Answer engine optimisation for retailers: winning the zero-click shelf

The zero-click shelf arrived late for retail — and most AEO advice was written for someone else

For two years, AEO was a publisher conversation. Google's AI summaries ate informational traffic first, and the agencies selling answer engine optimisation audits built their playbooks around that problem: article structure, question-and-answer formatting, machine-readable text files, "mentions" on third-party sites. Then in May 2026 Google published its first official documentation on optimising for generative AI features and spent an entire section saying most of those tactics do nothing for Google Search. It also did what the AEO industry has largely skipped: it named the thing that does move product visibility, and it isn't content.

This guide covers what AEO actually is once you strip the agency framing out, what the evidence does and doesn't support, and a three-shelf audit your merchandising and eCommerce teams can run in an afternoon without buying anything.

‍

What most retail AEO programs buy — and what that costs you

The standard engagement is an audit followed by a content sprint: an "AI visibility" scorecard against a competitor set, a recommendation to publish llms.txt, advice to chunk long pages, and a backlog of buying-guide articles written to be quoted by a language model.

Google's position on each is now on the record. You don't need llms.txt or any special markup, because Google Search doesn't use them. There is no requirement to chunk content, and no need to rewrite pages for AI systems. Seeking inauthentic mentions isn't as useful as it sounds, because the spam systems policing core ranking also police generative answers. Google goes further, telling site owners to evaluate third-party AEO and GEO services against its published guidance.

The real cost is not the invoice. It is that a content sprint aimed at your blog leaves the retrieval surface untouched. When a shopper asks an assistant for a waterproof pram cover that fits a specific model under $150, the system is not looking for your buying guide. It is looking for a product record that can answer a constrained query. If that record is missing the constraint, no amount of editorial fixes it.

‍

What answer engine optimisation actually is — and what it is not

Answer engine optimisation (AEO, also written as answer engine optimization) is the practice of making your information retrievable, extractable and attributable by systems that answer a question directly instead of returning a list of links. Generative engine optimisation is the same idea named from the model's side rather than the user's.

Google's view is that the distinction is largely marketing: its generative features are rooted in the core Search ranking and quality systems, using retrieval-augmented generation to pull pages from the same index and query fan-out to fire related sub-queries at once. There is no separate AI index. A page that cannot earn placement for a given intent cannot be retrieved, and therefore cannot be cited.

What is genuinely different for a retailer is the unit of optimisation. A publisher's answerable unit is a paragraph. Yours is a product record — the attributes, price, availability, variants, return terms and review signals that describe one item across your site and your feed. That is a merchandising asset, and it depends on the same product data consistency that governs pricing and stock everywhere else. Treating AEO as a content problem is the category error most of the market is currently making.

‍

What the evidence says

The click loss is well established — the shopping-specific loss is newer

SparkToro's analysis of Similarweb's US clickstream panel found 68.01% of Google searches between January and April 2026 ended without a click anywhere, up from 60.45% in 2024. Pew Research's browsing study of 900 US adults across 68,879 queries found people clicked a traditional result on 8% of visits with an AI summary against 15% without. Both are US measurements. Australian search behaviour has not been panel-measured at this resolution, so treat the direction as transferable and the magnitude as not.

Retail was insulated longer than publishing, and that insulation is ending. A study of 20.9 million shopping keywords by Visibility Labs — an AI SEO vendor whose commercial interest runs in the direction of the finding — reported AI Overviews on 14% of shopping queries in March 2026, up from 2.1% in November 2025. The methodology is disclosed and the keyword set defined structurally, by the presence of a shopping box. Treat the trend as credible, the exact figure as independently unverified.

Your product pages are probably the least machine-readable thing you own

Adobe benchmarked US retail pages with its own content visibility tool — a vendor-published diagnostic sold alongside an optimisation product, so read it as directional. The finding is still uncomfortable: homepages scored 75% machine-readable and category pages 74%, while individual product pages came in lowest at 66%. Returns pages scored 82%; FAQ pages 80%.

Read that ordering carefully. The pages retailers treat as compliance overhead are more legible to machines than the pages carrying the revenue — a natural consequence of how product detail pages get built: attributes live in a PIM, render through JavaScript components, and are laid out for a human eye scanning a template rather than for extraction.

What actually moves citation is relevance, position and extractable facts

The GEO evidence base is thinner than the industry implies, and a July 2026 critical survey of 45 studies is the best map of it. The foundational Princeton and IIT Delhi GEO paper, presented at KDD 2024, is sound within its design, but the famous "up to 40%" figure is a relative gain on a position-weighted word-count metric, measured on a source already placed in a five-document context. It establishes nothing about whether your page gets retrieved at all.

What replicates is narrower and more useful. A factorial experiment of 252,000 trials across six models and eighteen factors identified relevance and position in context as the primary determinants of the first citation, with explicit prices and recent dates producing measurable effects and formatting changes alone showing weak ones. A competing benchmark found only three of 54 method-and-domain combinations significantly positive, with gains eroding as competitors adopted the same tactics. Most instructive for anyone about to rewrite a catalogue: one end-to-end test found body-only optimisation reduced average top-20 retrieval presence by roughly 9%. A rewrite that reads better to a model can make the page harder to find.

Being known is not the same as being recommended

One 2026 preprint tested 112 products and found ChatGPT recognised 99.4% of them when named, but surfaced them in only 3.32% of open discovery queries; Perplexity fell from 94.3% to 8.29%. Single study, not replicated, so hold it loosely — but the distinction is structural. Brand recall inside a model is not distribution. If your visibility reporting runs on prompts that name your brand, it is measuring the wrong thing.

