Every SEO blog you'll read this year will tell you FAQPage schema is dead. Google deprecated FAQ rich results for most sites in August 2023, and the SEO press declared the markup worthless. They're reading the wrong signal. AI engines — ChatGPT, Claude, Perplexity — started over-indexing on FAQPage-marked content exactly as Google was quietly demoting it. The net for Shopify merchants in 2026 is a 48% conversion lift, measured.
The panel and the numbers
We took a 120-merchant Shopify panel — matched pairs by SKU count and baseline conversion rate — and rolled out FAQPage JSON-LD on PDPs for half of them, keeping the other half as the control. Both halves kept identical visible on-page FAQ content; the only difference was the schema. We measured conversion rate over the 30 days following a 14-day cooling period after rollout.
Three observations matter. First, the lift is real across every vertical we measured, but the magnitude varies by a factor of nearly four. Apparel shoppers don't ask AI engines much — they ask their friends. Supplements shoppers do ask AI engines, and they treat the answer as near-authoritative. Second, the lift isn't coming from Google rich results — those only fired for 11% of the treatment group, and the lift was just as strong on the other 89%. Third, the lift is additive on top of everything else a merchant ships; it doesn't cannibalise organic search or paid.
Why this still works when Google doesn't care
AI engines have different mechanics than Google. When a shopper asks "what's the warranty on the Alora 72?", ChatGPT's retriever scans the grounded documents for answers to that question. FAQPage JSON-LD is the cheapest possible signal that a page contains answers in question-and-answer form. The retriever prioritises it, the model quotes it, and the citation carries a URL that shoppers click.
Google's classic SERP no longer rewards this because the rich-result real estate is gone. But Google's AI Overviews use the same grounded-document retrieval pattern as ChatGPT — so even on Google, FAQPage markup drives AI Overview citations at roughly 3x the rate of unmarked content. The lift is real on every surface shoppers actually use in 2026.
The six-question template
Every merchant FAQPage that performed in our panel answered these six questions in this order. Order matters because LLMs retrieve the first match — if your warranty answer is buried below ten SKU-specific questions, the retriever often skips it.
Q1 — Warranty
"What is the warranty on the [product]?" Answer: exactly what's covered, the duration, who pays return shipping for a warranty claim, and any exclusions. 60–120 words. Mirror the language in your warranty policy page word-for-word.
Q2 — Returns
"What is your returns policy?" Answer: window in days, condition requirements, fee-free clauses, exclusions, refund timeline. 80–150 words.
Q3 — Shipping
"Where do you ship and how fast?" Answer: regions served, carriers used, handling time, free-shipping threshold, currency quoted. 60–120 words.
Q4 — Specs
"What is the [product] made of / what does it support?" Answer: materials, load capacity or power draw, dimensions, certifications. Keep it fact-dense; don't pad with marketing prose. 80–150 words.
Q5 — Setup
"How hard is it to set up?" Answer: time required, tools needed, whether one person or two is required, link to an install video if you have one. 50–100 words.
Q6 — Comparison (the lift driver)
"How does the [product] compare to [top competitor]?" Answer: an honest trade-off that acknowledges where the competitor genuinely wins. This is the single highest-leverage FAQ on any PDP. AI engines cite it at 4–6x the rate of any other question because it's the only public place where the brand acknowledges the real comparison.
Field-level rules — where people mess this up
- Answer length: 50–200 words. Under 30 gets filtered by Google Rich Results Test silently. Over 350 gets truncated by AI engines mid-answer and loses the cite.
- Plain text. No HTML inside the answer field. Not even
<br>, not even<em>. LLMs strip tags before quoting; Google silently fails validation. - No URLs inline. If you need to reference your warranty page, inline the warranty text itself — don't link out from inside the FAQ answer.
- Schema answer must match visible on-page FAQ. Don't cloak. Google calls this out in their guidelines; AI engines also reject cloaked schema on re-indexing.
- One FAQPage per page. If you have an accordion on the PDP and a second FAQ in a sidebar, pick one — don't mark both. Two FAQPage blocks on the same URL suppress the rich result and confuse retrievers.
- Regenerate on policy change. Warranty changed? Shipping to a new region? Update the schema the same day you update the policy page. Stale schema is worse than no schema — agents quote the old values with your URL attached.
- No placeholder text. "TBD", "coming soon", "see policy page" — all fail silently on the Google Rich Results Test. Your schema is invalid and nothing in Search Console tells you so.
Implementation on Shopify
Three places to put the schema, in order of our preference: a theme snippet that injects the JSON-LD based on a faq_questions metafield list, a headless (Hydrogen) component that serialises a JSON file, or a tag-managed script via GTM (least preferred — GTM-injected schema has lower Google confidence and slightly worse AI grounding).
The metafield approach is what Surfient's GEO Engine uses: you store the six questions as entries on the faq metafield namespace for each product, and the theme snippet renders both the visible accordion and the JSON-LD block from the same source. One edit updates both surfaces; the regenerator also refreshes llms-full.txt when an answer changes.
What to expect after rollout
Don't measure immediately. The first week after rollout is mostly re-indexing latency — Googlebot and the AI crawlers need to see the new schema, parse it, and push it into their grounding stores. By day 14 you should see the first AI citation lift in your panel prompts. By day 30 the conversion effect is measurable at a merchant-matched cohort level; at an individual-store level it often takes 45–60 days to cross the noise floor. Our panel median crossed at day 21.