Writing Ecommerce Meta Descriptions at Scale

Meta descriptions at catalogue scale pose a problem content sites never face: five thousand products and four hundred categories all need the snippet text that sells the click, and nobody is writing five thousand of anything by hand. The answer is a tiered system — templates that render real product data for the volume, human copy for the money — governed by one honest calibration: descriptions don't rank pages (they never have), they sell clicks on rankings you already hold, which prices the effort correctly and aims it at the pages with impressions. Here's the system.

The calibration first: what metas actually do

Not a ranking input — Google has said so for a decade, per the myths file — and frequently rewritten: Google substitutes its own snippet (pulled from page content matching the query) a large share of the time, more on long-tail queries. What your written meta controls: the default presentation on your primary queries, and the click-sell on the SERPs where CTR is the contest, per the packaging economics. Consequence: metas are a CTR project for pages with impressions — the high-impression list is the priority queue, and pages Google always rewrites for are pages whose meta effort was already refunded.

The tiered production system

1. Template tier (the long tail): attribute-driven generation from real product data — "[Product]: [key attribute], [benefit hook]. [Shipping/guarantee hook]." Rendered per-product from structured fields, it's honest variation (the data-not-spin rule applied to snippets); one template edit tunes thousands of SERO presentations. Category version: "[Category] — [selection breadth hook]. [Trust hook]."
2. Human tier (the money): top categories, hero products, comparisons and guides get written metas — 150ish characters of ad copy: the differentiator, the objection-handler (shipping, returns, range), a reason to click this result, per the specificity bias. The store's whole meta effort budget concentrates here, sized by the impression data.
3. The hooks inventory (what goes in either tier): the claims that move commercial CTR — price/value positioning, shipping speed and thresholds, guarantee/returns, stock breadth, expertise signals ("tested by...") — rotated by what the store can honestly claim; unsubstantiated superlatives read as noise to shoppers who've seen ten SERPs today.

The maintenance loop

Quarterly, inside the CTR mine: high-impression/low-CTR pages get meta (and title) rewrites with dated logging, read at four-plus weeks; template-tier tuning when the store's hooks change (new shipping threshold = one template edit = sitewide snippet update); and the truncation/accuracy sweep in the audit — metas citing dead promotions or old prices are micro trust-leaks at scale. Duplicate-meta warnings from crawlers get triaged sanely: duplicates across canonicalised variants are noise; across distinct indexable pages they're the template tier missing a differentiating field.

Frequently asked questions

Google rewrites most of our metas anyway. Why bother?

Because the rewrite rate concentrates on long-tail queries where your meta couldn't match anyway, while your primary commercial queries — the impressions that pay — show the written version far more often. Write for the head, template the tail, and let the rewrite machinery serve the queries you never targeted; that division is the whole system.

Should metas contain keywords for ranking?

For bolding, mildly: query terms in the snippet get emphasised, which aids scanning — a presentation effect, not a ranking one. Write for the click with the natural vocabulary the research surfaced and the bolding takes care of itself; stuffing a non-ranking field is effort with negative return.

Can AI write the human tier?

It drafts hooks competently and converges on generic ones at scale — the same supervision economics as every content tier: AI for volume drafts, human selection for the money pages where one sharpened claim moves real CTR. The clicks either way land on positions earned by the usual inputs — pages built right and the domain's earned authority (our line).