The traffic that does arrive is worth more than it was

Adobe reported that in March 2026, AI-referred visits to US retail sites converted 42% better than non-AI traffic, reversing a position twelve months earlier when the same channel converted 38% worse. This is vendor-published analytics from retailers on Adobe's own platform, and it compares channels rather than measuring a lift you can create.

Size it before you fund it. Take an Australian retailer running 500,000 monthly sessions with AI referrals at 1.5% — 7,500 sessions. At a 2% baseline conversion that is 150 orders; at 2.84%, 213. Sixty-three additional orders, and at the $96 average Australian online basket reported by Australia Post, roughly $6,000 a month. Real, but not a transformation. The channel is small and growing — fixing your feed does not conjure $73,000 a year.

Australian shoppers are using AI, and they don't fully trust it

Australia Post's 2026 report found Australians spent $82.6 billion online in 2025, up 14% year on year, with 32% of shoppers already using AI for shopping advice. The IAB Australia and Pureprofile Commerce & Discovery Report 2026 — a nationally representative survey of 1,079 Australians aged 18 to 70, fielded May 2026 — adds the counterweight: around eight in ten hold some concern about using AI to research products, and 69% still rely on their own research before buying. AI is entering the Australian research phase without displacing it. That argues for accuracy over volume: an assistant that misstates your price or stock availability costs you more than one that never mentions you.

‍

The three-shelf audit: what to check yourself, this week

Every check below can be verified by a merchandiser or eCommerce manager without an agency. Run it on your twenty highest-margin SKUs first, not the whole catalogue.

Shelf one: structured product data

1. Confirm you are actually eligible. A page must be indexed, eligible to show with a snippet, and included in Search generative AI features in Search Console. This is a settings check, and it is the single most common silent blocker.

2. Feed before schema. Google names Merchant Center feeds as a way to help products appear in AI responses. Note what Google also says: structured data is not required for generative AI features and there is no special schema to add, though it remains worth maintaining for rich results. With limited engineering time, spend it on feed completeness, not schema experiments.

3. Reconcile feed and page. Where your merchant listing markup and your Merchant Center feed disagree on price, availability or condition, you have created a conflict a retrieval system must resolve — and it may resolve it by dropping you. This is the same single source of truth problem that breaks omnichannel pricing.

4. Fill the constraint attributes. Size, colour, material, compatibility, dimensions, GTIN — what a constrained query filters on. If you run 20,000 SKUs and 30% are missing the attribute a shopper filters by, that is 6,000 products that cannot be matched, regardless of how well the page reads.

5. Put price and date in the visible text. The controlled evidence points to explicit prices and recent dates as extractable units. If your price exists only inside a JavaScript-rendered widget, it may not be there when it counts.

Shelf two: question coverage

6. Mine your own contact reasons. Your customer service queue is a free list of the questions shoppers actually ask before buying. Waterproof? Machine washable? Fits which model? Every unanswered one is a query your record loses.

7. Answer on the product page, not in a blog post. The answer has to sit on the record being retrieved for that product intent. This is question coverage, not content marketing — and Google's guidance warns against spinning up separate pages for every query variation, which its scaled content abuse policy covers.

8. Write the specifics only you have. Google is explicit that non-commodity content — first-hand, specific, not reproducible by a model — is what earns visibility. For a retailer that means fit notes, wear observations and real comparisons, not a restated spec sheet.

Shelf three: review signals

9. Make ratings machine-readable and accurate. Aggregate ratings are extractable units of exactly the kind that replicates in the evidence. Ensure the rating rendered to a shopper matches what is exposed in your markup.

10. Check your review practice against the ACL before amplifying it. This is risk US-written AEO advice ignores. The ACCC's guidance is direct: incentives must be offered regardless of whether the review is positive or negative, and must be clearly disclosed. Businesses that don't remove reviews they know or should know are fake may be breaching the law. Optimising review volume without fixing review governance turns an SEO project into a compliance exposure.

‍

What we still don't know

• Whether any of it durably improves discoverability. The 2026 survey's blunt conclusion: no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or on downstream clicks and conversions.
• Whether gains survive competition. Where visibility shares sum to one, an advantage available to everyone is an advantage to no one. Benchmarks already show gains eroding as adoption rises.
• Whether citation converts. The evidence connecting AI citation to revenue is the weakest link in the chain — and the one every AEO pitch deck treats as settled.
• What it looks like in Australia. Every serious measurement study cited here is US or European. Australian query mixes, catalogue sizes and seasonality differ enough that these figures should not transfer directly.
• How stable the measurement is. Audits find AI Overview source overlap of roughly 18% across two months, against 45% for organic Google. A single visibility score is a snapshot, not a rank.

‍

So — should you run an AEO program?

Yes, but not the one being sold to you, and not as a separate program.

If your product data is incomplete — missing constraint attributes, feed and page disagreeing, price rendered client-side — fix that first. It is the only intervention here supported by both Google's guidance and the controlled literature, and it pays off in paid shopping, on-site search and merchandising whether or not AI referrals become material. That is a bet with a floor.

If your product data is already clean, the marginal return on further AEO work is unproven. Spend the next quarter measuring instead: use the Generative AI performance report in Search Console, track unbranded discovery prompts rather than prompts that name you, and repeat measurements across several days before believing any number.

What you should not do is buy a content sprint aimed at a retrieval surface it cannot reach. The zero-click shelf is stocked from your catalogue, not your blog.

‍

Fix the record, not the article

Awayco works with enterprise retailers on the transaction layer, where product, price, availability and cart state have to agree across store, web and mobile. The inconsistencies that break a shared cart are the same ones that make a product record unusable to a retrieval system. For a read on where your product data breaks down, book a data consistency review with our team.