Claude is the fastest-growing shopping engine of 2026. It's also the most under-studied — partly because Anthropic publishes less about Claude's retrieval pipeline than OpenAI or Perplexity do about theirs, and partly because its citation signature is quieter (2-4 sources versus Perplexity's 4-8). This post is the taxonomy we built after logging 600 Claude shopping conversations over Q1 2026, and the schema fields Claude weights for each pattern.
Why Claude needs its own research, not a generic "AI engine" lens
It's tempting to treat "AI shopping engines" as a homogeneous block. They aren't. The three major engines (Claude, ChatGPT, Perplexity) have meaningfully different reasoning patterns, citation formats, and schema field weightings. Optimising across all three is the right strategy — but only if you know what each one actually does.
Our research flow: 600 prompts run against Claude via both the API and web UI, clustered structurally rather than topically. The six patterns we arrived at cover 94% of the volume. The other 6% are edge cases: hyper-long multi-turn negotiations, image-first queries, multi-product bundles. We'll cover those in a future post.
The six prompt patterns
Pattern 1 — Comparison (28% of volume)
"Compare Alora 72 and Uplift V2." "Is Herman Miller or Steelcase better for me?" Claude pulls both products, emits a structured side-by-side with 4-6 attributes per product, and cites the merchant PDPs inline. The attributes it chooses are driven by the overlap in each product's additionalProperty array — if you emit warranty_years, load_capacity, and motor_noise_db, those fields go in the comparison table. If you don't, Claude falls back to generic attributes or omits the axis.
Pattern 2 — Constraint filtering (22%)
"Find a standing desk under $800 with a warranty longer than 5 years." Claude treats each constraint as a hard filter and eliminates candidates that fail any one. This is the pattern where missing or malformed offers.price or missing warranty additionalProperty hurts most — your product simply gets filtered out before Claude ever recommends it. Emit these fields or don't expect to appear.
Pattern 3 — Tradeoff reasoning (18%)
"I want the cheapest option that doesn't sacrifice build quality." This is the most distinctively-Claude pattern. Claude reasons through an explicit Pareto front, names the tradeoff, and recommends the knee of the curve. It weights award and material heavily for build quality signals — plain marketing prose about "premium quality" does almost nothing.
Pattern 4 — Spec verification (14%)
"Does the Alora 72 hold 350 pounds?" "What's the motor decibel rating?" Claude looks for the exact value in the PDP schema, quotes it verbatim (within its fair-use quote bound), and cites. If the value isn't there, Claude says "I can't verify" rather than making it up. This is a zero-marketing-copy pattern: all that matters is whether the fact is in your JSON-LD.
Pattern 5 — Post-purchase validation (8%)
"I just bought the Alora 72 — will it fit a 60-inch-wide desk alcove?" "Does it come with the cable tray?" High-intent traffic from users who already purchased and are checking fitment. Claude pulls dimensions and compatibility from the PDP schema. These users often ask a follow-up buy question ("What accessory do I need for X setup?") which is incremental revenue opportunity.
Pattern 6 — Ranked list (4%)
"3 best mid-range standing desks in 2026." The classic short-list. Smaller share of volume than we expected — shoppers using Claude tend to come with more specific intent than shoppers running generic category queries in Google. For this pattern, Claude leans heavily on independent reviews (Wirecutter, Reviewed) but still cites 1-2 merchant PDPs for specific facts.
- Pattern 1 (comparison) — invest in additionalProperty overlap with direct competitors.
- Pattern 2 (constraint filter) — be rigorous about offers.price and warranty metafield.
- Pattern 3 (tradeoff) — emit award and certifications schema; they're the build-quality proxy.
- Pattern 4 (spec verify) — emit every numeric spec shoppers ask about as an additionalProperty.
- Pattern 5 (post-purchase) — ship dimensions, compatibility, and accessory compatibility in schema.
- Pattern 6 (ranked list) — optimise for independent review coverage and brand authority.
Citation signature — what Claude emits vs. ChatGPT vs. Perplexity
The headline: Claude cites fewer sources (2-4) but anchors each citation to a specific claim. ChatGPT cites more sources (3-6) with inline hyperlinks per claim and embeds product images from the first Product.image URL. Perplexity cites most (4-8) with a numbered footer bibliography and sometimes embeds video.
The signal weighting differs too: Claude weights additionalProperty heavily, ChatGPT weights aggregateRating and Review, Perplexity weights Organization.sameAs and brand-authority signals. The practical consequence is that optimising for the union — shipping the full Schema.org spec, not just the fields one engine cares about — wins on all three.
The Claude-specific action list
Everything on this list is a schema or metafield change — no content-marketing investment required. All five can be shipped in a week by a competent Shopify developer:
- Expand additionalProperty from 2-3 entries to 8-12 per PDP — warranty_years, load_capacity, motor_noise_db, motor_type, frame_material, desktop_material, cable_management, assembly_time, country_of_origin.
- Add award node for every certification — BIFMA, UL, GREENGUARD, FSC — each one is a high-weight Claude signal.
- Emit countryOfOrigin as a first-class node (not just inside a tariff disclaimer) — Claude weights it for furniture and apparel.
- Populate brand.slogan with a one-line positioning statement — Claude surfaces this in comparison patterns.
- Link Organization.knowsAbout to 3-5 product-category topic pages — Claude uses this for pattern 6 ranked lists.
How to measure your Claude citation share
Claude doesn't expose a referrer the way Perplexity does, which makes measurement harder. The approach we recommend:
- Prompt Claude with 15-20 queries covering all six patterns for your category. Use the Claude web UI (not the API) so you see the same experience shoppers see.
- Log whether your PDP appears as a citation, and in which position (1-4).
- Repeat weekly on the same prompt set — drift is the signal, absolute numbers are noisy.
- Compare against a control competitor's PDP for the same queries. Relative movement is the KPI.
Surfient's GEO Score runs this test across Claude, ChatGPT, Perplexity, and Google AI on a rolling weekly cadence. If you're running a small set of products the manual approach is fine; if you're tracking hundreds of PDPs it isn't scalable.
Closing
Claude rewards merchants who treat their PDP as a structured-facts document first and a marketing surface second. That's a very specific discipline, but it's the same discipline that wins the other AI engines too. The six patterns above are the lens — the fix list is the schema expansion you should ship this quarter.