Is focus on GEO worth it? What the research says for and against

“So what do you guys do for AI search?”

“How do we optimise for AI?”

“We want to show up on ChatGPT!”

“We need SEO and GEO.”

Australian business owners and C-suites know the concept of GEO/AI SEO/LLM SEO/AEO and often add it to their enquiries when reaching out for SEO services, but how we respond as an agency or consultant varies widely! 

I believe that GEO is worth understanding and rarely worth buying as a separate service. The research supports adjusting your SEO for AI surfaces, and the case for a standalone GEO retainer is still thin. In the US and global data available, most of the AI search affecting businesses is Google’s own: AI Overviews and AI Mode. The work that wins there is good SEO with a few sharper habits.

If you’re a marketing leader, you’ve probably been pitched GEO in the last six months. Maybe it came from an agency, maybe from your board. Marie Haynes recently called GEO a distraction, using GA4 data from 100 websites. Her data is useful, and so I wanted to build an article using this dataset as a foundation for the GEO conversation. I collected the strongest evidence on both sides, rated how much weight each piece can carry, and split it by industry where the data allows.

What GEO actually is

Generative engine optimisation (GEO) is the practice of improving how often a brand or page is retrieved, cited or named in AI-generated answers, such as ChatGPT, Gemini, Perplexity, and Google’s AI Overviews and AI Mode.

The term comes from a 2023 research paper by Aggarwal and colleagues at Princeton and IIT Delhi, presented at KDD 2024. You’ll also hear AEO (answer engine optimisation), LLM SEO and AI SEO. They overlap so heavily that you can treat them as one idea.

Google’s position on the matter has remained relatively consistent. Danny Sullivan, Google’s Search Liaison, told WordCamp US in August 2025 that good SEO is good GEO, and the same goes for AEO and every other acronym. John Mueller has said the label matters less than how your audience actually behaves, and that you should judge it with real usage data.

Take any platform’s guidance with some caution. Google has an interest in how SEOs behave, and past insights into its systems have been gamed, so it shares selectively. That’s why this article weighs Google’s statements alongside independent data.

How we’re rating the evidence for or against GEO

• High: peer-reviewed research, or a large primary dataset with a published method.
• Medium: a large independent or vendor dataset with a partly disclosed method.
• Low: a vendor claim with no published method, or a small or unpublished sample.

All figures are attributed to their source. Unless labelled otherwise, they are US or global data, not Australian.

The case for GEO

Claim	Evidence	Source	Strength
Content changes can shift how AI answers use a page	Citing sources, adding quotations and adding statistics lifted visibility 30 to 40% on a research benchmark	Aggarwal et al., KDD 2024	High, but in a lab simulation
AI visitors convert well	ChatGPT visitors converted at 13.8% vs 3.02% for Google organic (55 sites tracking conversions)	Marie Haynes, Oct 2026	Low: unpublished dataset, small conversion counts
Clicks are shrinking anyway	US zero-click searches rose from 60.45% (2024) to 68.01% (early 2026)	SparkToro/Similarweb via Search Engine Land	Medium
AI shortlists concentrate choice	74% of users picked the first-ranked item in an AI Mode study, and users chose a trusted brand when one appeared	Growth Memo user study	Medium (sample size not published)
Most AI categories are still unclaimed	89.3% of estimated AI demand sits in categories with no clear brand owner. Owners kept first place in 90.4% of month-on-month comparisons	Growth Memo, Semrush data, ChatGPT only	Medium
Off-site signals carry most of the weight	Third-party signals drive an estimated 85% of brand visibility in LLMs	AirOps	Low (method not stated)

The pro-GEO argument we find most persuasive comes from the behaviour data. When an AI tool hands someone a shortlist, they tend to take the top option or the brand they already trust. That makes being named more valuable than being cited, and it rewards brands that are already well known.

The conversion data looks encouraging, but the volumes are small. In Haynes’ dataset, Google organic drove 964 conversions to 36 from all LLMs combined. A high conversion rate on a small base tells you AI visitors arrive with intent. Whether AI justifies a separate budget needs larger volumes to answer.

