I came across Shopify’s recent research on AI-referred shoppers and one thing really stood out to me: 56% of AI-referred sessions landed directly on a product page, compared with 20% from organic search.
It got me thinking about what that means for the PDP.
Someone coming from ChatGPT may have already spent quite a bit of time researching and comparing products before they arrive. They may not need the whole sales pitch — they might just have one or two things left they’re unsure about before buying.
Shopify recommends making PDPs more useful for these shoppers with things like clearer use cases, comparisons, reviews, FAQs, shipping/returns, etc.
But adding everything a shopper might need can also make PDPs pretty crowded.
Curious what merchants here are actually doing:
Are you changing your PDPs because of AI-referred traffic yet?
And are you seeing any difference in how those visitors behave or convert?
The part that made this click for me was realising the assistant and the human need the same facts in two different places.
An assistant reading your PDP works off what comes back in the HTML. Anything sitting inside a tab, an accordion, or a review widget that only paints after JS runs is often not there at all. Easy check, curl your own product page and read what actually comes back. If your price, stock state, shipping table or review count only shows up after scripts run, the answer someone gets about your product is being built from a page you have never seen.
So the crowding worry mostly goes away. Put the full facts in structured data, Product with offers, availability, shipping and a return policy, and keep the visible page short. One plain FAQ block near the bottom covers the rest without turning the page into a spec sheet.
The other thing about landing straight on a product page is that no collection page did any pre-sell. The questions left are usually fit or sizing, how long shipping takes to their country, and the return window. Those three earn their space, most other additions do not.
Worth segmenting too, that traffic lands as a referral from chatgpt.com in GA4 so you can watch it on its own instead of guessing.
Are you seeing it convert better or worse than organic so far?
@lumine’s point that no collection page did the pre-sell has a second consequence, on basket size rather than conversion. The browse path is also where most stores do their merchandising — related products, complete-the-look, category cross-links — and AI-referred visitors skip all of it. So whatever basket-building happens has to sit on the PDP itself or after the add to cart.
Since that traffic already segments as a chatgpt.com referral, AOV by source is one report away. If AI-referred orders skew single-item, the answer isn’t a longer PDP — it’s where the second item gets offered. Has anyone actually pulled that number?
Hi @priyanks
I think the bigger shift with AI-referred shoppers is the level of intent they arrive with. If someone reaches a PDP after researching through ChatGPT, I’d expect them to be looking for confirmation rather than discovery, so I’d focus less on adding more content and more on making the key decision-making information immediately clear.
Things like who the product is actually for, the most common objection, real customer feedback, delivery expectations, and what happens if it doesn’t work out could probably have more impact than adding another section just for SEO.
I’d also be interested in comparing behavior beyond conversion rate, things like time on PDP, add-to-cart rate, and whether AI-referred visitors return before purchasing. That could tell us whether they’re genuinely more purchase-ready or simply using AI as another research step.
I’d segment first: GA4 channel groups matching chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, and claude.ai, along with utm_source=``chatgpt.com, placed above generic Referral. Run 30–90 days of CR and AOV against Organic. These numbers are a floor, since referrer stripping pushes some AI traffic into Direct.
If the segment moves the needle, a redesign is justified just for that segment. If not, you’re adding complexity for a sliver of your sessions. The same research notes that Organic still refers more traffic than every AI platform combined. I’ve seen less than 1% of traffic coming from LLMs on stores.
What I’ve seen, usually, is that these users are returning users. They may have already visited the store, had an initial session, and are now conversing with LLMs about your products. I would personalize the PDP based on that previous session in such a case.
I’d be careful about adding more to the PDP just because the traffic source is AI.
On our apparel brand, the pages that did better for high-intent visitors were the ones that answered the key buying questions fast: size/fit, shipping/returns, etc.
We moved the comparison and FAQ content lower on the page, kept the first screen tight, and put the fit answer right next to the size selector because that’s where the decision was actually getting made.
One thing I’d test is a short “why this over similar options” block that compares 2 or 3 real differences, not a giant feature dump. AI shoppers usually already know the category, so they’re looking for confirmation that they picked the right one, and a clean comparison plus a couple of review quotes that speak to the exact worry tends to do more than a long wall of copy.
I haven’t seen enough first-party data yet to make a strong claim on conversion differences, which is partly why I asked.
What I find more interesting is the second part of your point. If AI is doing more of the discovery/comparison upstream, the shopper may land with a much narrower set of unresolved questions.
My concern with the static approach is that those questions won’t be the same for everyone. One shopper may still need fit reassurance, another shipping timing, another evidence around quality/value.
I’m starting to think the bigger PDP opportunity may be less “add more information” and more “figure out what this shopper still needs and surface that selectively.”
Have you seen anyone actually testing that kind of adaptive PDP behavior yet?
