The Shopify fields nobody filled in

Why the count you are looking at is wrong

Before doing any of this we check how big the gap is. The obvious way is to ask Shopify how many products have a product type. That query is product_type:*, and it is a presence filter — it matches the empty string. On a catalogue where every single value was blank, it reported 100% healthy.

Four hand-checked samples proved it wrong in about a minute. That is the whole lesson: an aggregate count tells you a field exists, never that it contains anything. We dump the entire catalogue to a local file and read real values instead.

It is also why we can tell you the true gap for free before you commit to anything. The dump is an hour of machine time, not a consulting engagement.

The character budget is smaller than you think

Google shows roughly 60 characters of a title. Most Shopify themes append – Your Shop Name to every page, which on a typical store eats about 13 of those 60 before you have written a word. Write two thousand titles at 60 characters and every one of them is truncated in the result.

So the theme comes first. On the last rebuild we added one condition in layout/theme.liquid so the suffix is dropped on product pages only — collections, blog and homepage keep it, because there the shop name is the useful part. That single line returned 13 characters to every product title in the catalogue.

Decide the suffix before writing the titles, not after. Doing it in the other order means writing them twice.

Your SEO title is your Merchant Center title

This is the one that costs money. Shopify's Google channel sends seo.title as the feed title, not the product title. So a promotional SEO title is not a style choice — it is feed data.

"Buy X Online in India" reads to Google as non-product data and gets offers disapproved. On the last catalogue we found 38 products with promotional openers already live in the feed, and rewrote them. Nobody had connected the two systems, because in the admin they look like separate features.

The same rule kills most of the title formulas an agency will hand you. If it would look wrong on a Shopping card, it is wrong in seo.title.

The traps that make this skilled work rather than a bulk edit

Shopify replaces the whole seo object wholesale. Send a seo.description on its own and seo.title is silently set to null. Every write has to carry both fields. Our apply script refuses a partial seo object before it makes a network call, because by the time you notice you have wiped two thousand titles.

A wrong-but-real product id does not error, it writes to the wrong product. So every mutation echoes the handle back and the script compares it to what it meant to update.

Rule engines fail quietly. A singular-only pattern for filter missed every plural. A vendor whose name contained a category word filed that brand's adapter rings as filters. Instant film was classified as a camera. None of these show up in a summary total — they are only found by printing every row that matched a pattern and landed somewhere unexpected.

Collections created through the API are not published. They return a 200, they look correct in the admin, and they 404 on the storefront until they are explicitly published to the Online Store channel.

Duplicating a theme is asynchronous. It returns an id immediately, then keeps copying files for about twenty seconds. Write in that window and the file verifies clean — byte for byte — and is then silently overwritten by the background copy. That one cost us two publish cycles before we understood it, and it is now a poll-and-verify-twice step in the process.

The redirect map, and why fuzzy matching is dangerous

If the store moved platforms, the old URLs are usually still the ones Google sends people to. On one migration, 675 product URLs were serving 404s while still earning 67% of the site's search clicks. Nothing on the new store was wrong. The traffic simply arrived at nothing.

The obvious way to rebuild that map is string similarity between the old slug and the new one. It is also the wrong way. At a 0.82 similarity cutoff, the matcher we tested paired a Nikon Z6 III with a Z7 II, and a 35mm lens with an 85mm — both well above the threshold, both sending a buyer to the wrong product.

So matching runs on model tokens rather than overall similarity, and anything that genuinely needs a human decision is handed back as a list to be worked through, not guessed at. A redirect to the wrong camera is worse than a 404, because the 404 at least tells the truth.

Where we stop, and why

We do not guess. On the last catalogue, 23 products were genuinely ambiguous and were left with no product type rather than given a wrong one, because a wrong type is worse in a feed than an absent one.

Existing hand-written copy is read first and sent back unchanged. Every store has a handful of products someone actually wrote about, and overwriting those with generated text is the fastest way to make a rebuild a net loss. On the last one there were 74, and all 74 came through untouched.

We also say when something has to come off. On one store we found a fabricated "rated 4.5/5 from 650+ reviews" strip with stock-photo avatars sitting on every product page — while a real review app was already installed and rendering underneath it. That came off every page. Invented review claims are a legal exposure, not a conversion tactic.

Brand collections were built only for vendors with at least 20 products. 87 of 178 vendors had fewer than five, and a collection page with four items on it is a thin page, not an asset.