What an AI assistant can read on your product page (and what it can't)

Product pages take 6% of the sources AI assistants cite on buyer queries, across the 40 Shopify stores we track, and 19.3% of brand-domain citations in AirOps' data. They are never the page that wins the recommendation. This post is about the moment your page does matter: when the assistant opens it to verify what it is about to recommend.

Most product pages fail that moment on mechanics, not marketing. The facts sit in a JavaScript widget no AI crawler executes, or in the bottom half of the page that 44.2% of citations never reach, or in a spec sheet shipped as a JPEG. The fixes are cheap, and the evidence says they pay: in Princeton's GEO study, modest content changes lifted visibility in generative-engine answers by up to 40%, with the largest gains going to sites ranked lower in classic search. For a small store, page mechanics are a levelling opportunity.

A citation, as throughout this series, means the source an assistant names behind its answer. The numbers this post is built on:

Each number is unpacked, with method and caveats, in its section below.

Can the assistant read your page at all?

Start with the blunt finding. Vercel analysed how AI crawlers process the web, including more than 500 million GPTBot fetches, and found that none of the major AI crawlers render JavaScript. GPTBot, ClaudeBot and PerplexityBot download your script files; none of them run them. Anything your page draws in after load (a reviews widget, a tabbed spec panel, an app-injected size guide) does not exist for ChatGPT or Claude.

The exception is Google. Googlebot renders JavaScript, and Google's AI surfaces ride its index, so a JS-dependent page is visible to Gemini and AI Overviews while staying invisible to everyone else. That is the same per-engine split the engine teardown found in citation sources, now operating at the rendering layer.

For a standard Shopify store the default position is good: Liquid themes are server-rendered, so your catalogue ships as real HTML. The failure modes are the additions. Review apps that inject their content client-side put your social proof where only Google can see it. Headless builds without server rendering hide the whole catalogue. Bot-protection rules (a firewall app, a Cloudflare setting) can challenge AI user agents even when robots.txt welcomes them. And slow responses get dropped: the four-week sequence in the hub post starts with getting server response under 200ms for exactly this reason.

The one-minute test: open view-source on your best-selling product page, or curl it. If the spec table and the reviews are not in that raw HTML, they are not in any ChatGPT or Claude answer.

The first 30% decides what gets quoted

Kevin Indig analysed 3 million ChatGPT responses and 30 million citations, isolating 18,012 verified citations to study where in a page the quoted material lives. The distribution is steep: 44.2% of citations come from the first 30% of the content. The middle 40% of the page supplies 31.1%. The final third supplies 24.7%, with a sharp drop near the footer.

Source: Kevin Indig, 18,012 verified ChatGPT citations, as reported by Search Engine Land, February 2026.

The paragraph-level finding is the useful surprise: 53% of citations come from the middle of paragraphs, not the openers, and plain short sentences outperform dense prose. An assistant reads the way a skimming buyer does. It rewards pages that state facts early and plainly.

On a product page, that means the first 100 words carry the load. What the product is, the three numbers that matter (capacity, runtime, weight, whatever your category trades in) and who it fits. "Crafted for adventurers who demand more" extracts to nothing. "40-litre compressor fridge, 45 W draw, runs 18 hours on a 100Ah leisure battery" hands the model three facts it can quote. The same logic kills the spec-sheet JPEG: an image of a table is invisible to a text crawler. Real table, real text.

One scoping note from Evertune's data: the typical ChatGPT-cited page runs 941 words with four H2 sections and 15 external links. That is a buying guide, not a product listing. Don't inflate your product page to chase it. The product page's job is verification facts up top plus clean structured data; the citable long-form lives on your guides, comparisons and video transcripts.

Schema won't earn the citation. Do it anyway

The honest evidence first. Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, matched them against 4,000 control pages, and measured citations across AI Overviews, AI Mode and ChatGPT. Adding schema produced no citation lift on any platform. Cyrus Shepard's synthesis of 54 studies scores structured data 5.6 out of 10 as a citation factor. Anyone selling schema as the way to get cited is selling against the data.

Schema earns its place doing three other jobs. It feeds the shopping surfaces ChatGPT now attaches to 87% of product questions, which are built from product feeds and structured data rather than prose. It makes the page unambiguous when an assistant does read it: Product, Offer, AggregateRating, FAQ and Breadcrumb markup state your price, stock and rating in a form that can't be misread. And it keeps your entity consistent across every page that mentions you. Necessary plumbing; not a citation lever.

This is also the part of the audit we automate. CitoRank scans your live product pages, scores the JSON-LD you have today, and flags exactly which of those fields are missing or broken, alongside the citation tracking. How the audit runs covers both halves.

What Princeton measured: facts beat adjectives

The Princeton GEO study (KDD 2024) tested content changes against generative engines and measured the visibility effect. Adding statistics, quotations and citations from credible sources lifted visibility by up to 40%. Keyword stuffing, the reflex two decades of SEO trained, did not help. And the authors showed the gains skew toward sites ranked lower in classic search: the AI answer layer rewards content quality over accumulated domain authority more than the blue links ever did.

The merchant translation is one sentence: a sourced number beats an adjective, every time. "IP44-rated, 45 W measured draw, kept contents at 4°C for 18 hours in Outdoorsmagic's bench test" is a quotable claim with a named source. "Rugged, efficient and reliable" is filler the model skips. Write the product page the way this series quotes studies, and the assistants have something to work with.

Rebuild one product page this week

Pick the product your margin depends on and run the five steps:

Then measure whether it moved anything. CitoRank runs your buyer-intent queries against ChatGPT, Claude and Gemini, scores your product pages' schema, and names every source behind every answer, so a page rebuild shows up as a rank change you can date. See how the audit runs or compare plans.