The Shopify GEO checklist: what to fix before you chase citations

Most GEO advice starts at the content layer: write better answers, add FAQs, structure your headings. That advice is fine and roughly last in the order of operations. If an assistant's retrieval step never reaches your page, or reaches a version of it with no price in the HTML, the quality of your prose is irrelevant.

Here is the order we actually work in on a Shopify store, and why each item is on the list.

Retrieval: can the model get the page at all?

1. Decide crawler access separately for search, user-triggered retrieval and training. OpenAI identifies OAI-SearchBot as its search crawler and GPTBot as a training crawler; Googlebot controls access for Google Search AI features. Review the current documentation for each provider before changing robots.txt.
2. An llms.txt can provide an optional, concise map for tools that choose to read it. It is not a demonstrated ranking factor or a requirement for Google AI features. Crawlable HTML, internal links and an accurate sitemap come first.
3. Check what renders without JavaScript. Fetch your PDP with JS disabled. If the price, title, availability and variant list are not in that HTML, assume a share of retrieval passes never see them.
4. Kill the duplicate product path. Shopify serves the same product at /products/handle and /collections/x/products/handle. Canonicals usually handle it; internal links frequently do not, and the split wastes crawl budget on a large catalogue.
5. Get faceted navigation under control. Filter combinations generate crawlable URLs at a combinatorial rate. Pick the facets worth indexing, block the rest at the parameter level, and stop linking to the ones you blocked.

Trust: will the model state your facts as facts?

1. Reconcile schema with your Merchant Center feed. When your Product schema says $129 and the feed says $149, you have taught two systems that your data is unreliable.
2. Build the schema as a graph, not a pile. Organization, Product, Offer, AggregateRating, Breadcrumb and FAQ nodes that reference each other via @id read as one coherent entity.
3. Make identifiers consistent. GTIN, MPN and brand should match across catalogue, feed, schema and marketplace listings. Inconsistent identifiers are why a model hedges instead of naming you.
4. Consolidate the brand entity. One name, one description, one set of facts across your site, your directory listings, your marketplace profiles and Wikidata if you qualify.
5. Earn corroboration where your category already gets cited. Run the prompt set, look at which domains the assistants actually pull from, and work on those surfaces. It is usually three or four sites, and it is rarely the ones your PR agency pitched.

Extraction: is there anything worth quoting?

1. Write self-contained passages. A model lifts a paragraph, not a page. Every answer paragraph should survive being read with no surrounding context.
2. Put specifications in a real table, with units. Prose specs are hard to extract; a table with explicit units is trivially parseable and far more likely to be quoted accurately.
3. State who the product is not for. Negative qualification is rare, genuinely useful, and disproportionately quoted — models favour sources that make discriminating claims.
4. Answer the questions your support inbox already receives. Mine tickets, reviews and on-site search rather than a keyword tool. Those are the phrasings shoppers actually use with an assistant.

What to measure once it is done

Define a prompt set of real buying questions in your category — a hundred is plenty to start — and run it against each assistant on a schedule. Track how often you are named, how often you are linked, and which domains take the citations you did not. That baseline is the only way to tell whether any of this worked, and it is worth capturing before you change anything.

Put this into practice with Shopify SEO and Shopify GEO.