A product can have the right price, the right material and the right features, and still lose the sale to a worse product. Not because the shopper compared them. Because the AI doing the comparing could not see half the page.
That is the quiet failure in AI shopping. Nobody gets a rejection notice. The product simply never makes the shortlist.
Four moves, in the right order
A recent webinar recap on Search Engine Journal covered a session with Chris Sachs and Aris Vrakas of seoClarity on getting products found in Google and ChatGPT. They described four connected moves: make product pages readable, enrich content in shopper language, synchronise feeds and pages, and measure what changed.
None of that is exotic. What I found useful was the order. Most ecommerce teams I talk to start with the second step, rewriting descriptions, and skip the first one completely.
Access comes before persuasion
Vrakas started with a technical prerequisite. Product content has to be available to bots, not only visible to someone in a browser. He described JavaScript-heavy pages where shoppers could see details that an automated fetch never received.
This is the most common problem I see on modern ecommerce builds. The price, the size chart and the reviews load after the page renders, through scripts a crawler may not run. The human sees a complete page. The machine sees a title and a loading spinner.
The check is simple and almost nobody runs it. Fetch your own product page the way a bot does and read the raw HTML. If the core attributes are not there, no amount of copywriting will help, because the system never gets the copy. Prerendering or cached HTML fixes most of it.
I covered the broader version of this in making your site legible to machines. Product pages are where that problem costs the most, because every invisible attribute is a lost comparison.
Write for the question, not the spec sheet
The presenters used socks as a running example, which is perfect because socks are boring. Colour and fibre composition are standard attributes. But a shopper asking an AI assistant wants to know about warmth, thickness, durability and whether they work for hiking in November.
Conversational prompts combine several needs in one request. "Warm socks for long winter walks that do not slip in boots" is one sentence and four filters. A spec sheet that says "80% merino, 20% nylon" answers none of them directly.
The fix is to connect product characteristics to the decisions shoppers are trying to make. Say what the product is for, who it suits and where it fails. Vrakas also recommended attributable review evidence to support performance claims, which matters because AI systems weigh third-party proof more than brand adjectives.
One product, one version of the truth
Sachs warned against treating the product feed as a separate project owned by someone else. This is the organisational point, and it is the one that usually blocks progress.
In most retailers, the feed belongs to paid media, the product page belongs to merchandising or SEO, and structured data belongs to development. Each team improves its own version. The result is three slightly different descriptions of the same product, with different prices and attributes depending on which system updated last.
For an AI system revisiting the product, that inconsistency is a trust problem. Which price is correct? Which size range? The presenters called the fix closing the loop, where information added to the feed also updates the page and its structured data.
From an agency perspective, this is the hardest part to sell and the most valuable to fix. It requires SEO, paid media, merchandising and development to agree on one source of product truth. It is a governance decision dressed up as a marketing task.
Measure in separate columns
The measurement section was the most honest part of the session. The presenters listed three signals: product-grid visibility in shopping carousels, AI mentions and citations, and paid-feed performance in impressions, clicks and conversions.
Then they said the important thing. These are different signals, not interchangeable definitions of success, and an AI mention alone does not establish a sale.
I agree, and I would add one discipline. Report the three in separate columns and do not sum them into a single "AI shopping score". When a vendor gives you one blended number, you lose the ability to see which change actually moved revenue.
For the AI mention column specifically, GEOflux.ai tracks how often products and brands appear in ChatGPT, Gemini and Perplexity answers, and which sources drive those mentions. That tells you whether the access and enrichment work changed how machines describe you, before it shows up in sales.
Sachs tied it all to testing: the best SEO teams obsess about it. Change one thing on a group of product pages, keep a comparable group unchanged, and watch the difference. A before-and-after comparison on its own will mislead you, especially in a season like this one.
With AI shopping agents moving from demo to default, the product page now has two readers. The first one decides whether the second one ever sees it.