AI Search Visibility for Commerce: Getting Your PDPs Cited by ChatGPT, Perplexity & Gemini
Shoppers are asking assistants what to buy, and the assistant answers from sources it can read and trust. This playbook covers how to make a PDP citable: structured data the model can parse, named and quoted reviews it can lift, an llms.txt that points it at your best evidence, and content built to be quoted rather than ranked.
- 19 pages
- 15 min read
- For: seo lead, ecommerce leader, cmo
- Cited in AI search
- Star eligibility
- Tracked vs holdout
Playbook · 19pAI Search Visibility for Commerce: Getting Your PDPs Cited by ChatGPT, Perplexity & Gemini
What you’ll learn
- AI assistants answer from sources they can parse and trust, so a citable PDP beats a merely rankable one
- Emit clean Product, Review and AggregateRating JSON-LD so the model has structured facts to lift, not prose to guess at
- Named, dated, quoted reviews are the most citable evidence on a PDP: real attribution is what makes a model willing to quote it
- Publish an llms.txt that points assistants at your highest-evidence pages instead of leaving them to crawl blind
- Write the answer the shopper is asking for, in the words they ask it, so the model can lift a clean passage
Chapter previews
- Chapter 01
Citation is the new ranking
A shopper asking an assistant what to buy never sees a SERP. The assistant cites a handful of sources, and the job is to be one of them, which is a different game from ranking blue links.
- Chapter 02
Structured data the model can parse
Product, Review and AggregateRating JSON-LD give the assistant clean facts to lift instead of prose to interpret. Clean markup is the cheapest visibility you can buy in AI search.
- Chapter 03
Reviews as citable evidence
A named, dated, quoted review is the most quotable thing on a PDP. Real attribution is exactly what makes a model comfortable citing it.
- Chapter 04
llms.txt and pointing the model
A simple file that points assistants at your best-evidence pages, so they spend their crawl budget on what you most want quoted.
- Chapter 05
Content LLMs actually lift
Answer the question the shopper is asking, in their words, in a clean passage the model can quote without rewriting. Build for the lift, not the ranking.
- Chapter 06
Measuring AI-search presence
Track citations and referral traffic from assistants, and the share of your priority PDPs that are quoted, as a compounding asset rather than a one-off win.
Inside the playbook
A growing share of buying decisions now starts with a question to an assistant rather than a query in a search box. "What is the best foundation for oily skin," "which running shoe lasts longest," "is this jacket warm enough for winter." The assistant does not return ten blue links for the shopper to sift. It answers, and it cites a handful of sources it could read and trust. If your PDP is not one of those sources, you are invisible at the exact moment the decision is being made, no matter how well you rank in classic search.
Being citable is a different discipline from being rankable. A model does not reward keyword density or backlink counts in the way a search algorithm did. It rewards pages it can parse cleanly, facts it can verify, and evidence it can attribute. That means structured data it can read without guessing, reviews with real names and dates it is willing to quote, a clear signal pointing it at your best pages, and prose written to be lifted rather than ranked. This playbook is how you build each of those.
~88%
of shoppers consult reviews before a purchase decision, the evidence assistants lift
Representative range, Bazaarvoice / Bizrate Insights shopper surveys
rising
share of product research starting in AI assistants rather than a search box
Representative, flagged directional: eMarketer / Google commentary on AI search adoption
structured
data is what lets a model parse facts instead of guessing from prose
Per Google structured-data and rich-results guidance
named + dated
reviews are the most citable evidence a PDP can carry
Representative principle, aligned to Bazaarvoice review-integrity guidance
Rankable PDP
Tuned for keyword coverage and backlinks so it climbs the ten blue links a shopper used to scroll.
Wins at
- Familiar SEO discipline
- Works where classic search still drives traffic
- Measurable in rank trackers
Struggles with
- Invisible when an assistant answers instead of listing
- Prose the model has to guess at, not parse
- No attributable evidence to quote
- Nothing pointing crawlers at the best pages
Citable PDP
Built so an assistant can read the facts, trust the evidence and lift a clean passage into its answer.
Wins at
- Parseable Product + Review + AggregateRating markup
- Named, dated reviews the model will quote
- llms.txt pointing at the highest-evidence pages
- Question-led prose written to be lifted
Struggles with
- Needs real, attributable review evidence
- Markup has to validate, not just exist
- A newer discipline than classic SEO
Same product, two different optimisation targets. One is built to climb a list of links, the other to be quoted in an answer.
