How to Optimize Your PDP for AI Recommendations (Based on 5,000 Simulations)

Joe Kiernan’s team ran 5,000 simulations to learn what makes an AI shopping agent recommend one PDP over another.

The clearest result was not “add more copy.” Genuine customer review text produced the strongest lift. A 4,000-character block of repetitive, keyword-stuffed copy made the product less likely to be recommended.

Read the findings as directional evidence, not an Amazon ranking formula. The test used synthetic shirt PDPs and one agent setup at one moment in model development.

The Monday-morning version

If you have one hour, choose one research-heavy ASIN and do four things:

1. List the five questions a careful buyer must answer before purchasing.
2. Check whether the title, attributes, bullets, images, and A+ answer them.
3. Compare those answers with the themes in reviews, returns, and support tickets.
4. Fix missing or contradictory facts before adding more copy.

1. Start with category fit

Not every seller should invest the same amount of time in AI optimization today.

Across Digital Fuel Capital’s portfolio, surveys suggested AI touched at most 5% to 10% of shopping journeys. Referral data showed a sharper divide between categories.

High-consideration purchases were already receiving some AI traffic. These are products where shoppers compare materials, durability, compatibility, price, or other criteria before buying.

Impulse and taste-led purchases received little or no AI-referred traffic. Browsing a hundred apparel options can still be more useful than asking an agent to choose three.

Impulse buys see almost no AI traffic. High-consideration research is already moving to AI.

Joe explains which categories should act first:

The practical rule is simple. If shoppers research before buying, start now. If they mainly browse or buy on impulse, spend enough time to understand the channel and monitor it, but do not neglect proven acquisition work.

2. The three-stage AI Recommendation Funnel

Joe’s framework separates AI discoverability into three layers. The first is broad and difficult to influence. The last is where a seller has the most direct control.

Shape what AI already knows

Foundational training data can affect how a model understands a category or brand. The actions available to a seller look like conventional SEO: credible coverage, useful content, affiliates, and an authoritative web presence.

Joe treats this as the least important of the three layers. Many shopping agents use model training for reasoning, then run fresh searches to decide which products to recommend.

Win the live search

When an agent searches in the moment, the product must be discoverable and machine-readable. Complete marketplace data, conventional SEO, and valid Product schema help the agent find and parse the PDP.

The control surface depends on where the PDP lives. On Amazon, sellers control listing fields rather than page code. On an owned site, the team can also control schema, page structure, review display, and product feeds.

Earn the recommendation on the PDP

Discovery only earns consideration. The PDP must then give the agent enough evidence to decide that the product fits this shopper better than the alternatives.

This is where Joe focused the experiment, and where the largest practical opportunities appeared.

3. How the 5,000-simulation experiment worked

The team created roughly 50 synthetic PDPs for a fictional men’s Oxford shirt. The pages looked realistic, but no existing brand reputation or product demand could distort the results.

They added or removed one feature at a time, including images, a one-line value proposition, JSON-LD, long brand copy, keyword-stuffed copy, Q&A, and customer reviews.

For each comparison, an AI agent saw the changed PDP alongside the same competing products. The team repeated the choice across many trials and measured how often the changed PDP was recommended.

This design does not prove a universal ranking formula. It does isolate the direction in which each PDP feature moved the recommendation under the tested conditions.

4. What moved the recommendation

1. 1

   Full-text customer reviews

   Strongest lift

   Real review language gave the agent direct evidence from customers. A star rating and review count helped, but the written reviews mattered much more.

2. 2

   Buyer-specific Q&A

   Strong lift

   Answers to questions such as who should buy the product did more work than an equally long generic text block.

3. 3

   Valid product schema

   Helpful

   Structured product data improved the agent’s ability to interpret the PDP and may also have supplied a useful legitimacy signal.

4. 4

   Clear brand copy and a UVP

   Small lift

   Relevant brand information and a concise value proposition helped a little, but neither matched reviews or targeted Q&A.

5. 5

   Product images

   Small in this test

   Images produced only a slight lift with the tested agent. Joe cautioned that this result may be specific to the model and moment.

6. ↓

   Repetitive keyword stuffing

   Negative

   A 4,000-character block repeating that the shirt was the best reduced recommendations, possibly because it looked like spam or consumed useful context.

What the experiment does—and does not—prove

The test isolated the direction in which each feature moved recommendations under controlled conditions. It did not establish permanent weights for Rufus, ChatGPT, Gemini, or every category.

The pages were synthetic Oxford-shirt PDPs, and the transcript does not provide an exact effect size for every feature. Treat the ranking as a hypothesis to test, not a universal priority score.

