AI Search

Not All Pages Are Candidates in Google AI and AI Search

By Todd Gårdefors · · 8 min read

Why retrieval systems favor pages that reduce uncertainty. An in-house case study on ecommerce filters, use cases, and moving content from administration to infrastructure.

An abstract digital illustration showing select data blocks lighting up while others remain dark, representing content selection filters in AI retrieval layers

The ecommerce teams I’ve encountered these past years still think visibility works roughly the same way it did a few years ago. You rank well, you get traffic. *Simple enough.*

The problem is that ranking is no longer the final decision.

A page can rank perfectly well and still contribute almost nothing in AI search because the system does not really care about the page in the way we used to think about pages. It cares about whether parts of the page are useful enough to reuse.

That sounds subtle. Commercially, it is not.

Where this started getting interesting for us inhouse

When I was working inhouse as an SEO Manager, I started noticing this internally earlier in 2025 while working inside a stroller category ahead of Q2.

We tracked historically strong products with predictable seasonal demand. The same products had sold well the previous year too, which made comparisons reasonably fair—instead of the usual ecommerce habit of comparing Black Friday against a random Tuesday and calling it insight.

The interesting part was not rankings. Some products improved there, some barely moved. The interesting part was everything happening around the ranking:

Across the product set, organic revenue ended up roughly 15% higher year-over-year during peak season after the updates went live. We also saw higher visibility in AI-driven search traffic, especially in ChatGPT. It is a small sample size, not academic proof, and multiple variables were involved, fair enough. Still, the result is hard to ignore.

Especially because what we changed had very little to do with “AI optimisation” in the LinkedIn-post sense of the phrase. We mostly just made the pages more useful. That turned out to matter more than expected.

What we actually changed

None of the work was particularly exotic. Mostly, we stopped assuming the customer already understood the category. That assumption quietly breaks a surprising amount of ecommerce content.

We added clearer use cases to product pages

Instead of generic product copy, the pages explicitly explained:

Because phrases like “lightweight stroller with suspension” mean very little to someone buying their first stroller while panic-Googling at 01:14 in the morning.

Meanwhile, copy like “better suited for city parents regularly using public transport and small storage spaces” is both easier for humans to understand and vastly easier for retrieval systems to match against explicit user intent.

Key note: we added this to enhance core SEO and conversion, not to intentionally chase visibility in AI search. The reality is that the structural techniques are essentially identical.

We explained technical features properly

Most ecommerce product pages list technical specifications as if the customer already has deep category expertise. Most customers do not. Especially not first-time parents.

So instead of only listing features, we explained:

Just useful, grounded information. Machines seem to appreciate that far more than vague, emotional claims about “premium quality craftsmanship” written by someone detached from the product category.

We expanded filters and attributes heavily

This part is painfully underrated. We aggressively improved filter completeness, structured attributes, comparison logic, category relationships, and product-specific detail depth.

Again, it is not glamorous work—which is usually a decent sign that it matters commercially.

More complete attributes gave the products more retrievable dimensions. That becomes critical once AI systems start decomposing user queries into smaller sub-questions behind the scenes.

A customer may search for “best stroller,” but the retrieval system underneath effectively breaks that query down into a fan-out:

One generic product page struggles to cover all of that well. A page optimized with situational use cases, practical explanations, structured schema data, deep comparisons, localized FAQs, and clearer attributes suddenly becomes far easier to retrieve and reuse.

What I think is actually happening

I do not think AI systems are rewarding “good content” in the vague marketing sense people keep repeating.

I think they are favouring pages that reduce uncertainty efficiently. And that changes what qualifies a page as a useful candidate. Pages become easier to retrieve when they:

The retrieval layer is increasingly looking for usable sources, not just relevant URLs. There is a major difference.

Why some lower-ranking pages still get selected

You can see this pattern clearly once you stop looking only at rankings and start testing real purchase prompts. Two retailers can sell almost identical products:

Retailer A: Ranks higher, relies on boilerplate supplier copy, includes raw specifications, and maintains decent traditional technical SEO.

Retailer B: Explains exactly who the product is for, compares it directly against alternatives, explains engineering tradeoffs, answers common questions, and includes clearer contextual detail.

Historically, Retailer A could still do completely fine if authority signals were strong enough. That was the old game. Rank high enough, look credible enough, and you often get the benefit of the doubt.

Now, Retailer B is often more reusable inside AI-generated answers because the system has more granular substance to extract from. The page successfully contributes to more sub-questions.

When we tested this in practice on our stroller category in late Q1 2025, we saw the same pattern. Before we improved the product content, our more popular stroller products were barely visible when we ran the specific purchase prompts we cared about in ChatGPT.

In several cases, we were being outperformed by pages that were not obviously stronger on brand authority, but were significantly easier for the system to reuse because they gave it direct context and clearer decision support.

The operational problem underneath all of this

The uncomfortable part is that many ecommerce teams still treat product content as administration rather than infrastructure.

Supplier copy goes in. Someone tweaks the first paragraph slightly. Half the filters remain incomplete because nobody finished the attribute mapping three platform migrations ago. FAQs exist on only six random products, and comparison logic lives inside someone's head rather than on the site where customers can actually use it.

That was survivable when discovery depended mostly on traditional rankings. It becomes much more fragile once retrieval systems start deciding which sources are easiest to interpret, extract from, and trust. Machines are surprisingly unforgiving about structural ambiguity.

What smart ecommerce teams should probably audit now

Do not panic. Do not rebuild your entire SEO strategy around sensationalist AI headlines. But realistically assess whether your commercially important pages are actually useful as retrieval sources.

Questions worth asking during your next audit:

  1. Does the page help customers decide, or does it just describe the product?
  2. Does it explain why specific features matter?
  3. Does it explicitly answer adjacent buying questions?
  4. Does it contain original information that competitors do not possess?
  5. Are the attributes complete enough to support algorithmic retrieval layers?
  6. Would a machine have a definitive reason to prefer this source over ten near-identical retailer pages?

Those questions are becoming commercially critical surprisingly quickly. Because not all pages are viable candidates anymore. And increasingly, the systems deciding visibility seem to care less about whether your page exists and more about whether it deserves to be reused.

But wait, there's more. I have written more on the topic "Google AI and AI search":

More on AI Search