Why every reviews app just added an AI summary (and what should actually be in one)

That is a reasonable moment to ask: what should a review summary actually do? Because the basic version, a synthesised paragraph that says "shoppers praise the fit and packaging", is easy to build and not very useful.

What the basic version does

The standard AI summary implementation takes your review corpus, runs it through a language model, and outputs a one-to-three-sentence synthesis to display at the top of the reviews section. It is a compression exercise: a shopper who does not want to scroll 200 reviews gets the main themes in a sentence.

This is useful. The Baymard Institute has documented consistently that review section friction (too many reviews, no way to filter, no clear signal about what matters) is a real barrier at the decision stage. A paragraph that says "consistent praise for fit and sizing accuracy, frequent mentions of fast delivery, some concerns about packaging for fragile items" tells the shopper something actionable faster than scrolling does.

The problem is what it does not do.

Where the basic version stops

A synthesis paragraph with no citations is an assertion. A shopper who sees "shoppers praise the fit" has no way to verify that claim, or to find the three reviews that actually mentioned fit, without scrolling anyway. The summary replaced a reading task with a trust task: do you believe the summary?

The product-tag gap is the other problem. A summary that says "consistent praise for sizing accuracy" is not connected to a specific SKU or size range. A shopper buying a large in a different colourway has no signal from that summary about whether "sizing accuracy" applies to their variant. Product-tagged UGC, where the review or video is linked to a specific variant, is the evidence that the summary should be drawing from, but most review-app summaries aggregate across all variants and lose that signal.

The third gap is agent readability. AI shopping agents (ChatGPT Shopping, Google AI Mode, Gemini via Agentforce) read structured data, not just rendered HTML. A summary paragraph that lives inside a JavaScript widget and is not published in any structured format is invisible to those agents. The brands with AI-summarised reviews who are also showing up in agent recommendations are the ones who have published that content in a machine-readable layer, not just on the page.

What Iris does differently

Iris is Idukki's unified AI layer. Its review summary implementation starts from product-tagged UGC, meaning the source material for the summary is already linked to a specific SKU and variant, not to the product as a whole. The summary it generates is therefore variant-specific: if a shopper is on the large-cobalt colourway, the summary draws from reviews on that variant, not from reviews on the small-sage version that sits on the same PDP.

Each claim in the Iris summary links to the specific review (or set of reviews) that supports it. A shopper who sees "consistent praise for sizing accuracy" can click through to the four reviews that used that phrasing. The summary is a navigation aid, not a replacement for the underlying evidence.

The agent-readability part matters for a different audience. Iris publishes the summary content, along with the supporting review attribution, into the structured feed that Idukki maintains for AI agents: llms.txt, review JSON-LD, and the MCP endpoint. A ChatGPT Shopping agent asking about sizing for a specific product variant can resolve that query against the structured summary data, not just the rendered HTML. That is how a product page earns an AI recommendation, not by appearing in a search result.

The Agentforce Commerce GA piece covers why the agent-readability side of this has become urgent in the last month. The Idukki vs. Loox comparison now reflects the AI Summaries feature in Loox's updated feature set.