What AI Search Wants From Your Product Content

There's a new acronym doing the rounds and I want to tell you what's worth keeping from it and what you can safely discard.

GEO, or Generative Engine Optimisation, is the practice of structuring your content so that AI systems cite and surface it in their answers. It exists alongside SEO rather than replacing it, and the core idea is straightforward: when someone asks ChatGPT or Perplexity or Google's AI Overviews a question and the answer references your brand or product, that visibility has real commercial value.

The acronym is new. The underlying problem isn't. Getting your products discovered through the right channels has always been the game. The channel is changing.

What's Actually Shifting

Traditional SEO was built around a simple model: search engines crawl your page, index your keywords, and match queries to content based on relevance signals. The implicit audience was two things at once: a human who would read your product description and a crawler that would index it.

Generative search changes that model in a specific way. The AI doesn't just index your content. It reasons about it. It forms a view of what your product is, what it's good for, who it suits, how it compares to alternatives. Then it synthesises that into an answer for someone whose question may look nothing like your page title.

A product description that says "Classic Fit Chino Trousers, 98% Cotton, Available in Navy, Stone, and Khaki" is legible to a keyword crawler. It tells a generative AI very little about whether this is the right trouser for someone who works in a casual-smart office and needs something that works for both a formal meeting and an evening out. The question they actually asked.

The Evidence on What Helps

Research from Princeton and IIT Delhi, published in late 2023 and presented at ACM KDD 2024, established what content characteristics improve visibility in generative search responses. The findings are useful but not entirely surprising to anyone who's spent time thinking about content quality rather than content volume.

Adding citations, statistics, and quotations to content boosted visibility in generative engine responses by up to 40% in their testing. Improving fluency and readability delivered a further 15 to 30%. What didn't help: keyword stuffing. Adding more keywords performed no better than baseline, and in some tests made things worse. The same move that worked briefly for traditional SEO doesn't translate.

That said, the benchmark covered 25 diverse query domains and didn't specifically test ecommerce product pages. My reading of the evidence is that the direction holds (generative engines favour content that is verifiable, well-structured, and easy to process), but I'd be cautious about treating specific percentages as product-page targets.

Third-party credibility signals matter more than they used to. Generative engines tend to favour content that itself references credible sources, includes real data, and exists within a broader web of coverage. This is inconvenient for purely brand-owned content and advantageous for brands that have generated genuine press, review coverage, and independent commentary.

The Product Data Problem, Again

Here's the thing that I keep coming back to, and that I keep noticing is underappreciated in conversations about AI and ecommerce.

Most large product catalogues are in a state I would diplomatically describe as mixed quality. There are products with comprehensive descriptions, multiple use-case callouts, and good technical specifications sitting alongside products with a generic two-sentence description and three attributes. The latter were acceptable when the audience was a human who could look at the photos and make their own judgement. They are much less acceptable when the audience is an AI that has to synthesise an answer from what's written.

This matters for GEO but it matters more broadly. Amazon Rufus, which I wrote about earlier this year, occasionally hallucinates product specifications because the underlying product data is incomplete or inconsistent. The same risk exists for any AI system trying to reason about your catalogue.

The GEO conversation is partially a product data quality conversation in disguise. The underlying work is the same regardless of what you call it or what channel benefits: better specifications, richer use-case descriptions, more complete Q&A coverage.

What Not to Do

The worst version of this advice leads to keyword stuffing for generative engines: artificially inflating content with statistics, citations, and specific claims in hopes of being cited more often. The research is clear that this doesn't work, and the intuition tracks: a model that can reason about your content can also recognise when content looks like it was written to game an algorithm rather than to help someone.

The more durable version is the same thing it's always been: build product content that genuinely helps someone decide whether a product is right for them. Answer the real questions. Be specific about what something is and isn't suited for. Acknowledge limitations. The AI will get better at evaluating quality over time, and "looks like it was written to game an algorithm" is something these models detect more readily than you might think.

What I'm Actually Doing About It

At the moment, my most practical focus is on Q&A coverage. Most product pages have no Q&A content at all, or have the same generic questions applied across a category. Generating good Q&As (real questions actual customers would ask, answered specifically for each product) is an area where AI tooling can help generate volume while a human review layer maintains quality.

It also maps directly to what Amazon Rufus and similar systems use as a primary input. If your product has good Q&A data, the AI systems that reason about your catalogue have better material to work with.

It's not glamorous. It's product data hygiene dressed up with better questions. But that's what most of the genuinely valuable AI work in ecommerce looks like right now.

---

This piece was updated in May 2026. The 2025 follow-up on schema markup and structured data as AI infrastructure is at Structured Data Is Now AI Infrastructure.