The ownership finding is the one we’d act on. Most categories have no clear owner in AI answers yet, and in the ChatGPT data, owners mostly held first place month to month. That’s a reason to start early on a narrow category you can own.

The case against GEO

Claim	Evidence	Source	Strength
LLM traffic is small	ChatGPT sessions were 0.62% of Google organic sessions. All LLMs combined: 0.73% (100 sites, 145M+ sessions, 12 months of GA4)	Marie Haynes, Oct 2026	Medium (large but unpublished dataset)
Google still drives the outcomes	Google organic drove 96.5% of conversions across the channels analysed, 27x all LLMs combined	Marie Haynes, Oct 2026	Medium (unpublished; the Google share rests on 964 conversions, the LLM side on 36)
No technique has proven durable	A review of 45 GEO studies (2023 to 2026) found no technique with a stable, long-term, cross-platform causal effect on discoverability or downstream behaviour. It found that competition can erode individual gains	Martinez, 2026, arXiv	Medium (systematic review, preprint, not yet peer-reviewed)
The headline “40%” is narrow	The same review notes the 40% figure measures a relative gain in a simulator where the source is already one of five documents in context	Martinez, 2026	Medium (preprint)
ChatGPT referrals have stalled	Growth in ChatGPT sessions to websites ended around the GPT-5 release in August 2025 and has declined since	Marie Haynes, Oct 2026	Medium
“AI visibility” scores are noisy	AI citations and brand mentions correlated at -0.229. The most-cited domain was also the most-mentioned only 20.8% of the time	Growth Memo, Semrush data (1,094 US categories)	Medium

The 2026 review finds that GEO tactics can change how AI uses a page that has already been retrieved. The gap is evidence of those gains holding across platforms and over time. The original paper’s “up to 40%” describes a relative gain in a five-document simulation, which is much narrower than the promise of ranking in ChatGPT it’s often used for.

The measurement point matters just as much. Citations (your page is linked as a source) and mentions (your brand is named in the answer) barely move together. Any report that rolls them into one “AI visibility score” blurs two different signals.

Where both sides agree: Google’s AI is the AI search that matters

Advocates and sceptics agree on one thing: Google’s AI Overviews are taking clicks. The figures vary because each study measures something different:

1. Ahrefs: AI Overviews were linked to a 34.5% lower click-through rate for the top-ranking page in its April 2025 study. Its updated analysis of around 300,000 keywords (December 2025 data) puts the drop at 58%. Ahrefs notes this is correlational.
2. Seer Interactive: organic CTR on queries with an AI Overview fell to 1.3% in December 2025, then recovered to 2.4% in February 2026. Queries without an AI Overview sat at 3.8%.
3. Pew Research Center: users clicked a traditional result on 8% of visits when an AI summary appeared, against 15% without one. They clicked a link inside the summary on 1% of visits (900 US adults, March 2025 searches, reported July 2025).

So the AI search costing businesses clicks today is mostly Google’s. ChatGPT sends a fraction of a per cent of the traffic Google does. AI Overviews sit on top of the channel that still drives most of your leads.

We’re seeing the shift in our own Search Console too. 5twelve.com.au now gets long, conversational queries that read like prompts, for example, “can ai-generated content rank in search for ecommerce? (i’m located in australia)”. That points to AI Mode-style searching showing up inside Google, with questions getting longer and more specific. In our experience, pages that answer a specific situation clearly are the ones that pick these up.

Which industries are most affected by AI Overviews?

Industry	What the data shows	Source	What it means
Healthcare	AI Overviews on 51% to 88% of health queries, depending on the study and date	WebFX (130,000+ queries); BrightEdge (Dec 2025, as reported)	High exposure, especially for informational traffic
E-commerce	Contested: AI Overviews on 85% of queries with commercial modifiers in one panel, but around 13 to 14% of shopping keywords in another	Semrush (400-keyword panel) vs Semrush shopping analysis	Depends on query type. Check your own product and category terms
Local services	AI Overviews on around 68% of local business-type queries	Whitespark (2025)	High exposure on research queries. The map pack still carries “near me” intent
Legal	No dedicated study found. Brands in legal earn citations from informational content, but those citations don’t turn into “recommend a lawyer” mentions	Growth Memo topical focus study	Depth on core practice areas matters more than breadth of blog content
Real estate	Brands gain from citation breadth, such as content on adjacent suburbs	Growth Memo topical focus study	Area-level content can earn citations. Medium-low confidence

As you can tell, we read a lot of Growth Memo at 5 Twelve!