Where each tactic sits on citation lift and effort
The tactics, and what each one changes
AI-search visibility is not one switch. It is a small stack of changes, each of which alters a different thing the model relies on, and shows up in a different place. Read the table as a build order: structured data first, because it is the cheapest and most parseable signal, then the evidence and pointing layers on top.
| Tactic | What it changes | Where it shows |
|---|---|---|
| Product + AggregateRating JSON-LD | Gives the model clean, structured product facts and a rating to cite instead of prose to interpret | Rich results in classic search, parseable facts in AI answers |
| Named, dated, quoted reviews | Supplies attributable evidence the model is willing to quote directly | Pulled into assistant answers as cited customer proof |
| llms.txt pointing at best pages | Directs assistant crawlers to your highest-evidence content first | Influences which of your pages the model reads and cites |
| Question-led, liftable passages | Provides a clean answer in the shopper's words that the model can quote without rewriting | Quoted passages in assistant responses |
| Clean, fast, parseable HTML | Lowers the cost for a crawler to read and extract the page | Improves the chance the page is read at all |
Reviews as citable evidence
The most quotable thing on a PDP is a real customer saying something specific, with a name and a date attached. A model is far more willing to lift "verified buyer, March 2026: held up through a wet winter commute" than an unattributed marketing claim, because the attribution is what makes the quote defensible. This is also where the content-integrity line is sharpest: every review you expose as evidence must be a genuine, attributable quote with a real date. Synthesising reviews to pad an aggregate count does not just break trust, it poisons exactly the signal you are trying to build, because the assistants increasingly weight attribution and consistency.
The same structured review substrate does double duty. It earns star ratings in classic search and feeds assistants the customer evidence they quote. Build it once, cleanly, and it pays out across both surfaces from a single effort.
Is your PDP actually citable?
Before chasing citations, run the page through the test a model effectively applies: can it read the facts, trust the evidence, and find a clean passage to quote. The decision tree below is the same triage in question form. Most PDPs fail at the first gate, the markup, which is also the cheapest to fix.
Is your PDP citable?
Start here
Can an assistant parse clean Product + Review + AggregateRating JSON-LD on this page?
- No, the markup is missing or broken
Fix the structured data first.
Without parseable markup, the model is guessing from prose, and guessing is exactly what it avoids citing. Emit valid Product, Review and AggregateRating JSON-LD and validate it before anything else.
- Markup present but failing validation: Fix the validation errors: invalid markup is treated as absent
- AggregateRating with no real underlying reviews: Do not fake the count: collect real reviews first
- Markup valid and passing: Move to the evidence gate
- Yes, the markup is clean
Now check the evidence and the prose.
Parseable facts get you read. Attributable evidence and a clean, question-led answer get you quoted. This is where most of the citation lift actually comes from.
- Reviews are unnamed or undated: Surface named, dated, verified reviews the model can attribute
- No passage answers the shopper's actual question: Add a clear, liftable answer in the shopper's own words
- No llms.txt pointing to this page: Publish llms.txt and list your highest-evidence PDPs
AI-search readiness: how citable is your catalogue today
- 1
Invisible
You’re here ifNo structured data, or markup that fails validation. Reviews are unnamed and undated. The model has only prose to guess at.
Next moveEmit valid Product + Review + AggregateRating JSON-LD on the priority PDPs and validate it.
- 2
Parseable
You’re here ifClean, validating markup gives the model facts to lift, but the evidence is thin and no passage answers the shopper's real question.
Next moveSurface named, dated, verified reviews and write a liftable question-led answer.
- 3
Citable
You’re here ifFacts parse, reviews are attributable, and an llms.txt points crawlers at the highest-evidence pages.
Next moveTrack which PDPs get cited and treat cited share as a measured asset.
- 4
Compounding
You’re here ifPriority PDPs are cited by default, AI referral traffic is monitored, and the same review substrate earns stars in classic search too.
Next moveRun a continuous citability audit by lift-over-effort across the catalogue.
“In AI search the shopper never sees your ranking. They see whatever the assistant chose to cite, and the only way in is to be worth quoting.”
The 30-60-90 day plan
Becoming citable is a build order, not a single launch. This is the cadence that takes a catalogue from invisible to cited, leading with the cheapest, most decisive gate: the markup.
From invisible to cited in 90 days
- 01
Days 1-30
Audit the priority PDPs for citability, emit and validate Product + Review + AggregateRating JSON-LD, and fix the markup gate first. Never synthesise reviews to pad an aggregate count.
Markup gate
- 02
Days 31-60
Surface named, dated, verified reviews as quotable evidence, add a question-led liftable passage to each priority PDP, and publish an llms.txt pointing at the highest-evidence pages.
Evidence + prose
- 03
Days 61-90
Monitor AI referral traffic and which PDPs are cited in answers to the questions that matter, then extend the same markup-evidence-prose pattern down the catalogue by priority.
Measure + extend
Pointing the model, and measuring presence
An llms.txt file is a simple, low-cost way to tell assistant crawlers where your best evidence lives, so they spend their attention on the pages you most want quoted rather than crawling blind. Pair it with measurement: track referral traffic from assistants, monitor whether your priority PDPs are being cited in answers to the questions that matter, and treat the share of cited pages as a compounding asset. Visibility in AI search builds slowly and then holds, the same way organic equity did, so the teams that start cleanly now are the ones quoted by default later.
Sources and further reading
Free PDF, straight to your inbox.
Drop your email and we’ll send the full 19-page playbook now. We never sell your address. Unsubscribe in one click.
- AI assistants answer from sources they can parse and trust, so a citable PDP beats a merely rankable one
- Emit clean Product, Review and AggregateRating JSON-LD so the model has structured facts to lift, not prose to guess at
- Named, dated, quoted reviews are the most citable evidence on a PDP: real attribution is what makes a model willing to quote it
