Skip to main content
Autocomplete curation is currently in beta. The dashboard page is marked with a Beta badge. Behavior may change before general availability.

​
Overview

Layers groups customer search queries into clusters. Each cluster has a canonical text (for example, "running shoes") and a set of raw queries that rolled up into it. By default, any cluster that passes frequency and relevance thresholds can surface as an autocomplete suggestion. Autocomplete curation adds a second pass on top of that. Layers sends every candidate cluster — along with your store description and a natural-language brand prompt — to an AI model. For each cluster, the model decides:
  • Whether the suggestion should appear at all (allowed).
  • Whether the display text should be rewritten for customers (display_label).
Suppressed clusters disappear from the live autocomplete response. Re-labelled clusters still match the same underlying queries, but are displayed with the curated label.

​
When to use it

Use autocomplete curation when you want more editorial control over the typeahead without hand-managing every suggestion. Common scenarios:
  • Remove off-brand suggestions. Hide queries that surfaced from historical data but don’t fit your catalog or voice.
  • Standardize casing and phrasing. Force "nike" to display as "Nike", or collapse "mens tshirt" into "Men's t-shirts".
  • Clean up plurals and typos. Prefer "dress" over "dresss" as the customer-facing label while still matching both.
  • Enforce brand guidelines. Block suggestions that conflict with seasonal or legal guidance (for example, claims you can’t make about a product category).
If you only want to stop a handful of queries from appearing, a semantic redirect or manual block may be simpler. Autocomplete curation is most useful when you want rules that apply consistently across hundreds of suggestions.

​
How it works

Curation runs ahead of time rather than during each customer request. The live Autocomplete API returns the latest saved curation decisions for each cluster.
  1. Layers selects clusters that pass a minimum frequency threshold and have not yet been curated for the current version of your prompt and description.
  2. Layers evaluates those clusters using your store description and brand prompt.
  3. For each cluster, Layers decides whether the suggestion should be shown and whether its display text should change.
  4. Layers saves those decisions so the next autocomplete request can use them.
The curation version is a hash of your store description plus your brand prompt. Changing either field invalidates every cluster’s decision and triggers a full re-curation on the next curation run.

​
When curation runs

Curation is triggered automatically in three cases:
  • When you save a new brand prompt. If the prompt or description changes, Layers starts a new curation run for your store.
  • When you click Re-curate. This re-runs curation for every eligible cluster, regardless of version.
  • Daily, at 09:00 UTC. Layers re-curates any stale clusters for every store that has either a brand prompt or a description set.
Newly created clusters are curated as soon as they reach the frequency threshold, so new suggestions do not bypass your guidance.

​
What the customer sees

After curation runs:
  • Suppressed suggestions are hidden from Autocomplete API responses entirely.
  • Re-labelled suggestions appear with their curated text in query_text. Matching is unchanged — a customer typing "nike" still reaches the same cluster, they just see "Nike" in the dropdown.
  • All other relevance, stem deduplication, and semantic redirect behavior continues to apply on top of curation.

​
Writing a good brand prompt

The brand prompt is free-form text, up to 2,000 characters. The model already receives your store description, so the prompt should focus on editorial rules rather than restating what you sell. Good prompts are specific and rule-based. Examples:
  • “Capitalize brand names (Nike, Adidas, Lululemon). Suppress suggestions that mention competitor brands we don’t carry.”
  • “Prefer singular nouns in the display label (dress, not dresses). Suppress suggestions that are only a misspelling of another suggestion.”
  • “This store is kosher-certified. Suppress any suggestion referring to pork or shellfish, even if customers have searched for it.”
  • “Use sentence case for multi-word suggestions. Suppress any suggestion that looks like a SKU or part number.”
Avoid prompts that duplicate what the description already covers (“We sell shoes”). Focus on decisions the model has to make that only you know the answer to. If both the prompt and the store description are empty, curation is skipped entirely.

​
Previewing suggestions

The configuration page includes a live Preview panel. Enter any query and Layers shows the autocomplete pipeline with curation applied, so you can see exactly what a customer would see without waiting for the latest saved changes to finish applying. Preview results show:
  • The display label, with a strikethrough on any suggestion the model would suppress.
  • The original canonical text alongside a rewritten label, when the two differ.
  • The model’s short reason for its decision in the tooltip (visible to you, never to customers).
Previews are rate-limited to 10 requests per minute per user per store to control model costs.

​
Configuration

Configure curation in the Layers dashboard under Settings → Autocomplete. For step-by-step instructions, see Configure autocomplete curation.

​
Interactions with other features

  • Stem deduplication and relevance gating run before curation. Curation only sees clusters that already passed those filters, so your prompt doesn’t need to cover typos or obvious duplicates.
  • Semantic redirects run after curation, on the final suggestion text. A suggestion that is re-labelled can still trigger a matching redirect.
  • Query interpretation is unaffected — curation decides which clusters appear as suggestions, not how the query itself is interpreted.

​
See also