One thing to flag, though, is that this table is missing Australia. Almost all of the industry-split data is American, and much of it comes from vendors with something to sell. Our sources here tried to remain true to thought leadership in our industry without any paid product affiliation.

How people actually shop with AI

People now use AI at several points on the way to a purchase. Most of that use is research. Very little of it is the purchase itself. AI-assisted shopping falls into four broad modes:

1. Search engines with AI built in

Google’s AI Overviews and AI Mode answer the question on the results page, and Google remains where most purchase journeys start and finish. Pew found people clicked a result 8% of the time when an AI summary appeared, compared with 15% without one.

2. Generative engines

ChatGPT, Gemini, Perplexity and Copilot are used to shortlist, compare and ask follow-up questions. OpenAI’s own research puts conversations about “purchasable products” at about 2% of ChatGPT use. Measure Protocol’s broader tracking found 21.6% of ChatGPT interactions showed some commercial intent, with 7.1% showing high purchase intent (US, October 2025).

3. Marketplace assistants

These assistants live inside the store, such as Amazon’s Alexa for Shopping and Walmart’s Sparky. They have the biggest audiences of any AI shopping tool because shoppers are already there to buy.

4. AI agents

These act on your behalf: they browse, fill in forms and check out. They get the most attention but see the least usage, which the next section covers.

Australian surveys show the same split between research and buying:

• Manhattan Associates (August 2026, 500 consumers): 83% had used an AI assistant such as ChatGPT, Gemini or Copilot for shopping research, and 36% do so regularly.
• Commerce and PayPal (2026, 1,000 online shoppers): 63% were interested in trying an AI shopping agent, but only 4% wanted AI help at the point of payment.

But the traffic hasn’t followed the research. Across Marie Haynes’ 100-site dataset, all AI tools combined sent 0.73% of the traffic Google organic did (US/global). Across ten US Shopify stores, AI assistants drove 0.12% of revenue from January to August 2026 (Ethercycle). Shoppers ask AI, then check what it says. Product.ai found nine in ten shoppers who use AI for research verify its answers before buying, usually on Google, review sites and the brand’s own website.

Services follow the same path. Someone might ask ChatGPT to shortlist family lawyers or buyer’s agents on the Gold Coast, then search the names on Google, read the reviews and enquire through the website. AI shapes the shortlist, and search, reviews, and your site still win the enquiry.

So what’s real and what’s hype when it comes to AI agents/agentic commerce?

In search, agents are AI tools that act on your behalf to compare, shortlist and buy (also known as agentic commerce). The space is moving fast, and the promises are big, so here’s what has actually shipped as of October 2026, and where:

Shipped

Google has shipped more of the full journey than anyone else. Its version runs through Merchant Center and platforms like Shopify, so most retailers can take part without rebuilding their site.

US-only – Google’s Universal Commerce Protocol (January 2026)

Google released the Universal Commerce Protocol with Shopify, Etsy, Wayfair, Target and Walmart as co-developers. More than 20 payment partners have endorsed it, including Visa, Mastercard, Stripe, American Express and Adyen. It’s an open standard that lets an AI agent handle the whole purchase, from discovery through to checkout. Checkout inside AI Mode and the Gemini app runs on Google Pay. 

Google says the agent asks permission and only buys after you confirm the purchase and shipping details. Google also announced a Business Agent that lets shoppers chat directly with a brand in Search.

US-only – Google’s Universal Cart (October 2026)

At Google I/O in May, Google announced a single cart that follows you across Search and Gemini. It launched in the US on 15 October 2026 (later than the summer timing they were aiming for). The cart saves items from multiple retailers in one place, tracks price drops, finds promo codes, and suggests payment methods. You can check out through Google Pay or move the items to the retailer’s own site. Google named Nike, Target, Walmart and many Shopify merchants among the stores that support it. YouTube and Gmail were previewed as coming later, and Google still hasn’t given them a date. Google also said the protocol will extend beyond retail, starting with hotel bookings and local food delivery. Alongside it sits AP2, a payments framework that lets agents buy within limits you set. Google hasn’t said when AP2 will be widely available.