Models, merchant feeds, live browsing, and image understanding keep changing. The durable principle is to supply trustworthy decision evidence in a structure machines and shoppers can understand.

Reviews are decision evidence

The winning input was the full text of genuine customer reviews on the PDP itself. Sending the agent to a separate review page removes the reviews from the product-level context it is evaluating.

The reviews did not need to be uniformly positive. A natural mix can look more credible and help an agent judge whether the strengths and tradeoffs match a particular shopper.

On Amazon, sellers cannot paste customer reviews into bullets or A+. Improve the product, use Amazon-approved review mechanisms, and use review themes to strengthen the listing content you do control.

Specific answers beat more words

The Q&A block contained roughly the same number of characters as the spam block. Its advantage came from resolving buyer uncertainty, not from making the page longer.

Build questions around real decisions: who the product is for, where it works, what it works with, which constraints apply, and when a shopper should choose an alternative.

Hidden Q&A is a finding, not the playbook

In the simulation, hiding the Q&A from the human shopper did not reduce its effect on the agent. That shows the tested agent could process more text than a person would want to scan.

It does not make invisible SEO copy a safe tactic. Google classifies text hidden only to manipulate search as abuse and requires structured data to represent content visible on the page.

Use accessible accordions or tabs for detailed answers. They keep the PDP scannable while letting shoppers, assistive technology, search engines, and shopping agents reach the same information.

Read Google’s policy on hidden text and its structured-data guidelines.

Do not write off images

Images had little effect in this specific simulation. Joe noted that the tested model may explain the result, and both speakers expect multimodal agents to interpret product imagery more deeply over time.

Keep designing images for humans. Show scale, use, material, compatibility, and proof clearly. Better creative already helps people convert, and those same signals are increasingly readable by machines.

5. Translate the findings into seller-controlled actions

The experiment tested generic PDP features. An Amazon seller and a DTC site owner do not control the same things.

Evidence the agent needs	On an Amazon PDP	On an owned-site PDP
Product identity and facts	Complete category attributes, title, bullets, variations, price, and availability	Keep visible facts, product feeds, and Product schema consistent
Buyer questions	Answer recurring objections in bullets, images, video, and A+	Add visible, accessible Q&A based on real customer questions
Customer proof	Improve the product and use Amazon-approved review mechanisms	Display genuine review text on the product page
Visual evidence	Show scale, use, materials, compatibility, and constraints	Use the same evidence and descriptive image context
Brand credibility	Build a coherent Brand Store, storefront, and external presence	Publish useful brand, category, and product content

Do not manufacture reviews, copy customer language into endorsements, or hide text for machines. The goal is a clearer product record, not a second layer of invisible SEO copy.

6. Build a repeatable evidence loop

Joe has spent years building optimization systems at scale. At Perch and Razor Group, his team used fresh Amazon search-query data to identify meaningful shifts and refine affected listings each night.

The listings were refined automatically, every single night.

Joe explains how that system caught emerging long-tail searches:

The lesson is the cadence, not the automation itself:

1. Review search-query changes, reviews, returns, and support questions.
2. Identify a missing fact or new buyer concern.
3. Update the relevant listing field or creative asset.
4. Measure discovery, CTR, conversion, and returns before the next change.

7. The Amazon seller implementation checklist

1. Confirm category fit. Prioritize products with long research journeys, meaningful specifications, high prices, or several decision criteria.
2. Audit decision questions. List what buyers compare: use case, compatibility, size, material, durability, constraints, and alternatives.
3. Complete structured fields. Keep category attributes, variations, price, and availability accurate and consistent.
4. Turn objections into content. Use review, return, and support themes to improve bullets, images, video, and A+.
5. Protect review integrity. Improve the experience and use Amazon-approved review mechanisms; do not insert or manufacture endorsements.
6. Mirror the evidence off Amazon. On owned pages, use visible Q&A, genuine review text, accurate feeds, and valid product schema.
7. Protect the human experience. Keep the listing scannable and visual. Do not trade conversion clarity for AI-oriented copy.
8. Measure by category. Track AI referrals where available, query visibility, CTR, conversion, and returns. Treat the experiment as a starting hypothesis.

8. How this builds back to the Amazon Formula

This work belongs to the Impressions lever, specifically SEO and AEO. A product cannot earn the click or conversion if the shopping agent never discovers or recommends it.

The funnel also reaches beyond impressions. Reviews, answers, schema, and useful imagery make the PDP clearer for people, so the same work can support conversion after discovery.

As agents sit between brands and shoppers, brand quality becomes more important, not less. Commodity sellers compete on whatever fact the agent can compare. Trusted brands give the agent more reasons to recommend them.