download is standard practice. E-books and white papers were used by 51% of B2B marketers in the year to mid-2024, based on 980 B2B respondents in CMI and MarketingProfs' benchmark research, and white papers are the single most uploaded format on NetLine's syndication platform, which recorded 7.2 million content registrations in 2025. NetLine is a B2B content-syndication vendor describing its own supply, so read that as platform behaviour rather than a measure of the open web.
Put those together and the pattern looks like production order rather than negligence. Regional pricing, a local benchmark, an implementation timeline: those artefacts get built for a sales deck or a gated report first and reach the website second, if at all. They are also precisely the facts an AI answer about your category would most want to quote. That is the gap the structural self-assessment could describe in June without being able to measure it.
The first third of the page is doing most of the work
Placement matters almost as much as presence. Kevin Indig's research, cited in the Ahrefs guide, found 44.2% of AI citations came from the first 30% of a page, and that cited passages carried more named entities than typical prose while leaning on definitive language rather than hedging. NP Digital arrives at the same practical conclusion from a different direction, recommending that important original statistics surface within roughly the first third of educational and commercial pages. Its reading draws on a Search Engine Land analysis of AI citations, and Search Engine Land is owned by Semrush, which is worth knowing when you weigh it as a neutral authority.
Chunking is the mechanism behind both findings. AI search engines split pages into smaller sections, retrieve the ones relevant to a query, and cite only the passages that support an answer, which is why Ahrefs argues for atomic content and BLUF, Bottom Line Up Front. A passage needing three paragraphs of run-up to make sense does not survive being lifted out of context. Treat it as information architecture for AI search operating at paragraph scale.
Ahrefs' own 2024 cost-of-SEO survey is the worked example, and it was never built for this. It remains one of the most cited pages on its topic despite newer competing research, and the key insights shown in its graphics were also stated in the article text, so nothing had to be extracted from an image.
Currency does the rest of the work. Among the "best X" list articles ChatGPT cited, 57.1% had been updated since first publication and 79.1% were last updated within the study year of 2025, while URLs cited by AI assistants ran on average 25.7% fresher than URLs appearing in Google's organic results. A page that goes quietly stale stops being the source.
The one-afternoon extractability audit for a B2B site
Let's turn all of that into something you can run before your next content meeting. This sequence is ours rather than any vendor's, and it needs nothing beyond a browser and a blank document.
Think of the plain-text paste as your page's shadow. Whatever casts no shadow there does not exist for the system doing the citing.
1. Load the page with JavaScript disabled. Anything that vanishes was never guaranteed to be read in the first place.
2. Select the visible text and paste it into a plain text document. No styling, no images, no markup.
3. Walk your load-bearing facts one by one. The price, the benchmark, the sample size, the implementation timeline. Mark each one that survived the paste.
4. For the survivors, note where they sit. Anything material sitting below the first third of the document is a placement problem rather than a writing problem.
Four passes, one afternoon, and the output is a list instead of an opinion. Most teams find the same shape: the marketing copy survives beautifully and the proof does not.
There is some evidence the payoff is real, though it is thin locally. One Australian agency's tracking of 240 commercial queries found that pages cited inside an AI Overview recovered roughly 60% of the click-through rate lost on those queries, with average CTR on positions one through three down 18% where an AI Overview appeared. Small sample, single agency, undisclosed method, so treat it as directional only. It points the same way as everything else here. Being inside the answer is worth considerably more than sitting next to it, and rendering, accessibility and page structure are web build decisions as much as editorial ones.
Editorial hygiene, with a consequence it never had before
None of this asks you to write robotic copy for machines. Ahrefs lands in the same place, describing the work as making the answer easy to find, keeping claims precise, putting important information in visible text, and making sure key passages still read sensibly on their own. That is a description of good editing. The labels around it stay contested, and Google's public line has been that good SEO is largely good GEO, so you do not need to settle an acronym argument before doing the work.
For some organisations the work is already mandated. UK public sector bodies must meet WCAG 2.2 AA under the 2018 accessibility regulations and the Equality Act 2010, and that standard's first success criterion is a text alternative for non-text content. For those teams, the extractability fix and the accessibility obligation are the same afternoon's work. Australian organisations carrying equivalent accessibility expectations sit in a similar position, though the local obligation varies by entity type and is worth confirming with your own compliance advice rather than assuming.
Regulation is moving toward more visible attribution, not less. Under a conduct requirement the CMA imposed on Google, UK publishers get effective tools to prevent their content being used to power AI features such as AI Overviews, Google is required to make sure publisher content is properly attributed using clear links in AI-generated results, and publishers may opt out of their content being used for fine-tuning. Google has nine months to implement, with compliance reports every six months for the first year. The UK is legislating the citation. You still have to be citable to receive one.
Your best facts are almost certainly already written. They are sitting in a deck, a chart or a gated PDF, and moving them into visible text near the top of a page is the cheapest AI visibility work available to a growth-stage marketing team this quarter. It is the same job as earning AI recommendations from buyers, and it belongs in the standing content production rhythm rather than in a special project nobody schedules.
Key takeaways
- Run the JavaScript-off paste test on your five most commercially important pages before commissioning any new content. Whatever does not survive the paste cannot be cited.
- Move load-bearing numbers out of charts, decks and gated PDFs and into visible body text: pricing, benchmarks, sample sizes, implementation timelines. Keep the chart as illustration rather than as the container.
- Put the finding in the first third of the page. Kevin Indig's research cited by Ahrefs found 44.2% of AI citations came from the first 30% of a page, and NP Digital recommends the same placement for original statistics.
- Keep facts current and consistent wherever they appear. Among "best X" articles ChatGPT cited, 79.1% had been last updated within the study year of 2025.
- Treat every prevalence and CTR figure as directional. This measurement is contested and moves with the query set, so instrument your own pages rather than importing someone else's percentage into a board deck.
Frequently asked questions
What is SEO content writing in the age of AI answers?
SEO content writing now covers where a fact sits on the page and whether a machine can read it, alongside how the sentence reads. In an Ahrefs test, five major AI systems ignored JSON-LD, hidden Microdata and hidden RDFa and read visible HTML instead, so a claim that exists only in structured data, an image or a PDF cannot be lifted into an answer.
Does schema markup still matter if AI systems ignore it?
Yes, but not as a substitute for visible text. Ahrefs' test found five AI systems relied on visible HTML rather than JSON-LD, hidden Microdata or hidden RDFa, and its recommendation is that important information such as pricing or product features should also appear as text on the page, not that structured data should be removed.
Where should original statistics go on a page?
Early. Kevin Indig's research cited by Ahrefs found 44.2% of AI citations came from the first 30% of a page, and NP Digital recommends surfacing important original statistics within roughly the first third of educational and commercial pages.
Are AI answers affecting search traffic in Australia and the UK?
In the UK, measurably. SparkToro's Similarweb panel recorded 69.5% of UK Google searches ending without a click between January and April 2026, the highest of six countries measured. No equivalent Australia-segmented study has been published, though Google's AI Mode has been live in Australia since October 2025 and Digital Nomads HQ found AI Overviews on 47.8% of commercial-intent B2B services queries in its analysis of 116,918 Australian SERPs.
Start with what a machine can already read
Before you commission another word, find out which of your own facts a model can actually lift and where they sit on the page. See where you stand in search and AI, and what to do about it. Start a conversation with Amina.