At I/O, Google said its checkout built on the protocol will expand to Canada and Australia “in the coming months”, then the UK. Reported timing has varied since, and some coverage now points to next year for Merchant Center support here. The October US launch came with no confirmed international timeline. Even so, this agent capability is most likely to reach Australian shoppers first.

US-only – Amazon’s Alexa for Shopping (May 2026)

Amazon integrated its Rufus shopping experience into Alexa+ in May 2026 to create ‘Alexa for Shopping’. As of October 2026, U.S. customers can access it on the Amazon Shopping app and website and on Echo Show devices, allowing them to browse and shop the full Amazon store using voice, touch, or both.

U.S. customers who use Alexa for Shopping spend 40% more per order on average compared to shoppers who don’t use the AI assistant, making this shipped agentic commerce the strongest usage signal any platform has published.

Amazon is currently in court trying to keep Perplexity’s agent off Amazon.com. At the same time, its own Buy for Me agent completes purchases on other retailers’ websites without the shopper leaving the Amazon app. Amazon says tens of millions of products are eligible. Merchants are included by default and have to email Amazon to opt out. Some have complained about listings they never agreed to, wrong product images and orders for out-of-stock items. Amazon wants agents shopping everywhere, as long as they’re Amazon’s.

Currently, there is no news of a rollout to other countries or regions.

US-only – Perplexity’s Comet browser Instant Buy (March 2026)

Perplexity made its Comet browser free on desktop and mobile in March 2026. Comet can carry out multi-step tasks across tabs, and US users can buy in chat through Instant Buy with PayPal. 

Comet can handle the entire commerce journey, including product discovery, price comparison across multiple retailers, and autonomous checkout using PayPal. It remains US-only with no timelines available for other countries.

Amazon is suing Comet because it was caught shopping on Amazon. Although the court found that the Comet user, not Perplexity, accessed Amazon, the lawsuit continues, and the court said its ruling sets no broader rule for AI agents.

US-only – Microsoft’s Copilot Checkout (January 2026)

Microsoft launched Copilot Checkout on Copilot.com in January 2026, with payments through PayPal, Shopify and Stripe, and brought it to the Copilot mobile app in the US by April.

Microsoft claims to have a catalogue of 500,000 merchants, but has not shared any data on whether people are actually buying through Copilot. 

US, Canada and Japan only – Meta’s Muse (September 2026)

Meta launched Muse in September 2026, a personal agent that can book, fill in forms and buy. It pays with one-time cards through Stripe’s Link wallet. Muse is live in the US and Canada and has reportedly reached Japan, with no date announced for Australia. 

When users prompt the Muse agent to find an item they’re interested in, the agent browses retail sites and then offers suggested listings in the Muse app and on the website. Clicking a listing will either open the retail website directly or open in the Muse app, which will feature a button to check out with your agent.

Early reviews are mixed, as initial setup is time-intensive and can be complicated. Assisted checkouts where the agent prepares your cart on the merchant’s site is only available for partner merchants, including Walmart, Gap, Sephora, Wayfair and Best Buy.

Stalled

In January 2025, Sam Altman wrote that “in 2025, we may see the first AI agents ‘join the workforce’ and materially change the output of companies.” OpenAI has shipped a lot of agent features since then. Several have been scaled back or retired, and its newest attempt starts over on merchants’ own websites.

OpenAI’s Instant Checkout (scaled back March 2026)

OpenAI launched Instant Checkout in ChatGPT in late 2025, built on its Agentic Commerce Protocol with Stripe. In March 2026, it scaled back the feature after low merchant uptake, and purchases moved to retailer apps instead. Forrester analyst Emily Pfeiffer estimated around 30 participating Shopify merchants as of February 2026, according to trade coverage. The protocol now mostly powers product discovery, and shoppers complete purchases on the retailer’s own site.

