7 Shopify prompt classes ChatGPT actually cites

Between November 2025 and March 2026 we logged 3,912 commerce prompts to ChatGPT, Perplexity, Claude, and Google AI Overviews — sampled from merchant analytics, observed shopper traces, and a panel of 240 US and UK consumers who agreed to let us inspect their assistant history for two weeks. Seven intent classes cover 96% of them. The remaining 4% are typos, abandoned half-prompts, and one-off edge cases (someone asked Claude for a poem about their cat's dietary restrictions). For practical purposes, there are seven.

If you run a Shopify store, this is the taxonomy your product copy has to answer. Not one of the seven. All seven. Because assistants don't pick a favourite prompt type — shoppers do — and the ones you haven't written for are the ones where a competitor steals the citation. For the underlying retrieval mechanics we covered in an earlier post, see how ChatGPT actually cites commerce content; this piece is about the intent side of that same pipeline.

What we sampled and why it matters

The panel skewed toward considered purchases ($60+ AOV, non-impulse): furniture, small appliances, apparel with a fit concern, skincare, pet nutrition, replacement parts, outdoor gear. We excluded impulse categories (trending beauty, fast fashion, branded snacks) because those queries mostly still go to TikTok and Instagram search — the assistant funnel hasn't absorbed them yet, and OpenAI's own shopping rollout notes describe the same mid-to-high-consideration skew.

We also excluded research-only prompts (“how does a DC motor work”) and pure-navigation prompts (“Wayfair returns policy”). The 3,912 remaining prompts are what the industry calls buyable — the shopper is, in that moment, willing to be told which specific product to purchase. Every sentence the assistant writes back is either a citation Shopify stores can win or a citation they've already lost. If you want a refresher on how we measure citations at all, start with citation rate and share of AI voice in the glossary.

The seven intent classes, at a glance

Before we unpack each one, here is the distribution across the sample. The chart is the same data the rest of the post references — a screen reader will read the labels and values verbatim, and so will a retriever that scans the HTML.

Each class is defined by the ranking cue the assistant uses to pick a product — the specific sentence shape it prefers to quote. Miss the ranking cue and your PDP is indistinguishable from every other PDP in your category, and the model defaults to whichever brand it has the longest training-data history with (usually not yours).

1. Product discovery — “best X” (28.4% of the sample)

The starter prompt. “What's the best standing desk?”, “best indoor rower 2026”, “best cast iron skillet”. The shopper has a category in mind and wants a shortlist of three to five brands with a sentence each. This is the only class where assistants will freely name brands without a constraint — every other class narrows the answer set. Our engine-specific playbooks go deep here: rank Shopify in ChatGPT and rank Shopify in Perplexity.

Ranking cue: a self-contained sentence in the opening of your PDP that names the category, the shopper type, and one concrete differentiator. Assistants quote this sentence almost verbatim.

2. Comparison — “X vs Y” (17.1%)

“Uplift v2 vs Fully Jarvis”, “Allbirds vs Vivaia”, “Samsung Bespoke vs LG InstaView”. Two brands named; the shopper wants a structured tradeoff. Assistants will cite any page that explicitly compares the two, and they strongly prefer honest comparisons over marketing ones — if your page only says nice things about your own product, the retriever skips it in favour of a review site. The comparison page playbook covers the exact layout we recommend.

Ranking cue: a comparison table with 5–7 rows, one of which you willingly lose. The losing row is what makes the table quotable.

3. Budget-constrained — “best X under $N” (15.2%)

“best mechanical keyboard under $100”, “gaming chair under $400”, “espresso machine under $600”. The highest-conversion class in the sample. The shopper has already decided to buy — they're filtering. This is also the class Shopify stores under-index for the worst, because most PDPs don't state price as a first-paragraph fact. You put the price in the product metafield and on the buy button, and you leave it out of the prose. Assistants can't cite what isn't in the prose. Fix this at the data layer with a clean Shopify product feed for AI and at the copy layer with AI-first product descriptions.

Ranking cue: the price, stated as a sentence, in the opening paragraph. “The Surfient Desk Pro is $549 as of March 2026, which puts it in the upper end of the sub-$600 range.” Feels weird to write. Wins citations.

4. Use-case — “X for Y person / scenario” (14.0%)

“standing desk for a 6'3" engineer”, “yoga mat for a sweaty vinyasa practice”, “running shoes for wide feet and high arches”. Shoppers translate their own situation into natural language and expect the assistant to do the translation work. This class is where sub-optimised PDPs get crushed by Reddit threads — a random Redditor's bodyweight + their shoe size + “I love these” is more quotable than a 600-word product description written in marketing voice. See long-tail questions for ecommerce AI for the question shapes and how Shopify reviews feed AI citations for the social-proof half.

Ranking cue: an explicit “this is for…” sentence on the PDP that names the shopper's situation in their words, not yours.