I wouldn’t build a separate AI version of the PDP yet. What I’ve seen work is tightening the existing page around the final objections:
Put fit, compatibility, delivery estimate, and return window beside the buy box, not buried in tabs.
Add a short comparison table with 3 to 5 differences versus the closest alternatives. Keep it factual.
Check the page with JavaScript disabled and validate Product schema. Price, availability, variants, reviews, shipping, and returns should still be machine-readable.
Create an AI referral segment in GA4, then compare add-to-cart rate, conversion, and AOV against organic over at least 30 days.
Also annotate the date of any PDP change. AI traffic is often small and partially attributed to Direct, so weekly numbers can be misleading.
I’d test the adaptive idea without building a separate AI-only PDP first. Create a small “what’s still unclear?” choice (fit, delivery, returns, comparison) for a random sample of AI-referred visitors, then surface the matching answer beside the buy box. Keep a holdout seeing the standard page.
Measure the full path by source: PDP view → answer interaction → add to cart → checkout entry → purchase, with device and new/returning splits. Annotate every PDP change and run long enough to avoid weekly noise; AI traffic is often small and partly attributed to Direct. That tells you whether the extra guidance resolves a real objection or only increases clicks.
I’d view this as an opportunity to enhance the product details page rather than building-out a separate AI-centric experience. The product pages ought to answer the common last questions fast. What is it for? How do I use it? What are the key specs? What do other customers think? When will it be delivered? How do I return it? And are there useful comparisons?
For busy pages, I like progressive disclosure in the form of accordions or well defined sections. Then, through Shopify Analytics, examin comparisons on sessions, conversion rate and engagement by various referral sources across a period of time. The key is seeing if AI-referred shoppers have different questions or points where they bounce, and then adjusting the PDP to address those trends.
I wouldn’t build an “AI shopper” version of the PDP yet. An AI referral is a traffic source, not necessarily a distinct customer type.
The more useful question is what happened before the click:
If the product was absent from the candidate set, adding more PDP sections may not solve the real problem. It could be category association, incomplete commerce data, or weak outside evidence.
If it was included but compared using the wrong facts, make those specific facts clearer and consistent across the PDP, structured data, and relevant policy or collection pages.
If it was already recommended accurately, the PDP has a last-mile job: confirm fit, limitations, price and availability, delivery, returns, and the evidence needed to buy confidently.
That keeps the page from becoming a giant checklist. Add the smallest truthful fact that answers an observed objection, then measure behavior on that landing page.
I’d also keep referral performance separate from recommendation testing. A captured AI answer shows whether a product was found and compared; analytics shows what identifiable visitors did afterward. Neither one alone proves the other.
I like this framing — especially looking at whether AI-referred shoppers actually behave differently before deciding they need a different experience.
One thing I’m wondering about, though: analyzing referral-level trends still gets you to an aggregate view of what shoppers might be struggling with. What if the more useful opportunity for AI is at the individual session level?
For example, if one shopper appears uncertain about fit while another keeps looking for delivery or return information, AI could potentially interpret those behavioral signals and surface the most relevant existing information at that moment — rather than adding more content to the PDP for everyone.
Do you think there’s value in using AI that way to handle the “last mile” at scale, or would you still prefer optimizing the PDP around the aggregate patterns you see in analytics?
This makes a lot of sense, especially the distinction between AI referral as a traffic source rather than a distinct customer type.
Your point about the PDP having a last-mile job is interesting though. If the goal is to identify an observed objection and surface the smallest truthful fact that resolves it, I wonder if that’s actually where AI could be useful.
Doing that manually seems feasible at a page or campaign level, but much harder across thousands of shoppers and hundreds of products, where the unresolved question may be different each time — fit for one shopper, delivery for another, return risk, material, reviews, etc.
If AI could infer that remaining uncertainty from on-site behavior and surface the relevant existing evidence at the right moment, would you see value in that? Or do you think the inference itself would be too unreliable?
Before the click: was the product retrieved, associated with the right need, and included in the candidate set?
After the click: what uncertainty is preventing this particular shopper from buying?
Session-level personalization can help with the second problem. It cannot fix the first one if the product never entered the AI shopping decision in the first place.
The inference would also be more credible when it is based on explicit behavior—opening the size guide, checking delivery, revisiting the return policy, switching variants or viewing comparisons—rather than trying to infer intent from one weak signal.
At scale, I’d constrain the system to selecting and prioritizing merchant-approved evidence blocks, not generating new product claims. If confidence is low, it should fall back to the standard PDP. I’d also test it against a non-personalized control rather than assuming that personalization is automatically better.
So yes, I think there is value—but more as an evidence-routing and hypothesis-ranking system than an autonomous salesperson. The core PDP still needs to contain stable, complete and truthful facts for every shopper.