ChatGPT Atlas (retired August 2026)

ChatGPT Atlas, OpenAI’s AI browser, launched in October 2025. It offered an agent mode for Plus and Pro subscribers to perform browser tasks like filling out forms or shopping. OpenAI announced its retirement in July 2026, and the browser agent features moved into the ChatGPT desktop app and a Chrome extension.

dots (early, from September 2026)

OpenAI reportedly began rolling out its next attempt in late September 2026 to Pro, Business Premium, and Enterprise users. Each dot runs its own cloud browser and shops on the retailer’s own website. It buys with a card you’ve already saved there, and only after you approve the purchase. For now, ChatGPT shopping for most users stays conversational: it recommends, and you buy elsewhere.

Anthropic – no consumer checkout yet

Anthropic has taken a different route. Its Claude for Chrome extension can browse, compare and fill in forms, and it asks for your approval before any purchase. Its commerce work so far is aimed at businesses. In September 2026, Anthropic released shopper and merchant agent templates for retailers to build on, with Shopify, Visa, Mastercard and Accenture among the early users. It offers no consumer checkout.

What is holding agents back?

1. They still fail a lot

The Stanford AI Index 2026 reports agent success on OSWorld, a standard benchmark of computer tasks, rising from about 12% to about 66%. That’s fast progress, and it still means roughly one task in three fails. Most people won’t hand over their credit card on those odds.

2. Security isn’t solved

Prompt injection is when a web page hides instructions that hijack the agent reading it. OpenAI’s own security chief has called it “a frontier, unsolved security problem”. In December 2025, Gartner advised companies to block AI browsers for the foreseeable future, and that guidance is still widely cited.

3. Shoppers don’t trust agents with the final click yet

Accenture surveyed 25,590 consumers across 16 countries (June 2026). Only 9% were open to an agent making a fully autonomous purchase, including payment. 32% would let an agent choose a product as long as they approved the payment. Riskified’s Q1 2026 survey of about 2,000 US and UK shoppers found 55% weren’t comfortable letting an agent complete a purchase. Shoppers are happy to research with AI and still want to make the final purchase themselves.

4. Who gets to let agents in is now a political fight

Every platform wants agents to shop, but each one wants to decide which agents. Amazon shows this most clearly. Since August 2025, it has used its robots.txt file to block AI agents from OpenAI, Anthropic, Meta and Google, including Google’s shopping agent. Meta’s Muse has no user agent for robots.txt, to name one, so Amazon blocked it through its Conditions of Use instead. Meanwhile, Amazon’s own Buy for Me agent shops on other retailers’ websites by default.

Amazon also took Perplexity to court over Comet shopping on Amazon, as mentioned above. And Perplexity has its own history: in 2025, Cloudflare accused it of disguising its crawlers to get past sites that had blocked it.

The payments industry is working on a fix. Visa’s Trusted Agent Protocol, built with Cloudflare, lets an agent prove its identity so websites can distinguish a genuine shopping agent from a scraper. Mastercard and American Express are working on similar systems. It’s still early days, though.

5. Your website might be the weak link

Agents can only work with what they can read on your site, and they fill the gaps from somewhere else. Growth Memo tested how agents answered questions about 100 B2B software products. For integration and security questions, agents answered from the vendor’s own website 92 to 93% of the time. For pricing, that dropped to 79%. When the price wasn’t on the site, 45% of runs pulled it from a third party instead: a review site, a reseller or an old forum post.

Part of the problem is how agents read a page. Many AI crawlers fetch the raw HTML and don’t run JavaScript. A 2024 Vercel analysis found that the major AI crawlers, including OpenAI’s and Anthropic’s, didn’t render it. Anything that loads with scripts may be invisible to them, such as prices, service details and tabbed content. The same goes for information that only appears in a PDF, a pop-up or an image. Whatever the agent can’t read, it either leaves out, takes from someone else’s site or guesses.

For a service business, the fix is mostly good SEO housekeeping.

What we’d do instead of buying “GEO”

This is how we approach it with clients. Each step supports your SEO and your AI visibility at the same time, and all of it fits inside an SEO strategy that treats AI surfaces as part of search.

