Why AI Assistants Recommend Some Shopify Stores and Not Yours

AI answer engines can only recommend a shop they can quote. Why product pages built from photos and spec tables stay invisible — and what makes a catalogue legible to ChatGPT, Perplexity and AI overviews.

By Roman Korim · Iteractive.ai · 2026-08-23

Ask ChatGPT, Perplexity or Google's AI overview for "the best merino base layer for winter cycling" and you get a short answer naming two or three shops. Ask again tomorrow and it is usually the same two or three. If your store is not among them, the reason is rarely that your products are worse. It is that the assistant had nothing of yours to read.

An answer engine can only recommend what it can quote

A search engine ranks pages. An answer engine does something different: it writes a sentence and then attaches a source to it. That difference decides who gets named.

To attach a source, the model needs a passage it can point at — a stretch of text that actually says the thing the user asked about. A product page built the way most Shopify pages are built gives it almost nothing to work with:

• a product title,
• three to five photographs,
• a specification table (weight, material, sizes),
• and two sentences of description written to fit a card layout.

Everything a buyer actually wants to know — how it behaves at -5 °C, whether it holds odour after three days, whether to size up under a jacket — lives in the photographs, in the founder's head, or in a review thread somewhere else. None of it is text on your domain. So when the model composes its answer, it quotes the shop that did write those sentences down.

This is the whole mechanism, and it is unglamorous: no text, no citation.

Why catalogue size works against you

The problem compounds in a way that punishes exactly the stores that should win.

A shop with 200 SKUs has 200 pages that each need real prose to become quotable. A shop with 12 SKUs needs 12. The larger, more specialised catalogue — the one with genuine depth, the one that deserves to be recommended for the obscure query — is the one that runs out of writing capacity first.

So the thin, well-marketed store with a dozen products and a copywriter ends up being the one the assistant knows how to describe. Not because it is better. Because it is legible.

What being legible actually requires

Three things, in this order.

1. Prose on the page, in the buyer's language

Not keyword stuffing — sentences that answer the questions a buyer asks before purchase. If a customer emails you "will this fit under a shell jacket", that email is a content brief. The answer belongs on the product page as a sentence, because that sentence is what a model can lift.

Language matters more than most stores expect. A Hungarian buyer asking a Hungarian question gets an answer assembled from Hungarian sources. English-only product copy is invisible to that query, however good it is.

2. Structure the machines already agree on

Schema.org Product markup, honest availability and price fields, and a stable URL per product. Structured data does not make a model like you — it makes your claims unambiguous, so the model does not have to guess whether "£49" is the price, a discount, or a shipping threshold. Ambiguity is a reason to cite someone else.

3. Recency that is real

Answer engines lean on freshness signals when a query implies a current state ("best… in 2026", "in stock"). Recency has to be genuine: republishing the same page with a new date and no new sentences is a signal that decays, and it costs you the stable search record the page had already earned. A page that keeps changing its title never settles anywhere.

The part nobody sustains

None of the above is a secret, and none of it is technically hard. What defeats stores is the cadence.

Being quotable is not a project with an end date. Every new product needs its prose. Every seasonal shift changes which questions buyers ask. Every language you sell into needs its own sentences, not a translation of your homepage. Do it for six weeks and stop, and your catalogue slowly returns to being a wall of photographs.

This is the gap iteractive.ai was built for. Autopilot reads your existing catalogue, works out what each product actually needs said about it, and then keeps publishing — product prose, articles and social posts, in the language your customers buy in — on a schedule, without you writing the brief each week. You approve the direction once; the writing continues.

How to check where you stand right now

You do not need a tool to get the first read. Open an AI assistant and ask it, in the language your customers use, the question a buyer would ask before finding you — a category question, not your brand name. Then ask "which shops sell this?"

If it names competitors and not you, read what it says about them. You will usually find it is quoting a sentence those shops wrote and you did not.