Why Most "AI SEO" Doesn't Work — And What The Research Actually Says

Graham Hyndman · Visibility Australia · September 2026

There's a wave of agencies in Australia right now selling "AI SEO" and "AI visibility" services. Most of them are repackaging the same old SEO tactics with new labels. Some are selling things the evidence actively says don't work.

I know this because I spent months reading the actual research before I built anything. Not blog posts about the research. The research itself. And what it shows is uncomfortable for a lot of people selling services in this space.

Here's what the strongest available studies actually say.

The things being sold that don't work

Three tactics are still being pitched as AI visibility fixes across Australia. All three measure as negative, null, or bottom-decile in the strongest available studies.

Schema markup was tested by Ahrefs in a matched difference-in-differences study across 1,885 pages. The result: Google AI Overviews citations dropped 4.6%, ChatGPT rose 2.2%, AI Mode rose 2.4%. That's statistical noise — no meaningful effect in either direction. A related experiment found that AI systems extracted only visible HTML content and ignored JSON-LD schema entirely. The pages already had 100-plus citations each, so the test conditions were favourable. Schema still didn't move the needle.

llms.txt files were studied by Ahrefs across 137,210 domains. 97% of llms.txt files were never requested by any crawler. The idea that publishing a text file will make AI platforms notice your business has no support in the data.

GBP posts and Q&A seeding consistently rank in the bottom decile of local search ranking factors in the Whitespark 2026 survey of 47 local SEO practitioners. They're easy to automate, which is why agencies sell them. Easy to automate and effective are not the same thing.

If your current provider is selling any of these as their primary AI visibility strategy, that's worth a conversation.

What actually works — the six levers with evidence

The Forensic Audit we run at Visibility Australia is built on six levers. Each one has a specific study behind it. None of them are proprietary secrets — the research is published and anyone can read it. The difference is whether your provider has read it and built their methodology around it.

GEO content transforms have the strongest evidence in the field. A peer-reviewed study from Princeton and IIT Delhi, published at KDD 2024 (the premier data science conference), tested specific content changes across 10,000 queries. Adding quotations from credible sources lifted visibility by 41%. Adding statistics lifted it by 31%. Citing sources, 27%. These are controlled experiments with measurable effect sizes. Keyword stuffing, by contrast, measured negative.

Prompt-content alignment is the largest page-level effect measured to date. Discovered Labs analysed 2 million citation observations across 10,000 URLs, four AI engines, over six months. The finding: when a page's content semantically matches the language of the prompts customers actually use, citation rates increase significantly. The effect size (β +0.37) is three times larger than the next strongest page-level factor. This isn't about keywords — it's about matching the way real people phrase real questions.

Google Business Profile accuracy is the most underrated lever. The Whitespark 2026 Local Search Ranking Factors survey ranks primary GBP category as the number one local pack factor, ahead of proximity. Correct business hours, accurate service-area configuration, and proper category selection are all API-addressable and routinely wrong. Five of the top negative local ranking factors are GBP-related problems that most businesses don't know they have.

Review orchestration matters for both volume and recency. AI platforms weight recent reviews more heavily than old ones. A business with 200 reviews but nothing in the last 90 days looks dormant. The approach has to be ACCC-compliant — no gating (only asking happy customers), no incentivisation, no fake reviews. Ask every customer, neutrally, every time.

Citation consistency has moved into the top five factors for AI visibility specifically. When your business name, address, and phone number are inconsistent across directories, AI platforms lose confidence in your entity data. Consistent citations across major Australian directories compound over time.

Attribution and measurement is the lever nobody talks about. Search Engine Land tracked 51,000 AI Overview events and found that 22.4% of AI-driven traffic was misattributed to "Direct" rather than organic search. Most businesses are getting more AI-referred traffic than their analytics show — they just can't see it because the measurement is broken.

The uncomfortable truth about domain authority

Here's the part that most AI visibility providers leave out: domain authority matters roughly six times more than any page-level factor. That finding comes from the same Discovered Labs study. AI-perceived domain authority has a mean absolute SHAP value of 0.38, compared to 0.06 for the strongest page-level factor.

What this means in practice: a new business with no backlinks, no mentions, no directory presence, and no reviews can do perfect page-level optimisation and still not get cited. The page-level work is real and worth doing. But it's the smaller half of the equation.

The larger half — domain authority — comes from genuine third-party mentions, directory listings, reviews, and being referenced by other sites. There are no shortcuts to this. It compounds over time, which is why daily monitoring and management matters more than a one-off audit.

How to check where you actually stand

Before spending money on any AI visibility service, do this yourself. Open ChatGPT, Gemini, Perplexity, and Claude. Type in the question a customer would ask before calling a business like yours. Something like "best mortgage broker in [your suburb]" or "recommended electrician near [your area]."

Look at who gets named. If it's not you, that's not a mystery — it's a measurement. And now you know what to fix.

If you want a structured version of that test across twenty prompts, four platforms, with screenshots and a prioritised fix-it roadmap, that's what the Forensic Audit does. But the free version takes two minutes and tells you the truth.