5. Feature-specific — “X with Y feature” (11.3%)

“standing desk with memory presets and cable management”, “headphones with 30+ hours of battery”, “washing machine with a self-clean cycle”. Shoppers have done their research and are filtering on two or three non-negotiables. Assistants almost always return a list of three products here, with each product matched to two of the three features — they'll rarely claim a product meets all three unless the PDP says so outright. This is also where Schema.org Product markup earns its keep — see our Product JSON-LD guide for Shopify.

Ranking cue: a bulleted feature list in the opening fold where each bullet restates the feature name verbatim. Not “pre-set heights” — write “memory presets”, which is what the shopper typed.

6. Brand evaluation — “is [brand] good for X?” (6.8%)

“is Branch furniture worth the money”, “is Rothy's still good in 2026”, “is Purple mattress worth it for back pain”. The shopper already heard the brand somewhere (an ad, a friend, a prior ChatGPT answer) and is double-checking. A shockingly high percentage of these — 41% in our sample — are followed by a purchase within 48 hours. You want to win this one cleanly. This is the class where monitoring gets real — run monitor AI brand mentions weekly, not monthly.

Ranking cue: a review-style page about your own brand that names the tradeoff honestly. “Branch chairs are expensive for the category and fit best for home offices with >6 hours of seated use per day.” Assistants quote this sentence. The brand that writes this page wins the citation. The brand that writes “Branch chairs are the highest-quality office chairs on the market” gets skipped for a random blog.

7. Replacement — “alternative to X that [constraint]” (3.1%)

“alternative to IKEA Bekant that lasts longer”, “alternative to Stanley cups that don't leak”, “alternative to Jordan 1s that fit wide feet”. Small class, high intent. The shopper has a grievance; the first product the assistant names tends to win.

Ranking cue: a sentence that names the competitor and the specific shortcoming your product fixes. Yes, this means mentioning your competitor by name on your own PDP. It's fine. The assistant was going to mention them anyway.

The PDP pattern that covers all seven

Rewriting every PDP for every class is not practical. The pattern that covers all seven in a single opening fold is a four-paragraph structure we've seen work on 42 of the 44 Shopify stores we've migrated onto it so far — and which our anatomy-of-an-AI-cited-PDP teardown walks through line-by-line.

• P1 — One sentence: category + shopper + concrete differentiator (covers class 1).
• P2 — One sentence: price (in dollars, in prose) + the shopper segment it's priced for (covers class 3).
• P3 — A 3-item bullet list of features, each restated as the shopper would phrase it (covers class 5).
• P4 — A 'this is for' sentence naming the use case + a 'this is not for' sentence naming who should buy something else (covers classes 4 and 6).
• In a separate 'Alternatives' block below the fold — name 2 competitors and one tradeoff per competitor (covers classes 2 and 7).

Why this pattern holds across four different engines

The ranking cues above aren't engine-specific. We tested the same PDPs across ChatGPT, Perplexity, Claude, and Google AI Overviews — and while each engine has its own retrieval stack, the sentence shapes it prefers to quote converge. That's because every major assistant depends on the same three raw signals: a crawlable HTML body, JSON-LD structured data it can parse, and (where available) a machine-readable brief like llms.txt for Shopify.

Perplexity in particular is explicit about preferring shopping-ready data over marketing copy — their own shopping launch post calls out structured pricing and specifications as first-class ranking inputs. Google has been making the same argument about AI Overviews since the Search Central blog started covering generative results. For Shopify merchants, the Surfient features that most directly target these signals are the GEO audit engine (diagnosis) and prompt intelligence (measurement).

Measurement — how to tell the pattern is working

Because assistants mostly don't pass a referrer, you can't measure this in GA4. You measure it the way every GEO-aware team now does: maintain a prompt library of 40–60 commerce questions in your category (ten per intent class, minus class 7 which has fewer queries), run them weekly in ChatGPT, Perplexity, Claude, and AI Overviews, and log which brands each assistant names first. The full methodology is in our AI visibility metrics guide.

Three weeks after you ship the four-paragraph PDP pattern on your top 10 products, the share-of-voice on your priority prompts should move by at least 4 percentage points on one engine. If it moves on all four, you've shipped it correctly. If it moves on none, the prose is still too marketing — rewrite P1 as if you were describing the product to a friend in a single breath.

What to do this week

If you only have five hours this week, spend them on this:

1. Pick your top 3 revenue PDPs. Rewrite P1 and P2 using the four-paragraph pattern. The price has to be in the prose, in dollars, with the date. Yes, it's ugly. Do it anyway.
2. Ask ChatGPT the seven canonical prompts for your category — one per class. Note which brands it returns. That's your baseline. Repeat weekly using the panel format from the AI visibility metrics guide.
3. Add an “Alternatives” block below the fold on the same three PDPs. Name two competitors by name and state one tradeoff each — the tradeoff you'd concede in a sales call anyway.
4. Ship an FAQPage-schema block with the three questions your support team actually receives every week. Use exact shopper phrasing in the question field.

The first mile moves the needle because most of your competitors are still optimising for class 1 and assuming the rest will take care of itself. They won't. The assistants are asking seven questions, not one, and they're citing the stores that answer all seven.