1. Fix measurement first

Define a key event, such as an enquiry, a booking or a sale, and track it properly. Haynes found that 45 of her 100 sites weren’t tracking any key event at all, and this is extremely common across local businesses we meet. Without one, nobody can tell you whether AI traffic is worth anything, and every GEO pitch becomes a matter of faith.

What good looks like: key events are firing in GA4, and you can see them by source. AI referrals (chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com) are grouped into their own channel so they don’t get lost in “referral”.

Note that GA4 attribution isn’t great either. Going down to server logs with a tool like https://salience.com/ is much more reliable for attributing which AI crawlers and agents are visiting your site.

2. Report in three layers

• Exposure: mentions, citations and share of voice, tracked separately for each AI platform, because each one cites different sources
• Behaviour: branded search, direct traffic and AI referrals
• Outcome: leads and revenue

Each layer answers a different question. Exposure without outcomes is a vanity metric. Outcomes without exposure leave you guessing why things moved. 

Most AI influence will show up as branded search and direct visits, because people research with AI and then return to Google.

3. Win Google first, including AI Overviews and AI Mode

Based on available data, Google is where most clicks are lost and where most leads still come from. It’s also where AI shopping journeys end: when people shortlist with AI, they then verify on Google, review sites, and your website before they buy or enquire. Google’s agent checkout is also the one most likely to reach Australia first.

It is great if your money pages rank for the commercial queries that matter, and your pages answer the questions behind them clearly enough to be quoted in an AI Overview.

4. Earn third-party mentions, and keep them consistent

Use digital PR, reviews, directories and industry communities. What others say about you appears to carry real weight with AI tools, though the data here is early. When your own site doesn’t answer a question, agents look elsewhere: in the Growth Memo test, 45% of runs pulled missing pricing from a third party. So consistency matters as much as volume.

Your name, services, service areas and pricing approach should completely match across Google Business Profile, directories, review platforms and your site.

5. Publish assets worth citing

That means original data or benchmarks, clear definitions, a visible methodology, and URLs that never change. AI tools can easily replace generic content, because they can write the generic version themselves.

What good looks like: at least one asset in your category that only you could have published, such as a local price guide, a benchmark from your own data or a clear framework, and that others link to and quote.

6. Make your site readable and reachable for agents

Agents can only use what they can read and reach. Three checks cover most of it:

• Rendering: key information sits in the server-rendered HTML. Prices, service details and FAQs that load with JavaScript, or sit in PDFs and images, may be invisible to AI crawlers.
• Access: your robots.txt, CDN, and bot protection settings allow the AI crawlers and agents you want. Many sites block them by default without anyone deciding to. Make it a deliberate choice.
• Answers: pricing or price drivers are visible where compliance allows, and your enquiry form works for a person or an agent filling it in.

7. If you sell products, get ready for agent checkout

Google’s checkout built on the Universal Commerce Protocol is expected in Australia, though the timing is still unclear. It will work best for merchants with clean product data.

Your Merchant Center feed should be accurate and complete, your platform’s agentic commerce settings reviewed (Shopify is a co-developer of the protocol), and you’ve made a deliberate call on programs that include merchants by default, like Amazon’s Buy for Me in the US.

8. Pick a narrow category you can own

A service plus a region is a good starting unit. Most categories are still unclaimed, and in the ChatGPT data, brands that held a category mostly kept it month to month. Being the clear answer for one specific thing beats being one of many answers for a broad one.

‘Good GEO’ looks like a defined category, such as “conveyancing for off-the-plan buyers on the Gold Coast”, with the best page on the web for it and third-party mentions that reinforce it.

Good GEO is just good SEO…or is it?

GEO is a useful lens on SEO and a weak case for a separate budget. People are researching with AI, but the clicks, enquiries and sales still mostly run through Google and your own website. AI agents are real, mostly US-only and early. The platforms are still fighting over who gets to let them in.

So spend where the evidence points: winning Google, earning mentions, publishing good content, and making yourself accessible to AI bots and humans alike.

All of it compounds, no matter how AI search goes.

Although agentic commerce is still a long way from becoming common practice in Australia, GEO is still relevant to any conversation about SEO. Keen for a free strategy session? Book a call with 5 Twelve today.